new

Get trending papers in your email inbox!

Subscribe

Daily Papers

byAK and the research community

Nov 17

A Speech-LLM Framework for Enrollment-Free Wake-Word-Guided Target-Speaker Recognition in Complex Noisy Scenarios for Seamless Human-Machine Interaction

Target-speaker automatic speech recognition (TS-ASR) is essential for reliable human-machine interaction in complex acoustic environments, where overlapping speech and background noise make it difficult to identify and recognize the intended speaker. Existing approaches typically rely on speaker enrollment or cascaded processing pipelines, limiting their flexibility in practical applications. To address these limitations, we propose an enrollment-free Speech-LLM framework that leverages wake-word text as a lexical cue to guide target-speaker recognition. Specifically, a text-cue encoder extracts lexical representations from the wake word, which are integrated into a pretrained speech encoder through cross-attention to establish an implicit association between the wake-word occurrence and the target speaker. This enables the model to selectively recognize subsequent commands from the target speaker without requiring enrollment speech or explicit wake-word boundaries. Furthermore, a large language model (LLM) unifies command transcription and non-target speaker rejection within a single autoregressive framework. To improve robustness against complex acoustic interference, we develop a large-scale speech synthesis and simulation pipeline with a progressive four-stage training strategy. Experimental results demonstrate that the proposed framework achieves a command character error rate of 16.15% and a non-target speaker rejection rate of 91.56%, yielding a unified score of 87.70%, compared with 71.70% for the strongest enrollment-free baseline. These results demonstrate the effectiveness of wake-word-guided speaker selection and highlight the potential of integrating target-speaker recognition and rejection within a unified Speech-LLM framework for robust human-machine interaction.

  • 3 authors
·
Nov 14

The Character Error Vector: Decomposable errors for page-level OCR evaluation

The Character Error Rate (CER) is a key metric for evaluating the quality of Optical Character Recognition (OCR). However, this metric assumes that text has been perfectly parsed, which is often not the case. Under page-parsing errors, CER becomes undefined, limiting its use as a metric and making evaluating page-level OCR challenging, particularly when using data that do not share a labelling schema. We introduce the Character Error Vector (CEV), a bag-of-characters evaluator for OCR. The CEV can be decomposed into parsing and OCR, and interaction error components. This decomposability allows practitioners to focus on the part of the Document Understanding pipeline that will have the greatest impact on overall text extraction quality. The CEV can be implemented using a variety of methods, of which we demonstrate SpACER (Spatially Aware Character Error Rate) and a Character distribution method using the Jensen-Shannon Distance. We validate the CEV's performance against other metrics: first, the relationship with CER; then, parse quality; and finally, as a direct measure of page-level OCR quality. The validation process shows that the CEV is a valuable bridge between parsing metrics and local metrics like CER. We analyse a dataset of archival newspapers made of degraded images with complex layouts and find that state-of-the-art end-to-end models are outperformed by more traditional pipeline approaches. Whilst the CEV requires character-level positioning for optimal triage, thresholding on easily available values can predict the main error source with an F1 of 0.91. We provide the CEV as part of a Python library to support Document understanding research.

  • 3 authors
·
Apr 6

YazSes: An Offline, Privacy-First, Cross-Platform Hold-to-Talk Voice-Dictation System

Cloud voice-dictation services deliver strong accuracy but require streaming a user's speech to a remote provider, an unacceptable trade-off in privacy-sensitive professions and offline or air-gapped settings; the leading on-device alternatives are either platform-locked or aimed at expert scripting rather than plug-and-play dictation. We present YazSes, an open-source (Apache-2.0) hold-to-talk voice dictation daemon that runs entirely on-device, with a single codebase targeting Linux, macOS, and Windows through a protocol-based platform abstraction. YazSes transcribes speech locally with faster-whisper (CPU, int8) and injects the result into the focused application; a fast regex command grammar, backed by an optional small-language-model router, maps utterances to editor and terminal actions. Nothing leaves the machine: recording is push-to-talk rather than always-listening, there is no telemetry, and an opt-in personalization loop keeps its corpus encrypted on-device and proposes configuration changes instead of shipping data out. We describe the system architecture -- a staged pipeline behind a protocol-based platform abstraction with a JSON-RPC control plane -- and its privacy and threat model. We evaluate the shipping Python implementation on a single commodity Linux laptop; the macOS and Windows backends are implemented and unit-tested but not end-to-end evaluated here. On 200 LibriSpeech test-clean utterances spanning 40 speakers, word error rate ranges from 4.82% (tiny.en) to 2.59% (small.en) at a real-time factor of 0.520 for small.en, decoding faster than real time on CPU with no GPU. The command grammar reaches 100% action accuracy with a 0.0% false-positive rate on plain dictation at 0.021 ms per call, and the non-decode pipeline adds 0.289 ms of overhead. The system and the reproducible benchmark harness behind every number in this paper are public.

  • 1 authors
·
Jul 29

QuoteBench: How Matched Scores Can Hide Command-Path Failures

LLM coding agents issue Bash commands through interfaces that may serialize, wrap, and reparse model output. Matched execution scores alone cannot distinguish command-generation errors from failures introduced after generation. QuoteBench measures this boundary with exact final-state validation on 56 one-shot tasks from 14 incident-derived families, crossing the generation contract with the execution transport around one deliberately unescaped added parser. Escaping at the interpolation point reproduces each replayed reply's raw-path outcome, so any recovery under a disclosed boundary must come from the model changing its generation. Across eight same-window configurations, replaying the same reply through the added parser lowers success by 55.4 to 73.2 percentage points; disclosure recovers 30.4 to 60.7 points for six configurations, and zero or slightly negative for the other two. Raw generation is nearly saturated at the frontier; boundary adaptation is what still separates models. GPT-5.6-sol's matched gap of -3.6 points hides -64.3 points of damage and +60.7 points of compensation. The deployment configuration reorders models: one reversal among 26 comparable pairs is unambiguous and four more sit on single-task margins. Evaluations of command-issuing agents should report the model configuration, generation contract, execution path, operating point, and final-state validator rather than treat a matched score as an intrinsic model property.

  • 4 authors
·
Aug 12 4

CORAA: a large corpus of spontaneous and prepared speech manually validated for speech recognition in Brazilian Portuguese

Automatic Speech recognition (ASR) is a complex and challenging task. In recent years, there have been significant advances in the area. In particular, for the Brazilian Portuguese (BP) language, there were about 376 hours public available for ASR task until the second half of 2020. With the release of new datasets in early 2021, this number increased to 574 hours. The existing resources, however, are composed of audios containing only read and prepared speech. There is a lack of datasets including spontaneous speech, which are essential in different ASR applications. This paper presents CORAA (Corpus of Annotated Audios) v1. with 290.77 hours, a publicly available dataset for ASR in BP containing validated pairs (audio-transcription). CORAA also contains European Portuguese audios (4.69 hours). We also present a public ASR model based on Wav2Vec 2.0 XLSR-53 and fine-tuned over CORAA. Our model achieved a Word Error Rate of 24.18% on CORAA test set and 20.08% on Common Voice test set. When measuring the Character Error Rate, we obtained 11.02% and 6.34% for CORAA and Common Voice, respectively. CORAA corpora were assembled to both improve ASR models in BP with phenomena from spontaneous speech and motivate young researchers to start their studies on ASR for Portuguese. All the corpora are publicly available at https://github.com/nilc-nlp/CORAA under the CC BY-NC-ND 4.0 license.

  • 11 authors
·
Oct 14, 2021

Wieszcz-XIX: A 3.1-Billion-Word Corpus of Pre-1918 Polish and Temporally Bounded Language Models Trained From Scratch

Historical Polish is well documented as a language but annotated in machine-readable form only to about a million words for the period this paper covers; the rest sits behind optical character recognition of variable quality. We present Wieszcz-XIX, a corpus of 6.75 billion tokens (about 3.1 billion words) in 294,369 documents, most of them periodical issues, of Polish published from 1800 to 1918, assembled from Wolne Lektury and the Internet Archive by a pipeline that filters, deduplicates, audits for post-1918 leakage and splits at the document level. It is over three orders of magnitude larger than the annotated corpus of the same period, and we quantify its defects: recognition corruption against a false-positive floor, near-identical duplication, which is removed, and post-1918 leakage, which is excluded from the training corpus itself down to a known residue of 0.04 to 0.38% of its bytes, found in the transcribed source, so the published corpus is the trained one document for document. On a hand-corrected sample the character error rate is 0.68% where the text is legible, and 45% of the sampled passages cannot be corrected. On it we train a ladder of decoder-only models from 47M to 349M parameters from scratch, and measure their temporal boundedness. Against two modern Polish base models, one far larger, the 349M shows a crossover, as does the 107M against the comparator of its size: post-1918 vocabulary costs them about 3.1 bits per byte more than period vocabulary, a gap the comparators do not show, and period vocabulary costs them fewer bits than it costs the comparators. Shown period text, the models keep its spelling and the comparators only partly. Adding parameters gains about twice as much as a second pass over the data. We release the corpus, code and weights. Content warning: the models reproduce period prejudice, including antisemitic statements.

  • 1 authors
·
Oct 5

Breeze Taigi: Benchmarks and Models for Taiwanese Hokkien Speech Recognition and Synthesis

Taiwanese Hokkien (Taigi) presents unique opportunities for advancing speech technology methodologies that can generalize to diverse linguistic contexts. We introduce Breeze Taigi, a comprehensive framework centered on standardized benchmarks for evaluating Taigi speech recognition and synthesis systems. Our primary contribution is a reproducible evaluation methodology that leverages parallel Taiwanese Mandarin resources. We provide 30 carefully curated Mandarin-Taigi audio pairs from Taiwan's Executive Yuan public service announcements with normalized ground truth transcriptions. We establish Character Error Rate (CER) as the standard metric and implement normalization procedures to enable fair cross-system comparisons. To demonstrate the benchmark's utility and provide reference implementations, we develop speech recognition and synthesis models through a methodology that leverages existing Taiwanese Mandarin resources and large-scale synthetic data generation. In particular, we fine-tune a Whisper model on approximately 10,000 hours of Taigi synthetic speech data. Our ASR model achieves 30.13% average CER on the benchmark, outperforming existing commercial and research systems. By providing standardized evaluation protocols, diverse training datasets, and open baseline models, we offer a replicable framework with methodologies applicable to various linguistic contexts.

  • 8 authors
·
Feb 25

Benchmarking Akan ASR Models Across Domain-Specific Datasets: A Comparative Evaluation of Performance, Scalability, and Adaptability

Most existing automatic speech recognition (ASR) research evaluate models using in-domain datasets. However, they seldom evaluate how they generalize across diverse speech contexts. This study addresses this gap by benchmarking seven Akan ASR models built on transformer architectures, such as Whisper and Wav2Vec2, using four Akan speech corpora to determine their performance. These datasets encompass various domains, including culturally relevant image descriptions, informal conversations, biblical scripture readings, and spontaneous financial dialogues. A comparison of the word error rate and character error rate highlighted domain dependency, with models performing optimally only within their training domains while showing marked accuracy degradation in mismatched scenarios. This study also identified distinct error behaviors between the Whisper and Wav2Vec2 architectures. Whereas fine-tuned Whisper Akan models led to more fluent but potentially misleading transcription errors, Wav2Vec2 produced more obvious yet less interpretable outputs when encountering unfamiliar inputs. This trade-off between readability and transparency in ASR errors should be considered when selecting architectures for low-resource language (LRL) applications. These findings highlight the need for targeted domain adaptation techniques, adaptive routing strategies, and multilingual training frameworks for Akan and other LRLs.

  • 8 authors
·
Jul 2, 2025

Type-Safe Is Not Error-Free: A Constrained Decision Head Follows the Option Name, Not the Rubric Bound to It

Typed decision models are built for settings where model outputs are consumed directly by software. Instead of generating free-form text, they return a decision over a predefined set of options. By construction, every output conforms to the required schema. Yet this guarantee does not tell us whether the model interprets the options as intended. We study Jev and two Jev-like models with open weights by changing how option names are assigned to rubrics. Each option consists of an option name and a textual rubric that defines what the option means. We change only which option name is assigned to each rubric; the question, state, rubric wording, and set of option names remain exactly the same. On 1200 workflow decisions with task-specific rubrics, renaming the two options from 0/1 to no/yes changes 70.4 more answers per hundred (95% CI: [67.6, 73.1]) and shifts AUC from .94 to .23, revealing a systematic reversal in the decision ranking rather than simple uncertainty. The same operation has little effect with neutral option names. This pattern holds across all 4 predicates, where the effect is at least 7.4x larger than under the neutral control, and becomes stronger as the number of options increases. The effect also depends on the read-out geometry: a second model family that mean-pools over the full option span flips 4.1x less often. The hosted model exhibits the same behavior: the swap changes AUC from .8146 to .5806 and produces 24x as many answer flips as its test-retest floor. In contrast, replacing the option names with random character strings returns all model families to the neutral-control regime without reducing accuracy. The failure therefore depends on the semantic polarity of the option names rather than on the renaming operation itself. Across all conditions, the type-error rate remains 0%, even when decision accuracy degrades substantially.

  • 4 authors
·
Sep 22

Tracing LLM Reasoning Processes with Strategic Games: A Framework for Planning, Revision, and Resource-Constrained Decision Making

Large language models (LLMs) are increasingly used for tasks that require complex reasoning. Most benchmarks focus on final outcomes but overlook the intermediate reasoning steps - such as planning, revision, and decision making under resource constraints. We argue that measuring these internal processes is essential for understanding model behavior and improving reliability. We propose using strategic games as a natural evaluation environment: closed, rule-based systems with clear states, limited resources, and automatic feedback. We introduce a framework that evaluates LLMs along three core dimensions: planning, revision, and resource-constrained decision making. To operationalize this, we define metrics beyond win rate, including overcorrection risk rate, correction success rate, improvement slope, and over-budget ratio. In 4320 adversarial rounds across 12 leading models, ChatGPT-o3-mini achieves the top composite score, with a win rate of 74.7 percent, a correction success rate of 78.6 percent, and an improvement slope of 0.041. By contrast, Qwen-Plus, despite an overcorrection risk rate of 81.6 percent, wins only 25.6 percent of its matches - primarily due to excessive resource use. We also observe a negative correlation between overcorrection risk rate and correction success rate (Pearson r = -0.51, p = 0.093), suggesting that more frequent edits do not always improve outcomes. Our findings highlight the value of assessing not only what LLMs decide but how they arrive at those decisions

  • 8 authors
·
Jun 13, 2025

Instruct-Tuning Pretrained Causal Language Models for Ancient Greek Papyrology and Epigraphy

This article presents an experiment in fine-tuning a pretrained causal language model (Meta's Llama 3.1 8B Instruct) for aiding in three fundamental tasks of philological research: chronological and geographic attribution as well as text restoration in ancient Greek inscriptions and documentary papyri. Using a prompt-based instruct approach, the fine-tuned models surpass the state of the art in key metrics. For inscriptions, the models achieve a lower average character error rate (CER) of 22.5% (vs. 26.3%), while closely matching top-1 accuracy (60.9% vs. 61.8%) and top-20 accuracy (77.5% vs. 78.3%) for sequences up to 10 characters. They also provide a practical advantage by ignoring spaces during reconstruction, aligning better with the scriptio continua typically used in ancient written artifacts. In geographic attribution, the model outperforms previous benchmarks with a top-1 accuracy of 75.0% (vs. 70.8%) and a top-3 accuracy of 83.7% (vs. 82.1%). For dating, it achieves an average deviation of 26.2 years (vs. 29.3) and a median deviation of 1 year (vs. 3) from the actual date range. The models also set new baselines for documentary papyri, with a CER of 16.3%, a top-1 accuracy of 71.3%, and top-20 of 85.0% in text reconstruction; a top-1 accuracy of 66.4% and top-3 of 79.9% in geographic attribution; and, in chronological attribution, a deviation of 21.7 years from the actual termini post/ante quem, with a median deviation of 0 years.

  • 1 authors
·
Sep 20, 2024

T5Gemma-TTS Technical Report

Autoregressive neural codec language models have shown strong zero-shot voice cloning ability, but decoder-only architectures treat input text as a prefix that competes with the growing audio sequence for positional capacity, weakening text conditioning over long utterances. We present T5Gemma-TTS, an encoder-decoder codec language model that maintains persistent text conditioning by routing bidirectional text representations through cross-attention at every decoder layer. Built on the T5Gemma pretrained encoder-decoder backbone (2B encoder + 2B decoder; 4B parameters), it inherits rich linguistic knowledge without phoneme conversion and processes text directly at the subword level. To improve duration control, we introduce Progress-Monitoring Rotary Position Embedding (PM-RoPE) in all 26 cross-attention layers, injecting normalized progress signals that help the decoder track target speech length. Trained on 170,000 hours of multilingual speech in English, Chinese, and Japanese, T5Gemma-TTS achieves a statistically significant speaker-similarity gain on Japanese over XTTSv2 (0.677 vs. 0.622; non-overlapping 95% confidence intervals) and the highest numerical Korean speaker similarity (0.747) despite Korean not being included in training, although this margin over XTTSv2 (0.741) is not statistically conclusive. It also attains the lowest numerical Japanese character error rate among five baselines (0.126), though this ranking should be interpreted cautiously because of partial confidence-interval overlap with Kokoro. English results on LibriSpeech should be viewed as an upper-bound estimate because LibriHeavy is a superset of LibriSpeech. Using the same checkpoint, disabling PM-RoPE at inference causes near-complete synthesis failure: CER degrades from 0.129 to 0.982 and duration accuracy drops from 79% to 46%. Code and weights are available at https://github.com/Aratako/T5Gemma-TTS.

  • 2 authors
·
Apr 1 2

Prompt Design at Scale: How Format, Instruction Count, and Context Length Shape Instruction Adherence and Hallucination in Large Language Models

Practitioners make three prompt-design decisions with almost no controlled evidence behind them: how to format instructions and context (markdown, plain text, prose, or tabular), how many simultaneous instructions a system prompt can carry before compliance degrades, and how much context a model can hold before recall and honesty degrade. We report two controlled experiments crossing all three factors on one held, contamination-free synthetic corpus (the "Book of Veyra," 8,780 uniquely-named entities, deterministically regenerable from a fixed seed), evaluated across five models. Experiment 1 (960 calls/model) measures instruction-following decay as rule count N grows from 10 to 160, crossed with four formats and system-prompt vs. user-turn placement. Perfect-response rate collapses to zero by N=80 for every model, format, and placement. Placement produces effects at least as large as format at N=160 in most models, but the direction is model-specific. No model shows a reliable markdown advantage; one 35B model favors plain text instead. Experiment 2 (5,520 calls/model) measures recall accuracy, false-premise sycophancy, and absent-fact fabrication across a 2k-to-512k-token context ladder in the same four formats. Recall stays near ceiling through 64-128k tokens, then degrades sharply and format-dependently: one model's accuracy spread reaches 48 points at 128k tokens. Fabrication never occurs (0/5,760 probes), and sycophancy stays negligible (<=8.3%). What rises sharply near each model's context ceiling is outright refusal to answer (0% to 79-90%), distinct from sycophancy or fabrication. Neither pre-registered format ordering holds, and token overhead (+22% to +37% over plain text) further changes which format is preferable where accuracy spread is genuine. We release the full harness, corpus generator, and raw results (VeyraBench): https://github.com/iNetanel/veyrabench

  • 1 authors
·
Jul 20

Pervasive Annotation Errors Break Text-to-SQL Benchmarks and Leaderboards

Researchers have proposed numerous text-to-SQL techniques to streamline data analytics and accelerate the development of data-driven applications. To compare these techniques and select the best one for deployment, the community depends on public benchmarks and their leaderboards. Since these benchmarks heavily rely on human annotations during question construction and answer evaluation, the validity of the annotations is crucial. In this paper, we conduct an empirical study that (i) benchmarks annotation error rates for two widely used text-to-SQL benchmarks, BIRD and Spider 2.0-Snow, and (ii) corrects a subset of the BIRD development (Dev) set to measure the impact of annotation errors on text-to-SQL agent performance and leaderboard rankings. Through expert analysis, we show that BIRD Mini-Dev and Spider 2.0-Snow have error rates of 52.8% and 62.8%, respectively. We re-evaluate all 16 open-source agents from the BIRD leaderboard on both the original and the corrected BIRD Dev subsets. We show that performance changes range from -7% to 31% (in relative terms) and rank changes range from -9 to +9 positions. We further assess whether these impacts generalize to the full BIRD Dev set. We find that the rankings of agents on the uncorrected subset correlate strongly with those on the full Dev set (Spearman's r_s=0.85, p=3.26e-5), whereas they correlate weakly with those on the corrected subset (Spearman's r_s=0.32, p=0.23). These findings show that annotation errors can significantly distort reported performance and rankings, potentially misguiding research directions or deployment choices. Our code and data are available at https://github.com/uiuc-kang-lab/text_to_sql_benchmarks.

  • 4 authors
·
Jan 13

How Many Tries Does It Take? Iterative Self-Repair in LLM Code Generation Across Model Scales and Benchmarks

Large language models frequently fail to produce correct code on their first attempt, yet most benchmarks evaluate them in a single-shot setting. We investigate iterative self-repair (feeding execution errors back to the model for correction) across seven models spanning three families and both open-weight and proprietary providers: Llama 3.1 8B, Llama 3.3 70B, Llama 4 Scout (MoE, 16 experts), Llama 4 Maverick (MoE, 128 experts), Qwen3 32B, Gemini 2.5 Flash, and Gemini 2.5 Pro. On HumanEval (164 problems) and MBPP Sanitized (257 problems) with up to five attempts, self-repair universally improves pass rates: +4.9 to +17.1 pp on HumanEval and +16.0 to +30.0 pp on MBPP. Gemini 2.5 Flash achieves the highest final pass rates (96.3% HumanEval, 93.8% MBPP). Most gains concentrate in the first two rounds.Error-type analysis shows assertion errors (logical mistakes) are the hardest to repair at ~45%, while syntax and name errors are repaired at substantially higher rates, connecting to broader findings on the limits of LLM self-correction. Prior work found that weaker models fail at self-repair or require fine-tuning; we show that modern instruction-tuned models succeed with prompting alone, even at 8B scale. We also provide the first comparison of dense and MoE architectures for self-repair, and extend the repair-vs-resampling tradeoff analysis to modern models. A prompt ablation reveals chain-of-thought repair yields up to +5.5 pp additional self-repair gain (measured as improvement in repair delta) over minimal prompting for capable models.

  • 1 authors
·
Apr 11

Comparative Analysis of Multilingual Pre-trained Models for Nepali Automatic Speech Recognition

Multilingual pretrained models nominally support Nepali, yet no controlled benchmark has compared them under a single fine-tuning protocol. We fine-tune six pretrained models (XLSR-53, IndicWav2Vec, MMS-1B, Whisper-Medium, Whisper-Large-v3-Turbo, and Conformer-Hi) spanning CTC self-supervised, autoregressive encoder-decoder, and hybrid Conformer-CTC architectures, on the OpenSLR SLR54 Nepali corpus (~165 hours) using identical preprocessing, splits, optimizer, and family-matched learning-rate schedules. We evaluate Word Error Rate (WER), Character Error Rate (CER), and Real-Time Factor (RTF) on three independent test sets (OpenSLR, FLEURS, Common Voice). Whisper-Large-v3-Turbo (14.76% WER) and IndicWav2Vec (14.89% WER) tie at the top despite a 9x parameter gap and 40x pretraining-data gap, providing direct empirical evidence that language-family proximity in pretraining can substitute for raw scale for in-domain Nepali. CTC decoders run up to 29x faster than autoregressive Whisper at the same accuracy, flipping the practical deployment preference toward CTC under any latency budget. Massively multilingual pretraining (MMS-1B) yields the smallest out-of-domain degradation on FLEURS (+12.55 pp), indicating that scale buys robustness rather than peak in-domain accuracy. The resulting benchmark provides the first standardized, multi-model, efficiency-aware reference numbers for Nepali ASR.

  • 2 authors
·
May 30

Beyond Word Error Rate: A Failure Taxonomy of Streaming Dictation ASR

Speech recognition (ASR) systems are usually compared by word error rate (WER) on recorded files, after punctuation, capital letters and number format are removed. In dictation, the audio is streamed live and the text goes into the user's message as written, so what matters is which mistakes the user has to fix. The test set, streamed to eleven ASR systems, comprises 106 scripted English dictations from one speaker (read in a quiet room, with a fan on, and whispered), 20 clips in each of eleven other language and dialect sets, and 400 recordings that join clips in two languages. Every error was sorted by a fixed rule into four levels: recognition, output formatting, language handling and, for the self-hosted models, audio dropped before the model. Systems with similar WER failed in different ways. Zoom Scribe and OpenAI's GPT-Live-Transcribe have almost the same WER on quiet English (7.4% and 7.3%), but 34.7% and 8.8% on whispered English. Zoom and both AssemblyAI models left out 26-28% of whispered words, and still 21-26% with their end-of-turn detection relaxed or off; seven other systems left out at most 3%. Qwen3-ASR's 32% came from the speech detection in our own deployment (6% when relaxed). Across 107 EARS speakers, the same four systems lost 7-31% of whispered words, while six others lost under 2%. With only English selected, Gemini, Deepgram and our self-hosted Whisper lost the non-English part of a two-language recording in 86-99% of recordings, against 33-46% with both languages selected, while GPT-Live-Transcribe, OpenAI's realtime Whisper and Meta lost it in at most 20%. Running Whisper behind a Silero voice activity detector that drops audio judged to be silence raised whispered WER from 10.1% to 51.4%. Most English results come from one read speaker. The code, the recordings, both references and every transcript are released.

  • 4 authors
·
Nov 11

What Makes a Good Terminal-Agent Benchmark Task: A Guideline for Adversarial, Difficult, and Legible Evaluation Design

Terminal-agent benchmarks have become a primary signal for measuring the coding and system-administration capabilities of large language models. As the market for evaluation environments grows, so does the pressure to ship tasks quickly, often without thorough adversarial review of the verification logic. This paper is a guideline for writing good benchmark tasks, drawn from over a year of contributing to and reviewing tasks for Terminal Bench. Most people write benchmark tasks the way they write prompts. They shouldn't. A prompt is designed to help the agent succeed; a benchmark is designed to find out if it can. We argue that good tasks are adversarial, difficult, and legible, and that a large class of common failure modes -- AI-generated instructions, over-prescriptive specifications, clerical difficulty, oracle solutions that assume hidden knowledge, tests that validate the wrong things, and reward-hackable environments -- are predictable consequences of treating task authoring as prompt authoring. We catalog these failure modes, argue that real difficulty is conceptual rather than environmental, and discuss recent empirical evidence that over 15% of tasks in popular terminal-agent benchmarks are reward-hackable. We hope this serves as a useful reference for benchmark maintainers, task contributors, and researchers using benchmark scores as evidence.

  • 1 authors
·
Apr 29

VoiceCodeBench: Evaluating Exact Structured-Token Recovery in Automatic Speech Recognition

Automatic speech recognition (ASR) systems are commonly evaluated with word error rate (WER), yet many voice workflows depend on exact written values for identifiers, paths, and measured quantities. A transcript can appear fluent and achieve low WER while corrupting a value that a downstream system must parse, store, or execute. We introduce VoiceCodeBench, a benchmark for evaluating exact structured-token recovery in English ASR. It contains 300 human-recorded workplace segments spanning eight workflow domains and 1,482 audited target entities across 26 entity types, each with a canonical written form recoverable from the audio. Under a raw-audio-only protocol, systems receive audio bytes without additional context or metadata. Alongside WER, we evaluate Canonical Token/Entity Match (CTEM), Task Success Rate (TSR), and per-type exact recovery. Across 12 baseline ASR systems, lower WER generally corresponded to better structured-token recovery but did not fully determine it: Spearman correlations were -0.73 for both WER versus CTEM and WER versus TSR. The strongest baseline by TSR reached only 68.7%, leaving nearly one third of recordings with at least one unrecovered workflow-critical value. These results show that entity-sensitive metrics are needed to assess whether ASR output preserves exact values that production systems must parse, route, store, compare, or execute.

  • 7 authors
·
Aug 27

Evaluating LLMs at Detecting Errors in LLM Responses

With Large Language Models (LLMs) being widely used across various tasks, detecting errors in their responses is increasingly crucial. However, little research has been conducted on error detection of LLM responses. Collecting error annotations on LLM responses is challenging due to the subjective nature of many NLP tasks, and thus previous research focuses on tasks of little practical value (e.g., word sorting) or limited error types (e.g., faithfulness in summarization). This work introduces ReaLMistake, the first error detection benchmark consisting of objective, realistic, and diverse errors made by LLMs. ReaLMistake contains three challenging and meaningful tasks that introduce objectively assessable errors in four categories (reasoning correctness, instruction-following, context-faithfulness, and parameterized knowledge), eliciting naturally observed and diverse errors in responses of GPT-4 and Llama 2 70B annotated by experts. We use ReaLMistake to evaluate error detectors based on 12 LLMs. Our findings show: 1) Top LLMs like GPT-4 and Claude 3 detect errors made by LLMs at very low recall, and all LLM-based error detectors perform much worse than humans. 2) Explanations by LLM-based error detectors lack reliability. 3) LLMs-based error detection is sensitive to small changes in prompts but remains challenging to improve. 4) Popular approaches to improving LLMs, including self-consistency and majority vote, do not improve the error detection performance. Our benchmark and code are provided at https://github.com/psunlpgroup/ReaLMistake.

  • 15 authors
·
Apr 4, 2024

Data Generation for Post-OCR correction of Cyrillic handwriting

This paper introduces a novel approach to post-Optical Character Recognition Correction (POC) for handwritten Cyrillic text, addressing a significant gap in current research methodologies. This gap is due to the lack of large text corporas that provide OCR errors for further training of language-based POC models, which are demanding in terms of corpora size. Our study primarily focuses on the development and application of a synthetic handwriting generation engine based on B\'ezier curves. Such an engine generates highly realistic handwritten text in any amounts, which we utilize to create a substantial dataset by transforming Russian text corpora sourced from the internet. We apply a Handwritten Text Recognition (HTR) model to this dataset to identify OCR errors, forming the basis for our POC model training. The correction model is trained on a 90-symbol input context, utilizing a pre-trained T5 architecture with a seq2seq correction task. We evaluate our approach on HWR200 and School_notebooks_RU datasets as they provide significant challenges in the HTR domain. Furthermore, POC can be used to highlight errors for teachers, evaluating student performance. This can be done simply by comparing sentences before and after correction, displaying differences in text. Our primary contribution lies in the innovative use of B\'ezier curves for Cyrillic text generation and subsequent error correction using a specialized POC model. We validate our approach by presenting Word Accuracy Rate (WAR) and Character Accuracy Rate (CAR) results, both with and without post-OCR correction, using real open corporas of handwritten Cyrillic text. These results, coupled with our methodology, are designed to be reproducible, paving the way for further advancements in the field of OCR and handwritten text analysis. Paper contributions can be found in https://github.com/dbrainio/CyrillicHandwritingPOC

  • 5 authors
·
Nov 27, 2023

Compression Favors Consistency, Not Truth: When and Why Language Models Prefer Correct Information

Why do language models sometimes prefer correct statements even when trained on mixed-quality data? We introduce the Compression--Consistency Principle: next-token prediction favors hypotheses that allow shorter and more internally consistent descriptions of the training data. Truth bias emerges only when false alternatives are structurally harder to compress. We test this using small GPT-2-style character-level transformers (3.5M--86M parameters) on synthetic math corpora with controlled mixtures of correct and incorrect rules. In the random-error setting, models strongly prefer correct completions in paired evaluation: 83.1% accuracy at balanced data and 67.0% even when correct rules appear in only 10% of the corpus. Replacing random errors with a coherent but mathematically incorrect rule system largely eliminates the preference (near-chance accuracy). In a more natural-language-like synthetic world, the effect is weaker but still present (57.7%). Additional experiments show that embedding verification steps can restore preference for correctness even at small scale, while increasing the number of consistent rules produces a graded improvement in accuracy. Our results suggest that what appears as a "truth bias" is largely a side effect of compression pressure and preference for internal consistency, rather than an intrinsic drive toward truth. Full code and data are available at https://github.com/Rai220/compression-drives-truth.

  • 1 authors
·
Mar 12 2

Meta-Tool: Efficient Few-Shot Tool Adaptation for Small Language Models

Can small language models achieve strong tool-use performance without complex adaptation mechanisms? This paper investigates this question through Meta-Tool, a controlled empirical study comparing hypernetwork-based LoRA adaptation against carefully designed few-shot prompting. Using a Llama-3.2-3B-Instruct backbone, we evaluate four adaptation mechanisms--few-shot prompting, documentation encoding, hypernetwork-generated LoRA weights, and value-guided beam search--across four diverse benchmarks: Gorilla APIBench, Spider 2.0, WebArena, and InterCode. Our central finding is a well-supported negative result: despite generating non-trivial weight matrices, the 227.8M-parameter hypernetwork provides no measurable improvement over few-shot prompting alone. Comprehensive ablation studies reveal that few-shot examples contribute +21.5% to performance and documentation contributes +5.0%, while the hypernetwork adds 0%. A 3B model with well-designed prompts achieves 79.7% of GPT-5's average performance at 10 times lower latency. Error analysis across 722 failure cases spanning all shot counts (0--5) shows that at the 5-shot configuration (106 failures), failure modes are task-dependent: schema-heavy tasks (Spider 2.0, WebArena) show near-zero format errors with remaining failures semantic, while format errors dominate on Gorilla (100%) and InterCode (70%). These findings redirect practitioners toward prompt engineering and example curation rather than complex adaptation architectures.

  • 1 authors
·
Apr 21

What Single-Prompt Accuracy Misses: A Multi-Variant Reliability Audit of Language Models

Single-prompt accuracy is the dominant way to benchmark language models, but it can miss reliability failures that matter. We evaluate a 15-model open-weight corpus, with the main reliability analyses focused on 10 instruct models across five classification and reasoning benchmarks under five prompt variants each, measuring accuracy, token-probability calibration, verbal-confidence calibration, verbal parse rate, and prompt-perturbation spread for every (model x dataset x variant) cell. We find three broad results. First, evaluation design can materially change the conclusion. Switching Expected Calibration Error (ECE) token from a raw to a label-set-normalised definition changes per-cell calibration by a mean absolute 0.149. More strikingly, pairing a chain-of-thought prompt with a first-character evaluator on ARC-Challenge reduces apparent accuracy by 72-88% across all five primary models; two independent repair procedures recover 93.8% and 102.7% of the lost performance, indicating an evaluator-side rather than model-side failure. Second, confidence signals are fragile. On MMLU-Pro, every primary model verbally reports confidence substantially above both its accuracy and its token-probability confidence on the same rows, and verbal parse rate can collapse for a single model on a single prompt variant. Third, prompt robustness does not track parameter count reliably. Across 10 instruct models, the correlation between model size and prompt-perturbation spread ranges from -0.244 to 0.474 across benchmarks. Taken together, these results show that reliability conclusions for small language models depend not only on the model being evaluated, but also on the evaluation pipeline used to measure it. We argue that calibration definitions, evaluator logic, verbal parseability, and prompt robustness should be reported explicitly when making reliability claims.

  • 2 authors
·
May 2

CharBench: Evaluating the Role of Tokenization in Character-Level Tasks

Tasks that require character-level reasoning, such as counting or locating characters within words, remain challenging for contemporary language models. A common conjecture is that language models' reliance on subword units, rather than characters, contributes to their struggles with character-level tasks, yet recent studies offer conflicting conclusions about the role of tokenization, leaving its impact unclear. To address this gap, we introduce CharBench, a comprehensive benchmark of character-level tasks that is two orders of magnitude larger than existing alternatives. We evaluate a diverse range of leading open-weight and proprietary models on CharBench and find that it presents a significant challenge to modern LLMs, with an average accuracy of 43.6% and 32.3% on some tasks. We present an in-depth analysis of how intrinsic properties of words and their segmentations into tokens correspond to model performance. For counting tasks, we find that tokenization properties are weakly correlated with correctness, while the length of the queried word and the actual character count play a more significant part. In contrast, for tasks requiring intra-word positional understanding, performance is negatively correlated with the length of the token containing the queried character, suggesting that longer tokens obscure character position information for LLMs. We encourage future work to build on the benchmark and evaluation methodology introduced here as tools for improving model performance on such tasks.

  • 2 authors
·
Aug 4, 2025

RMCBench: Benchmarking Large Language Models' Resistance to Malicious Code

The emergence of Large Language Models (LLMs) has significantly influenced various aspects of software development activities. Despite their benefits, LLMs also pose notable risks, including the potential to generate harmful content and being abused by malicious developers to create malicious code. Several previous studies have focused on the ability of LLMs to resist the generation of harmful content that violates human ethical standards, such as biased or offensive content. However, there is no research evaluating the ability of LLMs to resist malicious code generation. To fill this gap, we propose RMCBench, the first benchmark comprising 473 prompts designed to assess the ability of LLMs to resist malicious code generation. This benchmark employs two scenarios: a text-to-code scenario, where LLMs are prompted with descriptions to generate code, and a code-to-code scenario, where LLMs translate or complete existing malicious code. Based on RMCBench, we conduct an empirical study on 11 representative LLMs to assess their ability to resist malicious code generation. Our findings indicate that current LLMs have a limited ability to resist malicious code generation with an average refusal rate of 40.36% in text-to-code scenario and 11.52% in code-to-code scenario. The average refusal rate of all LLMs in RMCBench is only 28.71%; ChatGPT-4 has a refusal rate of only 35.73%. We also analyze the factors that affect LLMs' ability to resist malicious code generation and provide implications for developers to enhance model robustness.

  • 9 authors
·
Sep 23, 2024

From Zero to OCR: A Data-Centric Framework for Unsupported Low-Resource Scripts

Optical character recognition (OCR) remains inaccessible to many low-resource scripts because annotated document images are unavailable and synthetic training data can fail to preserve script-specific visual structure. We present a data-centric framework that builds end-to-end OCR for such scripts from only an unannotated text corpus and digital fonts, without manually annotating a single training image. A synthetic data engine renders word, line and page images with correct complex-script shaping and automatically derived labels; a YOLOv8 word detector and a convolutional recurrent recognizer trained with connectionist temporal classification learn from these images; and a stage-gate evaluation protocol isolates recognition error before detection is introduced. We demonstrate the framework on Kashmiri, a Perso-Arabic Nastaliq script with approximately 6.8 million speakers that none of six widely used OCR engines supports. Three line-level recognizers, including one transferred from Arabic, stalled between 41% and 79% character error rate (CER), motivating a word-level formulation; subsequent improvements came primarily from rendering fidelity, diacritic-preserving crop margins and tuned non-maximum suppression. The resulting system achieves 0.39% CER on held-out words, 10.0% on synthetic pages and 13.5% on browser-rendered digital pages, and establishes reference results of 18.45% on typeset book pages and 49.8% on degraded historical documents. We publicly release the trained models, the inference code and a synthetic dataset of 285,864 Kashmiri text-line images, and we describe a step-by-step procedure for building OCR for other unsupported scripts.

  • 3 authors
·
Oct 17

RedCode: Risky Code Execution and Generation Benchmark for Code Agents

With the rapidly increasing capabilities and adoption of code agents for AI-assisted coding, safety concerns, such as generating or executing risky code, have become significant barriers to the real-world deployment of these agents. To provide comprehensive and practical evaluations on the safety of code agents, we propose RedCode, a benchmark for risky code execution and generation: (1) RedCode-Exec provides challenging prompts that could lead to risky code execution, aiming to evaluate code agents' ability to recognize and handle unsafe code. We provide a total of 4,050 risky test cases in Python and Bash tasks with diverse input formats including code snippets and natural text. They covers 25 types of critical vulnerabilities spanning 8 domains (e.g., websites, file systems). We provide Docker environments and design corresponding evaluation metrics to assess their execution results. (2) RedCode-Gen provides 160 prompts with function signatures and docstrings as input to assess whether code agents will follow instructions to generate harmful code or software. Our empirical findings, derived from evaluating three agent frameworks based on 19 LLMs, provide insights into code agents' vulnerabilities. For instance, evaluations on RedCode-Exec show that agents are more likely to reject executing risky operations on the operating system, but are less likely to reject executing technically buggy code, indicating high risks. Risky operations described in natural text lead to a lower rejection rate than those in code format. Additionally, evaluations on RedCode-Gen show that more capable base models and agents with stronger overall coding abilities, such as GPT4, tend to produce more sophisticated and effective harmful software. Our findings highlight the need for stringent safety evaluations for diverse code agents. Our dataset and code are available at https://github.com/AI-secure/RedCode.

  • 8 authors
·
Nov 12, 2024 1

Robust Tool Use via Fission-GRPO: Learning to Recover from Execution Errors

Large language models (LLMs) can call tools effectively, yet they remain brittle in multi-turn execution: following a tool call error, smaller models often degenerate into repetitive invalid re-invocations, failing to interpret error feedback and self-correct. This brittleness hinders reliable real-world deployment, where the execution errors are inherently inevitable during tool interaction procedures. We identify a key limitation of current approaches: standard reinforcement learning (RL) treats errors as sparse negative rewards, providing no guidance on how to recover, while pre-collected synthetic error-correction datasets suffer from distribution mismatch with the model's on-policy error modes. To bridge this gap, we propose Fission-GRPO, a framework that converts execution errors into corrective supervision within the RL training loop. Our core mechanism fissions each failed trajectory into a new training instance by augmenting it with diagnostic feedback from a finetuned Error Simulator, then resampling recovery rollouts on-policy. This enables the model to learn from the precise errors it makes during exploration, rather than from static, pre-collected error cases. On the BFCL v4 Multi-Turn, Fission-GRPO improves the error recovery rate of Qwen3-8B by 5.7% absolute, crucially, yielding a 4% overall accuracy gain (42.75% to 46.75%) over GRPO and outperforming specialized tool-use agents.

  • 9 authors
·
Jan 21 2

System X: An LLM Decision Model That Replies from Its Own KV Cache

Decision models return probabilities over supplied options, but a deployed agent may also need to reply in words. The present study introduces System X, a large language model (LLM) decision model that adds a text-reply option to the LLM-as-Jev readout. When this option wins, or confidence is low, the same model continues from the decision's key-value (KV) cache, avoiding a repeated prompt prefill. Qwen3-0.6B has been fine-tuned with LoRA on Reason-9k and Bitext reply rows. JevBench accuracy shows no statistically significant change (55.4% to 54.5%, p=0.87), while expected calibration error falls from 0.292 to 0.169. Held-out Bitext routing accuracy reaches 98.3%. Cache reuse saves 15-37% of latency on an Apple M4 Pro, with token-identical replies in the FP32 demonstration cases. Scoring options on copies of the prompt cache matches full-sequence scoring to within 1.4e-6 in probability. Adding dialogue, industry and complaint data (System X+) improves support-task results but lowers IFEval accuracy from 55.5% to 45.0%. System X v4 uses Qwen3-1.7B, rebalanced data and teacher-rewritten reply targets. Against its base LLM, JevBench accuracy rises from 55.0% to 61.0% (p=0.04), with no statistically significant IFEval change (p=0.71). A matched latency test on unadapted Qwen3-1.7B saves 17-39%. One LLM judge prefers v4 replies to an unadapted 1.7B second call (31 to 11 pairs, p=0.003); the other's preference is not significant. These reply comparisons also differ in prompts and generation limits. State conflicts and unsupported reply details remain failure modes. Further studies with several seeds, human raters and real support data are recommended.

  • 3 authors
·
Oct 17

When Correct Isn't Usable: Improving Structured Output Reliability in Small Language Models

Deployed language models must produce outputs that are both correct and format-compliant. We study this structured-output reliability gap using two mathematical benchmarks -- GSM8K and MATH -- as a controlled testbed: ground truth is unambiguous and the output contract is strict (JSON with required fields). We evaluate three 7-9B models under five prompting strategies and report output accuracy -- the joint event of mathematical correctness and valid JSON structure -- as the primary metric. A systematic format failure emerges: NAIVE prompting (no system prompt) achieves up to 85% task accuracy on GSM8K but 0% output accuracy across all models and datasets. REFERENCE prompting (a minimal hand-written JSON format prompt) fares little better, yielding 0% output accuracy for two of four models tested. Constrained decoding enforces syntactic validity but incurs 3.6x-8.2x latency overhead and in several settings degrades task performance substantially. To overcome this limitation, we developed AloLab, an iterative system-prompt optimizer (meta-agent: Claude Sonnet 4.5) requiring only black-box API access to the target model; it reaches 84-87% output accuracy on GSM8K and 34-40% on MATH across five independent runs per model, with 29/30 paired McNemar comparisons against the best static prompt significant at p < 0.05, at near-NAIVE inference latency and without model fine-tuning. The same format failure extends to GPT-4o (OpenAI, 2024), a proprietary closed-source model: REFERENCE achieves 0% output accuracy due to systematic markdown-fence wrapping, while AloLab reaches 95.2% [94.8, 95.6]. An ablation replacing the Sonnet 4.5 meta-agent with Claude 3 Haiku reduces mean output accuracy to 61.0% and increases run-to-run standard deviation from <1 pp to 21.8 pp, confirming that meta-agent capability is a primary driver of optimization quality.

  • 4 authors
·
May 3

ShIOEnv: A CLI Behavior-Capturing Environment Enabling Grammar-Guided Command Synthesis for Dataset Curation

Command-line interfaces (CLIs) provide structured textual environments for system administration. Explorations have been performed using pre-trained language models (PLMs) to simulate these environments for safe interaction in high-risk environments. However, their use has been constrained to frozen, large parameter models like GPT. For smaller architectures to reach a similar level of believability, a rich dataset of CLI interactions is required. Existing public datasets focus on mapping natural-language tasks to commands, omitting crucial execution data such as exit codes, outputs, and environmental side effects, limiting their usability for behavioral modeling. We introduce a Shell Input -Output Environment (ShIOEnv), which casts command construction as a Markov Decision Process whose state is the partially built sequence and whose actions append arguments. After each action, ShIOEnv executes the candidate and returns its exit status, output, and progress toward a minimal-length behavioral objective. Due to the intractable nature of the combinatorial argument state-action space, we derive a context-free grammar from man pages to mask invalid arguments from being emitted. We explore random and proximal-policy optimization (PPO)-optimized sampling of unrestricted and grammar-masked action spaces to produce four exploration strategies. We observed that grammar masking and PPO significantly improve sample efficiency to produce a higher quality dataset (maximizing the number of arguments while minimizing redundancies). Policy-generated datasets of shell input-output behavior pairs are used to fine-tune CodeT5, where we observe 85% improvements in BLEU-4 when constraining the action space to grammar productions with an additional 26% improvement when applying PPO. The ShIOEnv environment and curated command behavior datasets are released for use in future research.

  • 2 authors
·
May 23, 2025

Evaluating Large Language Models for Code Review

Context: Code reviews are crucial for software quality. Recent AI advances have allowed large language models (LLMs) to review and fix code; now, there are tools that perform these reviews. However, their reliability and accuracy have not yet been systematically evaluated. Objective: This study compares different LLMs' performance in detecting code correctness and suggesting improvements. Method: We tested GPT4o and Gemini 2.0 Flash on 492 AI generated code blocks of varying correctness, along with 164 canonical code blocks from the HumanEval benchmark. To simulate the code review task objectively, we expected LLMs to assess code correctness and improve the code if needed. We ran experiments with different configurations and reported on the results. Results: With problem descriptions, GPT4o and Gemini 2.0 Flash correctly classified code correctness 68.50% and 63.89% of the time, respectively, and corrected the code 67.83% and 54.26% of the time for the 492 code blocks of varying correctness. Without problem descriptions, performance declined. The results for the 164 canonical code blocks differed, suggesting that performance depends on the type of code. Conclusion: LLM code reviews can help suggest improvements and assess correctness, but there is a risk of faulty outputs. We propose a process that involves humans, called the "Human in the loop LLM Code Review" to promote knowledge sharing while mitigating the risk of faulty outputs.

  • 4 authors
·
May 26, 2025

Machine Translation Meta Evaluation through Translation Accuracy Challenge Sets

Recent machine translation (MT) metrics calibrate their effectiveness by correlating with human judgement but without any insights about their behaviour across different error types. Challenge sets are used to probe specific dimensions of metric behaviour but there are very few such datasets and they either focus on a limited number of phenomena or a limited number of language pairs. We introduce ACES, a contrastive challenge set spanning 146 language pairs, aimed at discovering whether metrics can identify 68 translation accuracy errors. These phenomena range from simple alterations at the word/character level to more complex errors based on discourse and real-world knowledge. We conduct a large-scale study by benchmarking ACES on 50 metrics submitted to the WMT 2022 and 2023 metrics shared tasks. We benchmark metric performance, assess their incremental performance over successive campaigns, and measure their sensitivity to a range of linguistic phenomena. We also investigate claims that Large Language Models (LLMs) are effective as MT evaluators by evaluating on ACES. Our results demonstrate that different metric families struggle with different phenomena and that LLM-based methods fail to demonstrate reliable performance. Our analyses indicate that most metrics ignore the source sentence, tend to prefer surface-level overlap and end up incorporating properties of base models which are not always beneficial. We expand ACES to include error span annotations, denoted as SPAN-ACES and we use this dataset to evaluate span-based error metrics showing these metrics also need considerable improvement. Finally, we provide a set of recommendations for building better MT metrics, including focusing on error labels instead of scores, ensembling, designing strategies to explicitly focus on the source sentence, focusing on semantic content and choosing the right base model for representations.

  • 8 authors
·
Jan 29, 2024

The Devil in the Details: Emergent Misalignment, Format and Coherence in Open-Weights LLMs

Prior work has shown that fine-tuning models on a narrow domain with misaligned data can lead to broad misalignment - a phenomenon termed "emergent misalignment" (Betley et al. 2025). While all tested models were susceptible to emergent misalignment, some models showed more resistance than others. Specifically the Qwen-2.5 family proved to be relatively resistant, while GPT-4o exhibited the strongest misalignment. In this paper we evaluate if current-generation open-weights models exhibit similar resistance to the Qwen-2.5 family and measure misalignment robustness over a range of model architectures and scales. We replicate the effect across nine modern open-weights models (Gemma 3 and Qwen 3 families, 1B-32B parameters). Models fine-tuned on insecure code generation show a 0.68% misalignment rate (compared to 0.07% for base models), matching the lower end of prior open-model results but dramatically lower than GPT-4o's 20%. We identify a critical format-dependent vulnerability: requiring JSON output doubles misalignment rates compared to natural language prompts (0.96% vs 0.42%). This suggests that structural constraints may bypass safety training by reducing the model's 'degrees of freedom' to refuse. These findings confirm emergent misalignment as a reproducible phenomenon in modern open-weights models, with rates substantially lower than observed in proprietary systems.

  • 1 authors
·
Nov 25, 2025

When Low CER is Not Enough: An Analysis of Hallucinations in Vision-Language OCR Systems on Historical Uruguayan Documents

Optical Character Recognition (OCR) is a key component in the digitization of historical archives. Recently, Vision-Language Models (VLMs) have emerged as strong alternatives to traditional OCR systems, achieving state-of-the-art performance on standard benchmarks. However, their suitability for archival transcription remains insufficiently understood. In this work, we benchmark traditional OCR systems and VLM-based approaches on the Berrutti dataset, a challenging collection of Uruguayan dictatorship-era documents derived from microfilm scans. While VLMs consistently outperform traditional methods in terms of Character Error Rate (CER) and Word Error Rate (WER), we show that these improvements hide a more complex picture. Through a detailed qualitative analysis, we uncover systematic failure modes that are invisible to standard metrics, including orthographic normalization, spurious content generation, and semantic substitutions that preserve fluency while altering meaning. Errors affecting named entities are particularly critical, as they can introduce substantial semantic distortions with minimal impact on CER and WER. These findings reveal a critical gap between quantitative OCR performance and transcription fidelity in real-world archival settings, and highlight the need for evaluation frameworks that go beyond character-level accuracy to capture the semantic reliability of generated transcriptions.

  • 6 authors
·
Sep 16

Quran Recitation Recognition using End-to-End Deep Learning

The Quran is the holy scripture of Islam, and its recitation is an important aspect of the religion. Recognizing the recitation of the Holy Quran automatically is a challenging task due to its unique rules that are not applied in normal speaking speeches. A lot of research has been done in this domain, but previous works have detected recitation errors as a classification task or used traditional automatic speech recognition (ASR). In this paper, we proposed a novel end-to-end deep learning model for recognizing the recitation of the Holy Quran. The proposed model is a CNN-Bidirectional GRU encoder that uses CTC as an objective function, and a character-based decoder which is a beam search decoder. Moreover, all previous works were done on small private datasets consisting of short verses and a few chapters of the Holy Quran. As a result of using private datasets, no comparisons were done. To overcome this issue, we used a public dataset that has recently been published (Ar-DAD) and contains about 37 chapters that were recited by 30 reciters, with different recitation speeds and different types of pronunciation rules. The proposed model performance was evaluated using the most common evaluation metrics in speech recognition, word error rate (WER), and character error rate (CER). The results were 8.34% WER and 2.42% CER. We hope this research will be a baseline for comparisons with future research on this public new dataset (Ar-DAD).

  • 2 authors
·
May 10, 2023

Making Agents More Consistent: Skills Should Form Habits for Repeat Tasks

On repeated work, agents are inconsistent. We ran 42 tasks three times each and found that, depending on the model, 38% to 74% returned answers that did not agree. Consistency is what a buyer, an auditor, or a regulator requires, and agents do not have it. They are wasteful too: 95.3% to 97.2% of what an agent generates goes to re-deriving a plan the system already knows. We propose skill habit formation. An agent mines its own execution history for candidate skills, deterministic variants that compete against the incumbent rather than replacing it. A candidate declares the region of input space it claims, so the common case runs as a script and the rest falls through to reasoning. Four gates of ascending cost admit candidates; the central one tests a candidate's execution trace against a retained reference, within a tolerance measured from that reference's own run-to-run variability. On text-to-SQL, three of four reasoning arms reproduced their own output on 11 to 13 of 42 repeated questions and the fourth on 26 of 42, while a habit-formed variant reproduced on all 456 dispatches we repeated and was non-inferior to every arm it replaced (p<0.0001). It also used 14% to 56% fewer tokens, turning net positive after 7 to 53 reuses. We measured what this costs in accuracy. The guard admitted work it should have deferred on 2.6% of natural paraphrases and 26% of inputs near its boundary, and 11 of 13 such failures were invisible to the trace-conformance gate at any threshold. Deterministic errors repeat exactly: a bad habit is as reliable as a good one, and that is the price of the property that makes the system auditable. Separating routing from parameter extraction raised end-to-end accuracy from 0.888 to 0.952 at 43% of the cost.

  • 2 authors
·
Sep 20

Correcting diacritics and typos with a ByT5 transformer model

Due to the fast pace of life and online communications and the prevalence of English and the QWERTY keyboard, people tend to forgo using diacritics, make typographical errors (typos) when typing in other languages. Restoring diacritics and correcting spelling is important for proper language use and the disambiguation of texts for both humans and downstream algorithms. However, both of these problems are typically addressed separately: the state-of-the-art diacritics restoration methods do not tolerate other typos, but classical spellcheckers also cannot deal adequately with all the diacritics missing. In this work, we tackle both problems at once by employing the newly-developed universal ByT5 byte-level seq2seq transformer model that requires no language-specific model structures. For a comparison, we perform diacritics restoration on benchmark datasets of 12 languages, with the addition of Lithuanian. The experimental investigation proves that our approach is able to achieve results (> 98%) comparable to the previous state-of-the-art, despite being trained less and on fewer data. Our approach is also able to restore diacritics in words not seen during training with > 76% accuracy. Our simultaneous diacritics restoration and typos correction approach reaches > 94% alpha-word accuracy on the 13 languages. It has no direct competitors and strongly outperforms classical spell-checking or dictionary-based approaches. We also demonstrate all the accuracies to further improve with more training. Taken together, this shows the great real-world application potential of our suggested methods to more data, languages, and error classes.

  • 5 authors
·
Jan 31, 2022

What Breaks When LLMs Code? Characterizing Operational Safety Failures of Agentic Code Assistants

Autonomous coding agents built on large language models (LLMs) are rapidly being integrated into development workflows, yet their operational safety properties remain poorly understood beyond evaluations of explicitly malicious inputs. In practice, high-impact failures arise during benign, goal-directed use through environment breakage, fabricated success reports, etc. that current benchmarks do not capture. What categories of operational safety failures actually occur when coding agents are used for everyday development tasks and what is their impact? We present an incident-driven empirical study grounded in two complementary evidence streams. We screen 68,816 papers from 22 premier venues, curating 185 safety-relevant studies, and mine 16,586 GitHub issues from widely deployed LLM-powered coding tools, manually confirming 547 genuine safety failures. Applying systematic open coding over both corpora, we derive a multi-dimensional safety taxonomy of 33 operational risk types organized across seven dimensions, and annotate each incident with contributing factors, task context, severity, and downstream impact. Our findings show that coding-agent failures are often severe, with 326 of 547 incidents rated high or critical. The dominant risks are constraint violations, destructive operations, authorization bypasses, and deception, and over 65% of incidents arise in bug fixing and setup or configuration, patterns largely missing from prior literature. These results have direct implications for SE tool designers and benchmark developers: guardrails must go beyond adversarial-prompt defenses to enforce environmental constraints, failure transparency, and safe-halt behaviors.

  • 2 authors
·
May 28

Can Large Multimodal Models Actively Recognize Faulty Inputs? A Systematic Evaluation Framework of Their Input Scrutiny Ability

Large Multimodal Models (LMMs) have witnessed remarkable growth, showcasing formidable capabilities in handling intricate multimodal tasks with exceptional performance. Recent research has underscored the inclination of large language models to passively accept defective inputs, often resulting in futile reasoning on invalid prompts. However, the same critical question of whether LMMs can actively detect and scrutinize erroneous inputs still remains unexplored. To address this gap, we introduce the Input Scrutiny Ability Evaluation Framework (ISEval), which encompasses seven categories of flawed premises and three evaluation metrics. Our extensive evaluation of ten advanced LMMs has identified key findings. Most models struggle to actively detect flawed textual premises without guidance, which reflects a strong reliance on explicit prompts for premise error identification. Error type affects performance: models excel at identifying logical fallacies but struggle with surface-level linguistic errors and certain conditional flaws. Modality trust varies-Gemini 2.5 pro and Claude Sonnet 4 balance visual and textual info, while aya-vision-8b over-rely on text in conflicts. These insights underscore the urgent need to enhance LMMs' proactive verification of input validity and shed novel insights into mitigating the problem. The code is available at https://github.com/MLGroupJLU/LMM_ISEval.

  • 5 authors
·
Aug 5, 2025 2

On the Limits of LLM Adaptability: Impact of Model-Internalized Priors on Annotation Task Performance

Large Language Models (LLMs) are increasingly used for zero-shot annotation and LLM-as-a-judge tasks, yet their reliability hinges on how model-internalized priors interact with user-provided instructions. We investigate three dimensions of this interaction: (1) how an LLM's familiarity with data and task definitions affects performance, (2) the extent to which additional information in prompts can correct zero-shot errors ("decision stickiness"), and (3) model susceptibility to misaligned task definitions. Through experiments on toxicity detection across diverse datasets (spanning social media, gaming, news, and forums) using both dense and mixture-of-experts models, we find that nearly two-thirds of zero-shot errors are resistant to correction, with an overall rescue rate (fraction of initial errors corrected by prompting) of only 34.8%. High-confidence errors prove especially resistant to correction. When given misaligned definitions, LLMs follow them while maintaining confidence levels unchanged from the aligned condition. Crucially, we introduce Definition-Specific Familiarity (DSF), which measures alignment between a model's internal concept and the task definition. After controlling for dataset-level confounds, DSF shows a positive association with model performance (partial r = +0.41), while three distinct memorization metrics (ROUGE-L, BERTScore, and embedding cosine similarity) all fail to show a positive association. These findings show the limitations of prompt-based correction in annotation tasks, highlighting the importance of definition alignment over text-level memorization.

PyMETA: Evaluating Student Code Diagnosis on and Beyond the First Execution Error

Large language models can diagnose a student program from its code, problem statement, and reference solution. Evaluating this ability requires a clear definition of what counts as the correct diagnosis. We introduce PyMETA, a Python error dataset with 48,646 student submissions to 155 problems. Every submission has a single label for the first execution error identified by an Online Judge, or No Error when the program passes all tests. A targeted subset of 97 submissions also has expert labels collected through iterative repair and re-execution. The taxonomy has three levels; its most detailed level contains 14 labels, including No Error, Logic Error, named Python exceptions, and an Other Errors category. We evaluate two finetuned models and two groups of prompted LLMs: four earlier models and four recent models. When evaluated against the first execution error, the recent prompted models reach 87.5--93.8% macro F1, above the strongest finetuned baseline at 80.6%. This is the opposite of the comparison obtained with the earlier prompted models. On the 97-item expert subset, however, exact-set match is only 43.3--48.5%, although sample F1 is about 79--81%. Output format also matters. On the same 45 audited submissions whose expert label sets do not contain Logic Error, none of the four recent models returns that label under single-error prompting, but 46.7--57.8% of their multi-error outputs include it, usually after an explicit-error label. The results show that model rankings and claims about label bias depend on the meaning of the gold label, the number of labels a model may return, and the scoring rule.

  • 6 authors
·
Sep 14

How Agent Skills Fail under Long Contexts: A White-Box Study in Code Auditing

Agent Skills package procedural instructions and checks for use by general-purpose agents, but loading a skill does not guarantee that every requirement remains active throughout a long tool-using trajectory. We study this problem in a production-derived, white-box code-audit workflow. Holding the task and 24 artifact checks fixed, we vary the surrounding context and classify where failures first become visible: lost requirements, editing drift, failed checking, or non-agent evaluator/runtime failures. Codex with gpt-5.4-mini passes 8/10 runs in a 10,991-character clean context but only 3/10 in both a 299,140-character relevant context and an equal-length irrelevant context. This 50-percentage-point difference is large but remains trend-level under two-sided Fisher tests (p = 0.0698). Requirement coverage nevertheless stays above 92% in both long conditions, showing that a few omissions can invalidate an otherwise complete artifact. A second task passes all clean and long runs, so the evidence does not support a universal context-length threshold. A detailed external checklist passes 10/10 runs, compared with 5/10 for a generic self-check (p = 0.0325). Coding-agent scaffolds may help by selecting a smaller working set, but they do not eliminate failures. We do not introduce context rot or a new general monitoring method; we provide a bounded failure classification and empirical case study for white-box code auditing.

  • 1 authors
·
Jul 19

Beyond Benchmarks: Can Small Open-Weight Models Get Real Coding Work Done?

Making an open-weight language model generate text is not the same as making it finish coding work. This retrospective empirical systems experience report follows my engineering work from July to October 2026 across coding clients, model-specific serving recipes, vLLM, SGLang, llama.cpp, and MLX. Models behind a nominally compatible endpoint still differed in tool-call formatting, turn control, context behavior, cache reuse, and reliability. A served model, coherent text, executed tool calls, and code that passes independent behavioral checks proved to be distinct levels of progress. The evidence includes 4,165 coding-agent gateway turns from my own sessions, 623 debug jobs, and three separate delegation studies with hidden behavioral checks and a fixed Claude Opus 5.5 parent. Saved cold-prefix requests had a median time to first token of about 69.4 s, versus 2.8 s for high-reuse requests (observational). In a 45-trial study, direct coding passed 15/15 and delegation to a quantized Qwen3-Coder 30B-A3B worker passed 12/15 and 14/15. On a second 15-task suite under a fixed artifact contract, the completed replacement cohort had 45/45 for direct coding and 12/45 for a Nemotron Cascade 2 worker (11 worker substitutions). Sixteen sessions stopped by a subscription limit before any model call were replaced; counting them as failures gives 37/45 and 8/45. A Qwen3-Coder worker passed 0/45. In both studies with a direct arm, direct coding had lower API list-price-equivalent frontier cost per verified success. A post-hoc diagnostic found that 19 of 36 artifacts rejected only for renamed parameters passed the hidden checks unchanged; they are not counted as successes. The contribution is an evidence-qualified account of serving, agent protocol, session continuity, artifact validation, and verification in practice. The studies are not pooled and do not rank models or runtimes.

  • 1 authors
·
Oct 24 7

RetryBench: Separating Retry Permission from Recovery Outcomes in GPT Tool Use

False retry permission (FRP) occurs when an agent retries a timed-out tool call without support from the visible contract. RetryBench separates this error from realized recovery. Successful protected recovery does not imply reliable rejection of unsupported retries. In a 720-response key control, valid keys usually enabled safe progress, but offered changed keys elicited unsupported retries in Sol (1/80), Terra (55/80), and Luna (60/80). A later 240-response control removed the candidate and required models to supply a key. Luna still made 8/80 unsupported dispatches; Sol and Terra made none in 80 responses each. Accepting a displayed bad key therefore cannot explain all observed errors, although changed output requirements and conditions preclude a pure display-effect comparison. Removing the weak-contract label increased retries in Sol and Terra's coded-action conditions in a six-endpoint study. A separate 1,536-response canonical-action block also found a positive Terra contrast (115/128 versus 124/128 retries), concentrated in two task/order conditions; the other five contrasts remained unresolved after multiplicity correction. An explicit-contract cue control elicited deferral in all 762 returned responses, leaving recovery unresolved. These results support reporting permission and recovery together across declared interface variants. The evidence covers four purposive tasks, fixed GPT endpoints, and one recovery decision; it identifies neither internal mechanisms nor episode-level safety.

  • 3 authors
·
Oct 10

Safety-Flag: A Unified Benchmark for the Reliability and Calibration of LLM Content Moderators

Large language models are increasingly used for content moderation, but most evaluations still report aggregate accuracy on individual benchmarks. We introduce Safety-Flag, which places seven widely used safety benchmarks (BeaverTails, XSTest, Ethics, WildGuard, Aegis, ToxiChat, and ToxiGen) into a single balanced flag / do-not-flag protocol. We release item-level decisions and confidence scores for six general-purpose LLMs and four dedicated guards, together with three reference models, evaluated on the same items. Safety-Flag measures three dimensions of moderator reliability: error direction, probability calibration, and confidence-based error ranking for human review. They often disagree. Aggregate accuracy does not reveal error direction: one model flags 85% of benign content, whereas another misses 54% of harmful content. All six general-purpose models are overconfident; fitting one temperature per model reduces calibration error by 2.8--6.0times without changing predicted labels or confidence ordering. Confidence-based abstention lowers selective risk for every model, although the gains depend on how well confidence ranks errors. Dedicated guards produce fewer false alarms and are better calibrated, but several have higher miss rates outside their documented coverage. We release the benchmark, fixed item lists, evaluation code, per-item model outputs, and leaderboard at: https://github.com/yibo-hu-lab/safety-flag-benchmark.

  • 1 authors
·
Aug 5

Denoising LM: Pushing the Limits of Error Correction Models for Speech Recognition

Language models (LMs) have long been used to improve results of automatic speech recognition (ASR) systems, but they are unaware of the errors that ASR systems make. Error correction models are designed to fix ASR errors, however, they showed little improvement over traditional LMs mainly due to the lack of supervised training data. In this paper, we present Denoising LM (DLM), which is a scaled error correction model trained with vast amounts of synthetic data, significantly exceeding prior attempts meanwhile achieving new state-of-the-art ASR performance. We use text-to-speech (TTS) systems to synthesize audio, which is fed into an ASR system to produce noisy hypotheses, which are then paired with the original texts to train the DLM. DLM has several key ingredients: (i) up-scaled model and data; (ii) usage of multi-speaker TTS systems; (iii) combination of multiple noise augmentation strategies; and (iv) new decoding techniques. With a Transformer-CTC ASR, DLM achieves 1.5% word error rate (WER) on test-clean and 3.3% WER on test-other on Librispeech, which to our knowledge are the best reported numbers in the setting where no external audio data are used and even match self-supervised methods which use external audio data. Furthermore, a single DLM is applicable to different ASRs, and greatly surpassing the performance of conventional LM based beam-search rescoring. These results indicate that properly investigated error correction models have the potential to replace conventional LMs, holding the key to a new level of accuracy in ASR systems.

  • 6 authors
·
May 24, 2024

Just Ask for a Table: A Thirty-Token User Prompt Defeats Sponsored Recommendations in Twelve LLMs

Wu et al. (2026) showed that most frontier large language models (LLMs) recommend a sponsored, roughly twice-as-expensive flight when their system prompt contains a soft sponsorship cue. We reproduce their evaluation on ten open-weight chat models plus the two of their twenty-three models that are still reachable today (gpt-3.5-turbo, gpt-4o). All reported rates in this paper are produced under the same judge the original paper used (gpt-4o); we additionally store every label under an open-weight (gpt-oss-120b) and a smaller proprietary (gpt-4o-mini) judge for an ablation. Three findings emerge. First, a prose description of an LLM evaluation pipeline is not, on its own, sufficient for accurate reproduction: we surfaced three silent implementation failures that each shifted a reported rate by tens of percentage points. Second, the central claims do generalise - the gpt-3.5-turbo logistic-regression intercept of alpha = 0.81 is within four points of the original alpha = 0.86, and 200 of 200 trials on gpt-3.5-turbo and gpt-4o promote a payday lender to a financially distressed user. Third, a thirty-token user prompt that asks the assistant for a neutral comparison table first cuts sponsored recommendation from 46.9% to 1.0% averaged across our ten open-source models, and from 53.0% to 0% averaged across the two OpenAI models. AI literacy and price-comparison portals are likely market-level mitigations; the harmful-product cell is bounded by neither. Raw data, labels and analysis scripts are at https://github.com/akmaier/Paper-LLM-Ads .

  • 5 authors
·
May 11

What Stops a Small Language Model From Driving a Database Agent

Small open-weight language models are assumed to fail at agentic database work because they lack the reasoning capacity for it. We test that against a production system. Over eleven days we drove the agent mode of an open-source SQL client with 39 open-weight models served locally and one hosted control, across six task surfaces: 8,199 runs, 110,711 ledger events, 14,008 refused tool calls. Of the 2,100 model-attributed agent-mode losses, 1,590, or 75.7%, came from runs that had invoked at least one tool. That majority is what survives resampling models rather than runs: it holds in 99.7% of clustered resamples and in 15 of the 22 models with at least twenty losses. Within it, transport, a run that used the tools and never got a deliverable through, is the largest class at 36.2% and capability, a run that invoked no tool at all, the smallest at 17.3%; we report that ordering as a property of this corpus rather than a general finding, since clustered by model it holds in only 74.5% of resamples. Transport failures decompose into a few mechanical argument shapes. Production ledgers record refusal codes and never the model's arguments, so these were invisible for ten days; capturing them exposed five server defects, one of which demanded a field on one tool, forbade it on the sibling that composed it, then failed the run for its absence. Five server changes, touching no model, prompt or sampling setting, moved six models by 6 to 21 cells out of 30. We also report a confound we believe affects published local-model benchmarks, ours included: with no context cap, one 7.1 GB model was admitted at its full 262,144-token window and held 51 GB on a 64 GB machine, producing runs indistinguishable in any ordinary log from a model timing out. The corpus, the scorer and a verifier that regenerates every figure are released.

  • 4 authors
·
Sep 17

Who Judges the Judges? A Chinese Safety QA Benchmark for Evaluating LLM Responses and Safety Judges

Safety benchmarks for large language models often assess the risk of a user query, although the outcome of question answering depends on whether the response violates a policy. This distinction is critical in Chinese harmful-content evaluation, where linguistic variation and adversarial transformations can obscure risky intent. We introduce C-SafeQA, a policy-grounded benchmark for response-level Chinese safety evaluation. It comprises 538 base queries and 8,877 adversarial queries answered by four full-model LLM deployments, yielding 37,660 query-response records labeled safe, unsafe, or disputed. Reference labels are generated through agreement-aware multi-model adjudication and blind audits of stratified subsets by three safety experts. C-SafeQA supports both evaluation of target-model safety and auditing of seven automated safety judges against shared reference labels. Unsafe-response rates range from 0.93% to 3.35% on base queries and from 11.68% to 30.05% on adversarial queries. On the adversarial subset, judges show substantial trade-offs between unsafe-response recall and risk-query-conditioned safe-response false positive rate, and no judge dominates all metrics. Both acrostic transformations reduce unsafe recall for all seven judges, revealing mechanism-specific evaluator weaknesses. Dataset records, metadata, verification code, and judge scripts are publicly released to support recomputation, while benchmark construction, target-response generation, and private adjudication remain outside the release boundary.

  • 10 authors
·
Aug 31

NL2SQL-BUGs: A Benchmark for Detecting Semantic Errors in NL2SQL Translation

Natural Language to SQL (i.e., NL2SQL) translation is crucial for democratizing database access, but even state-of-the-art models frequently generate semantically incorrect SQL queries, hindering the widespread adoption of these techniques by database vendors. While existing NL2SQL benchmarks primarily focus on correct query translation, we argue that a benchmark dedicated to identifying common errors in NL2SQL translations is equally important, as accurately detecting these errors is a prerequisite for any subsequent correction-whether performed by humans or models. To address this gap, we propose NL2SQL-BUGs, the first benchmark dedicated to detecting and categorizing semantic errors in NL2SQL translation. NL2SQL-BUGs adopts a two-level taxonomy to systematically classify semantic errors, covering 9 main categories and 31 subcategories. The benchmark consists of 2,018 expert-annotated instances, each containing a natural language query, database schema, and SQL query, with detailed error annotations for semantically incorrect queries. Through comprehensive experiments, we demonstrate that current large language models exhibit significant limitations in semantic error detection, achieving an average detection accuracy of 75.16%. Specifically, our method successfully detected 106 errors (accounting for 6.91%) in BIRD, a widely-used NL2SQL dataset, which were previously undetected annotation errors. This highlights the importance of semantic error detection in NL2SQL systems. The benchmark is publicly available at https://nl2sql-bugs.github.io/.

  • 5 authors
·
Jun 14, 2025

Test Coverage Analysis of Agentic Pull Requests

AI coding agents increasingly submit complete pull requests (PRs) with minimal human intervention, shifting software development from AI-assisted to autonomous workflows. As these agents become more prevalent, ensuring the code they generate is adequately tested, by existing tests or by tests the agents write, is critical to preventing regressions, yet little is known about testing in agentic PRs. To address this gap, we analyze 4882 agent-generated PRs from the AIDev dataset (532 Java and 4350 Python PRs) produced by five coding agents. We study (i) how often agents include test changes and (ii) how well covered are code changes by existing and agent-written tests. Agents include test changes in only 49.6% of PRs that change code under test files. Existing tests provide an incomplete safety net: they cover 61.5% of agents' changed executable lines in Java and only 27.0% in Python, where 64.8% of PRs have no changed line executed by any existing test. Agent-written tests improve coverage over existing tests, but only in a minority of PRs: 35.9% of Java and 22.5% of Python Code + Tests PRs show a coverage gain. Across both languages, error-handling constructs (e.g., try and catch blocks) are the most consistently under-tested, with miss rates reaching 86.0% in Java and 81.0% in Python. These findings motivate coverage-aware development practices, coverage feedback loops for coding agents, and evaluation benchmarks that measure test quality to better help agents reliably test their own code.

  • 4 authors
·
Jul 19