Datasets:

allenai
/

novelty-judge-bench

id int32 0 321	iclr_area stringclasses 20 values	idea stringlengths 417 2.1k	label stringclasses 2 values	idea_source stringclasses 2 values	title stringlengths 18 135	forum_id stringlengths 10 10 ⌀	rating float64 6 8 ⌀	contribution float64 3 3.75 ⌀	positive_signals listlengths 0 9	negative_signals listlengths 0 0
0	foundation or frontier models, including LLMs	Aligning Large Language Models (LLMs) with human preferences is crucial for safe and effective AI interactions. While popular methods like Direct Preference Optimization (DPO) have simplified alignment, they remain sensitive to data noise and overlook the differential importance of individual tokens. Existing token-lev...	POSITIVE	human	Token-Importance Guided Direct Preference Optimization	cMEnMVvMw9	6.5	3	[ "The paper introduces a novel Token-Importance Guided Direct Preference Optimization (TI-DPO) framework that integrates gradient-based token attribution with a triplet loss objective, aiming to achieve finer-grained preference alignment.", "Introduces a new approach that weights token importance and incorporates ...	[]
1	foundation or frontier models, including LLMs	Foundation models exhibit remarkable capabilities but remain static after pretraining, unable to efficiently incorporate emerging knowledge without expensive retraining or catastrophic forgetting. We introduce Temporal Knowledge Anchoring (TKA), a framework that enables continual knowledge integration in large language...	NEGATIVE	generated	LLM-foundation or frontier models, including LLMs-11	null	null	null	[]	[]
10	foundation or frontier models, including LLMs	Next token prediction has been the standard training objective used in large language model pretraining. Representations are learned as a result of optimizing for token-level perplexity. We propose Continuous Concept Mixing (CoCoMix), a novel pretraining framework that combines discrete next token prediction with conti...	POSITIVE	human	LLM Pretraining with Continuous Concepts	wTGcb3DxOn	7.33	3.33	[ "Great pre-training analysis with a novel architecture.", "The proposal of combining next token prediction with continuous concepts in the pretraining paradigm is novel.", "This idea of integrating an interpretability mechanism (concept) into pretraining frameworks through SAE comes with significant originality...	[]
100	applications to computer vision, audio, language, and other modalities	Robust 3D hand reconstruction is challenging in egocentric vision due to depth ambiguity, self-occlusion, and complex hand-object interactions. Prior works attempt to mitigate the challenges by scaling up training data or incorporating auxiliary cues, often falling short of effectively handling unseen contexts. In this...	POSITIVE	human	EgoHandICL: Egocentric 3D Hand Reconstruction with In-Context Learning	nwjy9BeorI	6	3.25	[ "The retrieval and in-context learning are novel and sound approaches.", "The application of in-context learning to 3D hand reconstruction is highly novel.", "Using a VLM for semantic retrieval (e.g., finding similar interactions or occlusion types ) rather than just visual similarity is a powerful and unique i...	[]
101	applications to computer vision, audio, language, and other modalities	Multi-modal learning has achieved remarkable success in aligning representations across vision, audio, and language, yet existing approaches predominantly learn static correspondences while overlooking the temporal dynamics inherent in multi-modal data streams. We introduce Temporal Cross-Modal Synchronization (TCMS), ...	NEGATIVE	generated	LLM-applications to computer vision, audio, language, and other modalities-11	null	null	null	[]	[]
102	applications to computer vision, audio, language, and other modalities	Physical simulation relies on spatially-varying mechanical properties, typically laboriously hand-crafted. We present the first feed-forward model to predict fine-grained mechanical properties, Young's modulus ($E$), Poisson's ratio ($\nu$), and density ($\rho$), throughout the volume of 3D objects. Our model support...	POSITIVE	human	VoMP: Predicting Volumetric Mechanical Property Fields	aTP1IM6alo	7	3	[ "The introduction of MatVAE- a VAE used to learn a low-dimensional latent space of valid real-world material triplets- is a significant contribution.", "The core design is novel and effective.", "Proposes a framework for predicting voxel-level material properties from arbitrary 3D representations, defining a no...	[]
103	applications to computer vision, audio, language, and other modalities	Cross-modal understanding typically emphasizes alignment and retrieval, yet many applications require generating detailed content in one modality from another—such as synthesizing realistic audio from silent video or hallucinating plausible visual dynamics from sound alone. We introduce Cross-Modal Causal Synthesis, a ...	NEGATIVE	generated	LLM-applications to computer vision, audio, language, and other modalities-7	null	null	null	[]	[]
104	applications to computer vision, audio, language, and other modalities	Multimodal learning typically relies on correlation-based fusion strategies that combine information from vision, audio, and language without explicitly modeling their causal interdependencies. We introduce Causal Multimodal Representations (CMR), a framework that learns directional influence relationships between moda...	NEGATIVE	generated	LLM-applications to computer vision, audio, language, and other modalities-9	null	null	null	[]	[]
105	applications to computer vision, audio, language, and other modalities	Video diffusion models (DMs) have enabled high-quality video synthesis. Yet, their substantial computational and memory demands pose serious challenges to real-world deployment, even on high-end GPUs. As a commonly adopted solution, quantization has proven notable success in reducing cost for image DMs, while its direc...	POSITIVE	human	QVGen: Pushing the Limit of Quantized Video Generative Models	XJXZXuTj11	6.8	3	[ "The first full QAT method for video generation models I'm aware of.", "The authors observe that the auxiliary module $\\Phi$ exhibits rank decay during quantization-aware training and effectively leverage this property through their rank-decay mechanism. This insight is both intuitive and novel.", "This paper ...	[]
106	optimization	Modern neural network optimization treats parameter spaces as Euclidean, ignoring the rich geometric structure of loss landscapes and architectural symmetries. We introduce Adaptive Riemannian Manifold Optimization (ARMO), a framework that learns and exploits low-dimensional Riemannian manifolds embedded within high-di...	NEGATIVE	generated	LLM-optimization-6	null	null	null	[]	[]
107	optimization	Regret matching (RM)---and its modern variants---is a foundational online algorithm that has been at the heart of many AI breakthrough results in solving benchmark zero-sum games, such as poker. Yet, surprisingly little is known so far in theory about its convergence beyond two-player zero-sum games. For example, wheth...	POSITIVE	human	Convergence of Regret Matching in Potential Games and Constrained Optimization	EOV1q1U23N	7	3	[ "The paper provides the first general-purpose guarantees for RM+ beyond zero-sum settings.", "this is a nice work that offers new characterization of the last-iterate convergence properties of regret matching variants beyond zero-sum games", "Establishing the first polynomial-time convergence guarantee for alte...	[]
108	optimization	We introduce Neural Preconditioned Optimization, a framework that learns adaptive preconditioners for iterative optimization algorithms while preserving convergence guarantees. Classical preconditioning techniques for gradient descent and quasi-Newton methods require extensive domain knowledge and manual parameter tuni...	NEGATIVE	generated	LLM-optimization-1	null	null	null	[]	[]
109	optimization	The minimum-cost bipartite matching between two sets of points $R$ and $S$ in a metric space has a wide range of applications in machine learning, computer vision, and logistics. For instance, it can be used to estimate the $1$-Wasserstein distance between continuous probability distributions and for efficiently matchi...	POSITIVE	human	Efficient algorithms for Incremental Metric Bipartite Matching	wnIanx0r0w	7	3.25	[ "builds upon interesting known ideas in a novel way", "The introduction of the \"push-relable\" subroutine instead of augmenting paths, as well as the careful analysis of the \"Find-Admissable-Edge\" subroutine allow a clean generalization of the static algorithm.", "The use of push–relabel techniques to design...	[]
11	foundation or frontier models, including LLMs	We introduce Neurosymbolic Scaffolding, a framework that augments foundation models with the ability to dynamically construct and reason over explicit symbolic structures during inference. While large language models excel at pattern recognition and generation, they struggle with tasks requiring verifiable logical reas...	NEGATIVE	generated	LLM-foundation or frontier models, including LLMs-4	null	null	null	[]	[]
110	optimization	Modern deep learning optimizers face a fundamental trade-off: adaptive methods like Adam excel at handling ill-conditioned landscapes but may converge to sharp minima with poor generalization, while sharpness-aware minimization (SAM) seeks flat minima but uses fixed perturbation scales insensitive to local geometry. We...	NEGATIVE	generated	LLM-optimization-0	null	null	null	[]	[]
111	optimization	Dynamically configuring algorithm hyperparameters is a fundamental challenge in computational intelligence. While learning-based methods offer automation, they suffer from prohibitive sample complexity and poor generalization. We introduce AutoEP, a novel framework that bypasses training entirely by leveraging Large La...	POSITIVE	human	AutoEP: LLMs-Driven Automation of Hyperparameter Evolution for Metaheuristic Algorithms	hit3hGBheP	6.5	3	[ "I have to acknowledge this paper as a novel perspective, with reasonable usage of LLMs to relieve human efforts required for tuning optimization algorithms.", "Introducing an ELA-grounded, LLM-reasoning framework for online parameter control is conceptually new and methodologically sound.", "The method of usin...	[]
112	optimization	We introduce Temporal Landscape Memory Networks (TLMNs), a framework for optimization in non-stationary environments where objective functions evolve over time. Unlike classical optimizers that treat each iteration independently, TLMNs maintain a learned representation of historical loss landscapes to predict future ge...	NEGATIVE	generated	LLM-optimization-4	null	null	null	[]	[]
113	optimization	This paper studies the complexity of finding an $\epsilon$-stationary point for stochastic bilevel optimization when the upper-level problem is nonconvex and the lower-level problem is strongly convex. In this work, we show that faster rates are achievable for higher-order smooth problems. We first reformulate F$^2$SA ...	POSITIVE	human	Faster Gradient Methods for Highly-smooth Stochastic Bilevel Optimization	XIAta0WOJ6	6	3	[ "The $\\textsf{F}^2\\textsf{SA}$ method was originally proposed as a penalty-based approach. In this paper, the authors offer a novel interpretation by viewing it as a forward difference approximation of the hypergradient.", "The finite-difference modification of F2SA is both novel and conceptually elegant, offer...	[]
114	optimization	Federated Learning (FL) enables clients to collaboratively train a global model without sharing their private data. However, the presence of malicious (Byzantine) clients poses significant challenges to the robustness of FL, particularly when data distributions across clients are heterogeneous. In this paper, we propos...	POSITIVE	human	Byzantine-Robust Federated Learning with Learnable Aggregation Weights	lXSrulux48	6	3.25	[ "The paper presents a genuinely novel approach to Byzantine-robust federated learning by treating aggregation weights as learnable parameters.", "This isn't just a minor tweak to existing methods—it represents a meaningful shift in how we approach the aggregation problem.", "The method seems novel and provides ...	[]
115	optimization	Contemporary learned optimizers struggle to generalize across diverse loss landscapes and training horizons due to their reliance on local gradient information and short-term meta-training objectives. We introduce GeometricFlow, a framework that learns to optimize by modeling the intrinsic Riemannian geometry of parame...	NEGATIVE	generated	LLM-optimization-3	null	null	null	[]	[]
116	optimization	Score matching has become a central training objective in modern generative modeling, particularly in diffusion models, where it is used to learn high-dimensional data distributions through the estimation of score functions. Despite its empirical success, the theoretical understanding of the optimization behavior of sc...	POSITIVE	human	Convergence Dynamics of Over-Parameterized Score Matching for a Single Gaussian	VDIH6L8ANo	6.67	3	[ "Their analysis of the gradient dynamics in the small noise regime requires novel techniques.", "Even if the chosen true data distribution is really simple, the authors highlight a complex (and technical) analysis which has not been considered in the literature." ]	[]
117	optimization	Modern neural network optimizers treat each update in isolation, discarding rich geometric information accumulated along the optimization trajectory. We introduce Trajectory-Aware Optimization (TAO), a framework that maintains an explicit memory of the path traversed through the loss landscape and leverages this histor...	NEGATIVE	generated	LLM-optimization-5	null	null	null	[]	[]
118	optimization	Computing the polar decomposition and the related matrix sign function has been a well-studied problem in numerical analysis for decades. Recently, it has emerged as an important subroutine within the Muon algorithm for training deep neural networks. However, the requirements of this application differ sharply from cla...	POSITIVE	human	The Polar Express: Optimal Matrix Sign Methods and their Application to the Muon Algorithm	yRtgZ1K8hO	8	3.25	[ "A novel method is proposed and analyzed mathematically.", "I found this result very interesting and new.", "I feel that the idea is quite novel and can be useful beyond the Muon use case.", "the presented algorithm seems novel and interesting" ]	[]
119	optimization	Modern deep learning optimizers rely on element-wise adaptive learning rates that ignore the geometric structure of neural network loss landscapes. While second-order methods exploit curvature information for faster convergence, their computational cost remains prohibitive for large-scale models. We introduce Trajector...	NEGATIVE	generated	LLM-optimization-2	null	null	null	[]	[]
12	foundation or frontier models, including LLMs	While Large Language Models (LLMs) hold promise to become autonomous agents, they often explore suboptimally in sequential decision-making. Recent work has sought to enhance this capability via supervised fine-tuning (SFT) or reinforcement learning (RL), improving regret on the classic multi-armed bandit task. However,...	POSITIVE	human	When Greedy Wins: Emergent Exploitation Bias in Meta-Bandit LLM Training	ZRO79e2BHI	6	3	[ "their proposed reward designs are novel, from what I can tell", "This paper introduced token-level reward derived from GAE, which improved upon prior work." ]	[]
120	datasets and benchmarks	Static benchmarks inadequately reflect the temporal dynamics inherent in real-world machine learning deployments, where data distributions evolve continuously due to seasonal variations, cultural trends, and shifting user behaviors. We introduce TemporalShift, a benchmark framework designed to evaluate model robustness...	NEGATIVE	generated	LLM-datasets and benchmarks-16	null	null	null	[]	[]
121	datasets and benchmarks	We introduce spatio-temporal omni-object video grounding, dubbed $\textbf{OmniSTVG}$, a new STVG task aiming to localize spatially and temporally all targets mentioned in the textual query within videos. Compared to classic STVG locating only a single target, OmniSTVG enables localization of not only an arbitrary numbe...	POSITIVE	human	OmniSTVG: Toward Spatio-Temporal Omni-Object Video Grounding	azcQJtcYTE	6.67	3	[ "Pioneering a New Direction: The work identifies the critical limitation of existing STVG tasks (i.e., single-target grounding) and defines a more complex and practical new research direction (i.e., multi-target grounding).", "Contribution of a Core Resource (BOSTVG): It provides the field's first large-scale (10...	[]
122	datasets and benchmarks	Embodied AI has developed rapidly in recent years, but it is still mainly deployed in laboratories, with various distortions in the Real-world limiting its application. Traditionally, Image Quality Assessment (IQA) methods are applied to predict human preferences for distorted images; however, there is no IQA method to...	POSITIVE	human	Image Quality Assessment for Embodied AI	azj53PLJRL	7	3.75	[ "It defines the image quality assessment (IQA) problem in embodied intelligence as \"image usability for robots,\" transcending traditional frameworks based on human or machine vision systems (HVS/MVS).", "It not only provides a novel benchmark dataset but also successfully establishes critical empirical links be...	[]
123	datasets and benchmarks	Machine learning benchmarks traditionally assume static data distributions, yet real-world deployments face continuous temporal shifts that degrade model performance unpredictably. We introduce Temporal Shift Decomposition Benchmarks, a new paradigm for dataset construction that explicitly disentangles multiple sources...	NEGATIVE	generated	LLM-datasets and benchmarks-13	null	null	null	[]	[]
124	datasets and benchmarks	Existing machine learning benchmarks typically aggregate human annotations into single ground-truth labels, discarding valuable information about disagreement and treating annotator variance as noise. However, many real-world tasks involve inherent subjectivity where multiple valid interpretations coexist—from content ...	NEGATIVE	generated	LLM-datasets and benchmarks-11	null	null	null	[]	[]
125	datasets and benchmarks	Precise, object-aware control over visual content is essential for advanced image editing and compositional generation. Yet, most existing approaches operate on entire images holistically, limiting the ability to isolate and manipulate individual scene elements. In contrast, layered representations, where scenes are ex...	POSITIVE	human	Referring Layer Decomposition	AlgRVfd1z7	6.5	3	[ "the paper proposes a large-scale, real-world dataset for an interesting task", "a paper that defines a new task and proposes a new dataset", "This paper introduces Referring Layer Decomposition (RLD), the pioneering task that explores layer decomposition guided by multi-modal referring inputs.", "This paper ...	[]
126	datasets and benchmarks	Modern machine learning benchmarks primarily assess correlational pattern matching on static test distributions, offering limited insight into whether models capture genuine causal mechanisms or exploit spurious features. We introduce Causal Intervention Benchmarks (CIB), a framework for constructing hybrid evaluation ...	NEGATIVE	generated	LLM-datasets and benchmarks-8	null	null	null	[]	[]
127	datasets and benchmarks	Large language models (LLMs) have demonstrated remarkable performance on single-turn text-to-SQL tasks, but real-world database applications predominantly require multi-turn interactions to handle ambiguous queries, execution errors, and evolving user requirements. Existing multi-turn benchmarks fall short of capturing...	POSITIVE	human	BIRD-INTERACT: Re-imagining Text-to-SQL Evaluation via Lens of Dynamic Interactions	nHrYBGujps	7.5	3.25	[ "The combination of CRUD support, HKB reasoning, and dynamic interaction is unique.", "The paper addresses an important gap between single-turn text-to-SQL evaluation and real interactive database usage, providing a benchmark with clear practical relevance.", "the paper makes text-to-SQL evaluation much more re...	[]
128	datasets and benchmarks	Approximately 445.2 million individuals worldwide are living with traumatic amputations, and an estimated 31.64 million children aged 0–14 have congenital limb differences, yet they remain largely underrepresented in human pose estimation (HPE) research. Accurate HPE could significantly benefit this population in appli...	POSITIVE	human	InclusiveVidPose: Bridging the Pose Estimation Gap for Individuals with Limb Deficiencies in Video-Based Motion	SyQqXAdWUq	7.33	3.67	[ "the paper addresses a clear gap in research", "The paper presents a dataset of approximately 327,000 video frames featuring various types of limb deficiencies. It also introduces a new metric, called Limb-specific Confidence Consistency (LiCC), designed to help evaluate whether pose-estimation models can tell th...	[]
129	datasets and benchmarks	Modern benchmarks face a fundamental tension: static test sets enable contamination and eventual saturation, while human-generated alternatives are costly and slow to produce. We introduce ProceduralBench, a framework for generating procedurally-synthesized evaluation instances with cryptographic guarantees of novelt...	NEGATIVE	generated	LLM-datasets and benchmarks-6	null	null	null	[]	[]
13	foundation or frontier models, including LLMs	Foundation models demonstrate impressive capabilities yet lack interpretable, compositional reasoning structures. We propose Modular Reasoning Primitives (MRP), a framework for training large language models to develop explicit, reusable reasoning components that can be flexibly composed across diverse cognitive tasks....	NEGATIVE	generated	LLM-foundation or frontier models, including LLMs-2	null	null	null	[]	[]
130	datasets and benchmarks	Modern machine learning benchmarks primarily assess pattern recognition capabilities but fail to rigorously evaluate compositional generalization—the ability to understand and recombine primitive concepts in novel configurations. We introduce CompoGen, a dynamic benchmark framework that systematically tests whether mod...	NEGATIVE	generated	LLM-datasets and benchmarks-9	null	null	null	[]	[]
131	datasets and benchmarks	Online services rely on CAPTCHAs as a first line of defense against automated abuse, yet recent advances in multi-modal large language models (MLLMs) have eroded the effectiveness of conventional designs that focus on text recognition or 2D image understanding. To address this challenge, we present Spatial CAPTCHA,...	POSITIVE	human	Spatial CAPTCHA: Generatively Benchmarking Spatial Reasoning for Human-Machine Differentiation	8LAc5kzjhz	6	3	[ "The paper addresses a timely and critical real-world problem. With conventional CAPTCHAs becoming increasingly vulnerable to AI, the proposed method of targeting a well-established weakness in MLLMs—spatial reasoning—is an insightful and promising direction for the next generation of human-verification systems.", ...	[]
132	datasets and benchmarks	Open-ended segmentation requires models capable of generating free-form descriptions of previously unseen concepts and regions. Despite advancements in model development, current evaluation protocols for open-ended segmentation tasks fail to capture the true semantic accuracy of the generated descriptions. We empirical...	POSITIVE	human	Benchmarking Open-ended Segmentation	QSmwRnw8GP	6	3	[ "The paper tackles a highly relevant and underexplored problem—how to objectively evaluate open-ended segmentation models where outputs are free-form text rather than fixed labels.", "The proposed Lexical Alignment Curve (LAC) introduces a novel evaluation perspective that bridges semantic granularity levels (exa...	[]
133	datasets and benchmarks	Machine learning benchmarks traditionally consist of fixed datasets that quickly saturate as models improve, providing diminishing insight into capability gaps and progress. We introduce Procedural Evaluation Graphs (PEGs), a framework for generating infinite families of benchmark tasks with controllable difficulty and...	NEGATIVE	generated	LLM-datasets and benchmarks-15	null	null	null	[]	[]
134	datasets and benchmarks	Evaluating tabular generators remains a challenging problem, as the unique causal structural prior of heterogeneous tabular data does not lend itself to intuitive human inspection. Recent work has introduced structural fidelity as a tabular-specific evaluation dimension to assess whether synthetic data complies with th...	POSITIVE	human	TabStruct: Measuring Structural Fidelity of Tabular Data	XOPH34Extq	7	3.5	[ "The paper introduces a new metric that complements the assessment of data generators on real-world tasks", "It identifies a blind spot (structural fidelity) in the current evaluation metrics and proposes new metrics to measure it.", "the proposed metric is novel (global utility) and interesting.", "The motiv...	[]
135	datasets and benchmarks	Existing benchmarks evaluate model performance on static test sets, failing to capture the temporal dynamics and compositional shifts inherent in real-world deployment scenarios. We introduce ChronoShift, a benchmark framework designed to systematically evaluate how models adapt to evolving compositional relationships ...	NEGATIVE	generated	LLM-datasets and benchmarks-7	null	null	null	[]	[]
136	datasets and benchmarks	Machine learning benchmarks typically consist of fixed datasets that quickly saturate as models improve, providing diminishing insight into true generalization capabilities. We introduce a framework for procedurally generated compositional benchmarks that dynamically synthesize evaluation instances by systematically re...	NEGATIVE	generated	LLM-datasets and benchmarks-0	null	null	null	[]	[]
137	datasets and benchmarks	Recent advancements in multimodal slow-thinking systems have demonstrated remarkable performance across various visual reasoning tasks. However, their capabilities in text-rich image reasoning tasks remain understudied due to the absence of a dedicated and systematic benchmark. To address this gap, we propose OCR-Reaso...	POSITIVE	human	OCR-Reasoning Benchmark: Unveiling the True Capabilities of MLLMs in Complex Text-Rich Image Reasoning	aH7eyx64pC	6.5	3.25	[ "OCR-Reasoning is the first benchmark to systematically assess reasoning processes in text-rich image scenarios, addressing a long-overlooked need.", "The paper demonstrates strong originality by introducing a new benchmark, OCR-Reasoning, that evaluates multimodal large language models on text-rich image reasoni...	[]
138	datasets and benchmarks	Modern machine learning benchmarks suffer from rapid saturation and limited coverage of compositional reasoning capabilities across modalities. We introduce CompGen, a generative benchmark framework that synthesizes evaluation instances through systematic recombination of atomic visual, linguistic, and logical primitiv...	NEGATIVE	generated	LLM-datasets and benchmarks-4	null	null	null	[]	[]
139	datasets and benchmarks	We introduce Gaia2, a benchmark for evaluating large language model agents in realistic, asynchronous environments. Unlike prior static or synchronous evaluations, Gaia2 introduces scenarios where environments evolve independently of agent actions, requiring agents to operate under temporal constraints, adapt to no...	POSITIVE	human	Gaia2: Benchmarking LLM Agents on Dynamic and Asynchronous Environments	9gw03JpKK4	8	3.33	[ "A key innovation is the shift from static to dynamic, asynchronous environments.", "introducing Gaia2, the first benchmark unifying asynchronous execution, temporal reasoning, noise robustness, ambiguity resolution, and multi-agent collaboration under a verifiable evaluation framework directly usable for RLVR" ]	[]
14	foundation or frontier models, including LLMs	We introduce Reflexive Reasoning Networks, a framework that endows large language models with internalized verification capabilities for multi-step reasoning tasks. Rather than relying on external verifiers or outcome-based supervision, our approach trains models to simultaneously generate reasoning traces and evaluate...	NEGATIVE	generated	LLM-foundation or frontier models, including LLMs-17	null	null	null	[]	[]
140	datasets and benchmarks	Current machine learning benchmarks provide static snapshots of model performance, leading to rapid saturation and encouraging overfitting to specific test distributions. We introduce the concept of adaptive co-evolutionary benchmarks that dynamically adjust their difficulty and composition in response to the improving...	NEGATIVE	generated	LLM-datasets and benchmarks-10	null	null	null	[]	[]
141	datasets and benchmarks	Morphological traits are physical characteristics of biological organisms that provide vital clues on how organisms interact with their environment. Yet extracting these traits remains a slow, expert-driven process, limiting their use in large-scale ecological studies. A major bottleneck is the absence of high-quality ...	POSITIVE	human	Automatic Image-Level Morphological Trait Annotation for Organismal Images	oFRbiaib5Q	6.67	3.33	[ "The combination of SAE and MLLM is a new approach to auto-generating trait captions for biological images.", "BIOSCAN-TRAITS (80k traits / 19k images) fills a gap in large-scale, trait-level supervision for biodiversity ML." ]	[]
142	datasets and benchmarks	The rapid progress of AI and robotics may profoundly transform society, as humans and robots begin to coexist in shared communities, bringing both opportunities and challenges. To explore this future, we present Virtual Community—an open-world platform for humans, robots, and society—built on a universal physics engine...	POSITIVE	human	Virtual Community: An Open World for Humans, Robots, and Society	Qo0OZZoTLh	7.33	3.67	[ "A rare integration of city-scale environments, human–robot co-presence, and social graphs with time-aligned schedules, enabling questions that neither indoor-only simulators nor purely social agents can address.", "I overall find the simulator quite novel and believe it provides a useful platform for studying em...	[]
143	datasets and benchmarks	Current machine learning benchmarks evaluate tasks in isolation, measuring performance on individual datasets without considering how models handle multiple objectives simultaneously or compose knowledge across domains. We propose Compositional Multi-Task Benchmarks (CMTB), a new evaluation paradigm that explicitly mea...	NEGATIVE	generated	LLM-datasets and benchmarks-14	null	null	null	[]	[]
144	datasets and benchmarks	Medical question answering (QA) benchmarks often focus on multiple-choice or fact-based tasks, leaving open-ended answers to real patient questions underexplored. This gap is particularly critical in mental health, where patient questions often mix symptoms, treatment concerns, and emotional needs, requiring answers th...	POSITIVE	human	CounselBench: A Large-Scale Expert Evaluation and Adversarial Benchmarking of Large Language Models in Mental Health Question Answering	8MBYRZHVWT	6.67	3.33	[ "The second component, COUNSELBENCH-ADV, is highly innovative.", "represents one of the largest expert-annotated datasets for open-ended mental health QA" ]	[]
145	datasets and benchmarks	Machine learning benchmarks traditionally consist of static datasets that quickly become saturated as models improve, limiting their utility for distinguishing state-of-the-art systems and identifying fundamental capability gaps. We propose Adaptive Adversarial Benchmarks (AAB), a framework that dynamically generates e...	NEGATIVE	generated	LLM-datasets and benchmarks-1	null	null	null	[]	[]
146	datasets and benchmarks	Modern benchmarks suffer from rapid saturation and hidden shortcuts that allow models to achieve high scores without genuine understanding. We introduce Adaptive Adversarial Benchmarks, a framework where evaluation datasets co-evolve with model capabilities through continuous adversarial example discovery. Our approach...	NEGATIVE	generated	LLM-datasets and benchmarks-12	null	null	null	[]	[]
147	datasets and benchmarks	Fingerspelling plays a vital role in sign languages, particularly for conveying names, technical terms, and words not found in the standard lexicon. However, evaluation of two-handed fingerspelling detection and recognition is rarely addressed in existing sign language datasets—particularly for BANZSL (British, Austral...	POSITIVE	human	BANZ-FS: BANZSL Fingerspelling Dataset	GMR9BUsPbq	7	3.5	[ "The proposed BANZ-FS dataset addresses a tangible linguistic and practical gap by focusing on two-handed BANZSL fingerspelling, a setting largely underrepresented in current literature.", "This is the first large-scale fingerspelling dataset specifically for the two-handed BANZSL system. It fills a critical gap,...	[]
148	datasets and benchmarks	Modern machine learning benchmarks provide static snapshots of tasks and data distributions, failing to capture the temporal dynamics that characterize real-world deployment scenarios. We introduce the concept of Living Benchmarks: continuously evolving evaluation frameworks that automatically ingest, curate, and incor...	NEGATIVE	generated	LLM-datasets and benchmarks-5	null	null	null	[]	[]
149	datasets and benchmarks	The rapid development and widespread adoption of Audio Large Language Models (ALLMs) require a rigorous assessment of their trustworthiness. However, existing evaluation frameworks, primarily designed for text, are not equipped to handle the unique vulnerabilities introduced by audio's acoustic properties. We find that...	POSITIVE	human	AudioTrust: Benchmarking The Multifaceted Trustworthiness of Audio Large Language Models	E823AY0taq	6.8	3.6	[ "The proposed framework addresses a significant gap, namely, the need for audio-specific trustworthiness evaluation of ALLMs.", "the proposed evaluation strategies (e.g., construction of test data designed to probe specific vulnerabilities) are unique and valuable to future work.", "proposing the largest and mo...	[]
15	foundation or frontier models, including LLMs	Vision-Language Models (VLMs) enable powerful multi-agent systems, but scaling them is economically unsustainable: coordinating heterogeneous agents under information asymmetry often spirals costs. Existing paradigms, such as Mixture-of-Agents and knowledge-based routers, rely on heuristic proxies that ignore costs and...	POSITIVE	human	Why Keep Your Doubts to Yourself? Trading Visual Uncertainties among Vision-Language Models	zeqCjGQB4U	6.5	3	[ "The core concept of a decentralized market for uncertainty is a powerful and novel metaphor that drives the entire paper.", "I appreciate the idea of treating uncertainty as a tradable commodity, which is novel seems to me.", "The paper's main strength is its core idea: \"market-based uncertainty trading.\" Th...	[]
150	datasets and benchmarks	We introduce the concept of Living Benchmarks, a new paradigm for dataset curation and model evaluation that addresses the temporal rigidity of traditional static benchmarks. Unlike conventional benchmarks that freeze data and tasks at publication time, Living Benchmarks continuously evolve by incorporating fresh data ...	NEGATIVE	generated	LLM-datasets and benchmarks-3	null	null	null	[]	[]
151	datasets and benchmarks	Although Multimodal Large Language Models (MLLMs) have demonstrated proficiency in video captioning, practical applications require captions that follow specific user instructions rather than generating exhaustive, unconstrained descriptions. Current benchmarks, however, primarily assess descriptive comprehensiveness w...	POSITIVE	human	IF-VidCap: Can Video Caption Models Follow Instructions?	lBXJexaC8a	6	3.25	[ "IF-VidCap is the first benchmark to explicitly evaluate instruction-following in video captioning (27 constraint types), addressing a critical gap beyond traditional accuracy/fluency metrics.", "creation of IF-VidCap, the first benchmark systematically evaluating instruction-following in video captioning with co...	[]
152	datasets and benchmarks	Large Language Models (LLMs) are conversational interfaces. As such, LLMs have the potential to assist their users not only when they can fully specify the task at hand, but also to help them define, explore, and refine what they need through multi-turn conversational exchange. Although analysis of LLM conversation log...	POSITIVE	human	LLMs Get Lost In Multi-Turn Conversation	VKGTGGcwl6	8	3.5	[ "The paper makes an important conceptual contribution by decomposing overall performance degradation into two distinct components: aptitude (best-case capability, measured as 90th percentile performance, A90) and unreliability (variance across runs, measured as the 90-10 interpercentile range, U90/10).", "This ex...	[]
153	datasets and benchmarks	We introduce ProGen-Bench, a procedurally generated benchmark framework that addresses the rapid saturation and limited diagnostic value of static evaluation datasets. Rather than fixed test sets, ProGen-Bench employs parameterized generative programs to create infinite task variations across vision, language, and reas...	NEGATIVE	generated	LLM-datasets and benchmarks-2	null	null	null	[]	[]
154	generative models	Existing generative models struggle to satisfy multiple compositional constraints simultaneously, often producing outputs that inadequately balance competing objectives or fail to respect fine-grained user specifications. We introduce Compositional Constraint Diffusion (CCD), a framework that decomposes the generation ...	NEGATIVE	generated	LLM-generative models-14	null	null	null	[]	[]
155	generative models	The importance of mobile phone GPS trajectory data is widely recognized across many fields, yet the use of real data is often hindered by privacy concerns, limited accessibility, and high acquisition costs. As a result, generating pseudo–GPS trajectory data has become an active area of research. Recent diffusion-based ...	POSITIVE	human	TrajFlow: Nation-wide Pseudo GPS Trajectory Generation with Flow Matching Models	BDOldEjwCE	6.5	3.25	[ "The paper presents the first flow-matching-based generative framework.", "This paper presents the first application of the flow-matching paradigm to the task of GPS trajectory generation, which addresses the problem that the performance of existing models, especially diffusion models, degrades dramatically when ...	[]
156	generative models	We present a novel geometric perspective on the latent space of diffusion models. We first show that the standard pullback approach, utilizing the deterministic probability flow ODE decoder, is fundamentally flawed. It provably forces geodesics to decode as straight segments in data space, effectively ignoring any intr...	POSITIVE	human	The Spacetime of Diffusion Models: An Information Geometry Perspective	qCsbYJZRA5	6.5	3	[ "Defining geometry on the spacetime manifold (x_t, t) using the Fisher–Rao metric is both conceptually straightforward and surprisingly underexplored, making the contribution feel natural yet novel.", "The work introduces a novel and interesting notion of latent spacetime.", "The second application (transition ...	[]
157	generative models	Inverse problems in scientific imaging—from medical tomography to computational microscopy—are fundamentally ill-posed, requiring strong priors to resolve ambiguities and suppress noise. Classical regularization methods impose hand-crafted priors that often fail to capture the complexity of natural structures, while pu...	NEGATIVE	generated	LLM-generative models-0	null	null	null	[]	[]
158	generative models	Diffusion-based image generation models excel at producing high-quality synthetic content, but suffer from slow and computationally expensive inference. Prior work has attempted to mitigate this by caching and reusing features within diffusion transformers across inference steps. These methods, however, often rely on r...	POSITIVE	human	Evolutionary Caching to Accelerate Your Off-the-Shelf Diffusion Model	z9MlGsQbzR	6	3	[ "The paper introduces a principled optimization perspective (Pareto frontier with NSGA-II) that replaces heuristic-based caching rules. This reframing is conceptually elegant and well-motivated.", "framing diffusion caching as a multi-objective Pareto optimization problem to balance quality and efficiency is nove...	[]
159	generative models	We introduce Compositional Multi-Modal Generation (CMMG), a framework that learns to dynamically orchestrate generation across multiple modalities and abstraction levels. Unlike conventional generative models that produce outputs in a fixed modality, CMMG employs a learned policy network that decides which representati...	NEGATIVE	generated	LLM-generative models-12	null	null	null	[]	[]
16	foundation or frontier models, including LLMs	Multimodal Large Language Models (MLLMs), which couple pre-trained vision encoders and language models, have shown remarkable capabilities. However, their reliance on the ubiquitous Pre-Norm architecture introduces a subtle yet critical flaw: a severe norm disparity between the high-norm visual tokens and the low-norm ...	POSITIVE	human	The Unseen Bias: How Norm Discrepancy in Pre-Norm MLLMs Leads to Visual Information Loss	GVVNG2EMQv	6	3	[ "The paper presents a novel perspective, using effective theoretical and empirical analysis (like inter-layer cosine similarity) to confirm the impact of \"norm discrepancy\" on the model's \"asymmetric update dynamics\".", "By introducing the concept of \"representational inertia,\" the authors provide a new the...	[]
160	generative models	We present LongLive, a frame-level autoregressive (AR) framework for real-time and interactive long video generation. Long video generation presents challenges in both efficiency and quality. Diffusion and Diffusion-Forcing models can produce high-quality videos but suffer from low efficiency due to bidirectional atten...	POSITIVE	human	LongLive: Real-time Interactive Long Video Generation	nCAODkpsPJ	6.67	3.33	[ "A new KV-recache strategy is proposed for frame-level AR with multiple prompts.", "several techniques are simple (which is good) yet novel and effective, such as KV recache.", "This paper tackles an important practical problem—interactive long video generation—and provides the first system that delivers >20 FP...	[]
161	generative models	Generative models excel at synthesizing individual modalities but struggle to maintain semantic coherence when composing across vision, language, and structured data simultaneously. We introduce Compositional Cross-Modal Diffusion, a framework that enables coherent multi-modal generation by learning a shared semantic m...	NEGATIVE	generated	LLM-generative models-2	null	null	null	[]	[]
162	generative models	Reinforcement learning (RL) algorithms have been used recently to align diffusion models with downstream objectives such as aesthetic quality and text-image consistency by fine-tuning them to maximize a single reward function under a fixed KL regularization. However, this approach is inherently restrictive in practice,...	POSITIVE	human	Diffusion Blend: Inference-Time Multi-Preference Alignment for Diffusion Models	M2DXbwO8le	6.5	3.25	[ "The problem is novel and well-motivated. The paper formalizes the inference-time multi-preference alignment problem from the perspective of MORL, which is practically important and underexplored in diffusion literature.", "The studies of conducting inference time alignment on diffusion models are novel and inter...	[]
163	generative models	Current generative models produce outputs through predetermined sampling trajectories, whether via autoregressive decoding or fixed denoising schedules. We introduce Reflective Generation, a framework that endows generative models with metacognitive capabilities to dynamically adapt their generation process. Our approa...	NEGATIVE	generated	LLM-generative models-3	null	null	null	[]	[]
164	generative models	The probabilistic diffusion model (DM), generating content by inferencing through a recursive chain structure, has emerged as a powerful framework for visual generation. After pre-training on enormous data, the model needs to be properly aligned to meet requirements for downstream applications. How to efficiently align...	POSITIVE	human	Half-order Fine-Tuning for Diffusion Model: A Recursive Likelihood Ratio Optimizer	AZ6lqcvHLX	6.67	3	[ "The paper presents a novel fine-tuning scheme and devises gradient estimators for the diffusion model's chain-of-thought, which appears genuinely innovative.", "This paper proposes the Recursive Likelihood Ratio (RLR) optimizer, a novel \"Half-Order\" (HO) fine-tuning paradigm that successfully overcomes these c...	[]
165	generative models	Contemporary generative models synthesize content through monolithic processes that entangle high-level compositional decisions with low-level rendering details, limiting interpretability and fine-grained control. We introduce Compositional Hierarchical Generation (CHG), a framework that explicitly decouples semantic p...	NEGATIVE	generated	LLM-generative models-13	null	null	null	[]	[]
166	generative models	Score-based methods, such as diffusion models and Bayesian inverse problems, are often interpreted as learning the data distribution in the low-noise limit ($\sigma \to 0$). In this work, we propose an alternative perspective: their success arises from implicitly learning the data manifold rather than the full distribu...	POSITIVE	human	When Scores Learn Geometry: Rate Separations under the Manifold Hypothesis	34V0IZytle	6.67	3	[ "The rate separation results (Theorem 4.1) are novel and important for this line of research.", "Under standard assumptions, the results give new insight into why learning geometry information is \"easier\" than distribution learning in score-based models.", "The authors' main hypothesis - that recovering manif...	[]
167	generative models	We introduce Compositional Latent Decomposition Networks (CLDNets), a generative framework that learns to factor complex data distributions into hierarchical, reusable latent components. Unlike monolithic generative models that encode entire scenes into unstructured latent codes, CLDNets discover interpretable factoriz...	NEGATIVE	generated	LLM-generative models-8	null	null	null	[]	[]
168	generative models	Fast and high-quality language generation is the holy grail that people pursue in the age of AI. In this work, we introduce Discrete Diffusion Divergence Instruct (DiDi-Instruct), a training-based method that initializes from a pre-trained diffusion large language model (dLLM) and distills a few-step st...	POSITIVE	human	Ultra-Fast Language Generation via Discrete Diffusion Divergence Instruct	mtdyZsa47V	6.67	3.33	[ "DiDi-Instruct achieves new state-of-the-art results on the OpenWebText benchmark, consistently delivering lower PPL across 8-128 NFEs with over 20x faster distillation time than competing methods.", "This method is different from time-matching distillations and provides a route to few-step discrete generation." ...	[]
169	generative models	We introduce a generative framework that learns to iteratively revise and refine its outputs through explicit self-critique and localized editing. Unlike conventional generative models that produce outputs in a single forward pass or through uniform denoising schedules, our approach decomposes generation into an initia...	NEGATIVE	generated	LLM-generative models-9	null	null	null	[]	[]
17	foundation or frontier models, including LLMs	Foundation models demonstrate impressive capabilities but lack principled mechanisms for uncertainty quantification, leading to overconfident predictions and hallucinations on out-of-distribution inputs. We introduce Epistemic Adapters, a parameter-efficient framework that retrofits pretrained large language models wit...	NEGATIVE	generated	LLM-foundation or frontier models, including LLMs-0	null	null	null	[]	[]
170	generative models	While exocentric video synthesis has achieved great progress, egocentric video generation remains largely underexplored, which requires modeling first-person view content along with camera motion patterns induced by the wearer's body movements. To bridge this gap, we introduce a novel task of joint egocentric video and...	POSITIVE	human	EgoTwin: Dreaming Body and View in First Person	QFJkvv3zMi	7	3.25	[ "The design of head-centric motion tokenization is novel.", "To the best of my knowledge, EgoTwin is the first work to explicitly model joint egocentric video and human motion generation.", "The paper introduces a novel and clearly defined problem of joint egocentric video and human motion generation, with st...	[]
171	generative models	Diffusion models excel at generating high-quality images but struggle with compositional reasoning and fine-grained attribute control. We introduce Compositional Latent Diffusion, a framework that bridges structured scene representations with diffusion-based generation. Our approach learns a disentangled latent space w...	NEGATIVE	generated	LLM-generative models-6	null	null	null	[]	[]
172	generative models	We introduce Deliberative Generation Networks, a framework that equips generative models with explicit multi-step reasoning capabilities during the generation process. Unlike conventional approaches that produce outputs through direct sampling or fixed refinement schedules, our method decomposes generation into a seque...	NEGATIVE	generated	LLM-generative models-7	null	null	null	[]	[]
173	generative models	To model discrete sequences such as DNA, proteins, and language using diffusion, practitioners must choose between three major methods: diffusion in discrete space, Gaussian diffusion in Euclidean space, or diffusion on the simplex. Despite their shared goal, these models have disparate algorithms, theoretical structur...	POSITIVE	human	A Unification of Discrete, Gaussian, and Simplicial Diffusion	1taAXRcm21	6	3.5	[ "The idea of introducing repeats is interesting and novel, as are the theoretical convergence results.", "being able to tie together disparate concepts into a unified framework will be helpful for future researchers aiming to navigate the field", "The paper reveals an intriguing connection between neutral evolu...	[]
174	generative models	Existing generative models excel at producing high-fidelity samples but lack interpretable mechanisms for compositional control over semantic attributes. We introduce Compositional Latent Flow Networks, a generative framework that decomposes the generation process into modular, reusable latent components corresponding ...	NEGATIVE	generated	LLM-generative models-4	null	null	null	[]	[]
175	generative models	Diffusion-based generative models have reformed generative AI, and have enabled new capabilities in the science domain, for example, generating 3D structures of molecules. Due to the intrinsic problem structure of certain tasks, there is often a symmetry in the system, which identifies objects that can be converted by ...	POSITIVE	human	Quotient-Space Diffusion Models	3JPAkwSVc4	7.5	3.5	[ "The formulation of diffusion models via horizontal lifting on quotient spaces is innovative. It generalizes previous work on equivariant diffusion and provides a unifying geometric perspective that can be applied to various symmetry groups.", "The paper provides a powerful and unifying theoretical lens to explai...	[]
176	generative models	We introduce Neural Program Rendering (NPR), a generative framework that decomposes synthesis into symbolic program induction and differentiable execution. Rather than directly generating pixels or tokens, our approach first learns a compact domain-specific language whose programs capture compositional structure and hi...	NEGATIVE	generated	LLM-generative models-1	null	null	null	[]	[]
177	generative models	Recent image editing models have achieved impressive results while following natural language editing instructions, but they rely on supervised fine-tuning with large datasets of input-target pairs. This is a critical bottleneck, as such naturally occurring pairs are hard to curate at scale. Current workarounds use syn...	POSITIVE	human	Learning an Image Editing Model without Image Editing Pairs	OHqZ61ZqNO	6	3	[ "The idea of implementing unpaired training setup using the prior knowledge of VLM and manifold constraint is a novel approach.", "The model proposes VLM-editing loss and DMD loss, which seems novel.", "The paper is, to my knowledge, the first to train a general instruction-following image editor using **gradie...	[]
178	generative models	End-to-end backpropagation requires storing activations throughout all layers, creating memory bottlenecks that limit model scalability. Existing block-wise training methods offer means to alleviate this problem, but they rely on ad-hoc local objectives and remain largely unexplored beyond classification tasks. We prop...	POSITIVE	human	DiffusionBlocks: Block-wise Neural Network Training via Diffusion Interpretation	pwVSmK71cS	6	3	[ "The method seems very novel. Converting these disparate tasks (classification, image generation, text generation) into denoising tasks and separating the forward/backprop for the network into different groups of layers (blocks) seems to solve a problem that prior block-wise training methods could not.", "The mai...	[]
179	generative models	Modern generative models excel at producing high-fidelity samples but often lack compositional reasoning and struggle to respect hard logical constraints during generation. We introduce Program-Guided Diffusion, a hierarchical generative framework that decomposes synthesis into explicit planning and rendering stages. O...	NEGATIVE	generated	LLM-generative models-5	null	null	null	[]	[]
180	generative models	Generative diffusion models excel at producing high-fidelity samples but struggle with compositional reasoning—generating outputs that simultaneously satisfy multiple interdependent constraints expressed as logical specifications. We introduce Constraint-Guided Diffusion, a framework that incorporates differentiable co...	NEGATIVE	generated	LLM-generative models-11	null	null	null	[]	[]
181	generative models	Massive Activations (MAs) are a well-documented phenomenon across Transformer architectures, and prior studies in both LLMs and ViTs have shown that they play a substantial role in shaping model behavior. However, the nature and function of MAs within Diffusion Transformers (DiTs) remain largely unexplored. In this wor...	POSITIVE	human	Massive Activations are the Key to Local Detail Synthesis in Diffusion Transformers	tOOAWDRjrb	6	3	[ "The proposed DG strategy is effective and novel.", "The finding of MA's role of synthesizing is novel and supported." ]	[]
182	generative models	Standard discrete diffusion models treat all unobserved states the same way, typically mapping them to an absorbing [MASK] token. This creates an "information void" where global semantic information that may be inferred for the masked tokens from the unmasked tokens is not directly passed from one denoising step to ano...	POSITIVE	human	Continuously Augmented Discrete Diffusion model for Categorical Generative Modeling	JNAZ3e7Bwt	6.5	3.25	[ "It bridges a known gap and is conceptually novel. This specific combination of discrete diffusion with continuous latent vectors when the tokens are masked had not been previously explored.", "I find the idea of augmenting MDMs with continuous latent vectors to provide semantic information is both novel and well...	[]
183	generative models	We introduce Physics-Informed Latent Diffusion (PILD), a generative framework that synthesizes physically plausible scientific data by embedding conservation laws and governing equations directly into the diffusion process. While conventional generative models excel at capturing statistical patterns, they frequently vi...	NEGATIVE	generated	LLM-generative models-10	null	null	null	[]	[]
184	generative models	We introduce Compositional Manifold Diffusion (CMD), a generative framework that learns to decompose the data manifold into semantically disentangled subspaces without explicit supervision. Rather than modeling generation as a single denoising trajectory, CMD performs diffusion over a product space of learned factor ma...	NEGATIVE	generated	LLM-generative models-15	null	null	null	[]	[]
185	generative models	Few-step diffusion or flow-based generative models typically distill a velocity-predicting teacher into a student that predicts a shortcut towards denoised data. This format mismatch has led to complex distillation procedures that often suffer from a quality--diversity trade-off. To address this, we propose policy-base...	POSITIVE	human	pi-Flow: Policy-Based Few-Step Generation via Imitation Distillation	1vAte8dsap	6	3	[ "The policy-based perspective is a conceptually elegant and novel reinterpretation of few-step distillation, bridging diffusion/flow modeling and imitation learning.", "Distilling a policy function that predicts an entire trajectory, while conceptually related to operator learning in DFNO (Deep Fourier Neural Ope...	[]
186	learning on time series and dynamical systems	Understanding complex temporal phenomena requires models that are both predictive and interpretable through the lens of dynamical systems theory. We introduce Neural Dynamical Manifolds, a framework that learns low-dimensional latent representations where data evolves according to interpretable differential equations w...	NEGATIVE	generated	LLM-learning on time series and dynamical systems-0	null	null	null	[]	[]
187	learning on time series and dynamical systems	The design of training objective is central to training time-series forecasting models. Existing training objectives such as mean squared error mostly treat each future step as an independent, equally weighted task, which we found leading to the following two issues: (1) overlook the label autocorrelation effect amon...	POSITIVE	human	Quadratic Direct Forecast for Training Multi-Step Time-Series Forecast Models	vpO8n9AqEG	6	3	[ "The proposed method to improve the training objectives for time series forecasting is novel and applicable to related problems in this domain.", "The paper introduces a technically novel strategy to solve an empirically and theoretically motivated problem that is much relevant in neural, (direct) non-autocorrela...	[]
188	applications to physical sciences (physics, chemistry, biology, etc.)	Discovering interpretable governing equations from experimental data remains a central challenge in the physical sciences, particularly when measurements are noisy, sparse, or high-dimensional. We introduce a neural-symbolic framework that combines the representational power of neural operators with the symbolic reason...	NEGATIVE	generated	LLM-applications to physical sciences (physics, chemistry, biology, etc.)-2	null	null	null	[]	[]
189	applications to physical sciences (physics, chemistry, biology, etc.)	Recently, many generative models for de novo protein structure design have emerged. Yet, only few tackle the difficult task of directly generating fully atomistic structures jointly with the underlying amino acid sequence. This is challenging, for instance, because the model must reason over side chains that change in ...	POSITIVE	human	La-Proteina: Atomistic Protein Generation via Partially Latent Flow Matching	RDerF20JYT	8	3.25	[ "examining for clash rates and rotamer frequencies sets a new evaluation metric that is meaningful for this subfield", "This is a novel solution not seen in prior protein diffusion models, which either treated all atoms explicitly (incurring huge complexity) or attempted fully latent approaches.", "The paper pr...	[]

End of preview. Expand in Data Studio

Old Ideas, Novel Problems:
The Instability of LLM-Based Novelty Evaluation

Rewording the judge prompt changes GPT-5.4's and Opus 4.6's accuracy

Benchmark instances for the paper Old Ideas, Novel Problems: The Instability of LLM-Based Novelty Evaluation. Code for running judges on this benchmark and reproducing the paper's experiments is at allenai/novelty-eval.

AI-based ideation systems are often evaluated on the novelty of their ideas, and that judgment is increasingly delegated to large language models. Such judges are typically built ad hoc and validated, if at all, on human-authored papers rather than on the generated ideas they are meant to score. So, how do novelty judges perform?

Not well. We present a systematic controlled study of novelty evaluation design choices. We build an evaluation set automatically, mining OpenReview for passages where reviewers explicitly affirm or dispute a paper's originality and keeping only submissions with unanimous agreement at the extremes of their research area; we pair these with ideas from a vanilla LLM generator. Across six judges, we find that seemingly small prompt design choices have large consequences; e.g., simply telling the judge that reviewers found one idea novel and the other not can change its verdict on more than half of the identical idea pairs it is shown, shifting pairwise accuracy by over 50 points and occasionally pushing it below chance. The same change helps one judge and hurts another. Retrieval and larger reasoning budgets help little, and two purpose-built novelty evaluators are outperformed by our cheapest prompted baseline. These results raise questions about reported novelty gains of automated ideation systems, and call for robust novelty evaluation methods.

Setups

Novelty evaluation data creation pipeline

The benchmark consists of two data setups: Human-Only and Human+Generated. Both setups consist of two separate idea "pools": a high-novelty pool D_+ and a lower-novelty pool D_-, where ideas from D_+ can be assumed to be more novel than those in D_-. The two setups share the same D_+ and differ in where D_- comes from.

Setup	`D_+` (novel)	`D_-` (lower-novelty)
Human-Only (`human-only_*`)	154 validated-novel	145 validated-lower-novelty
Human+Generated (`human-plus-generated_*`)	154 validated-novel	154 weakly-labeled-lower-novelty

Where:

validated-novel (both setups) — ICLR 2026 submissions where a majority of reviewers explicitly praised the originality of the contribution itself and none disputed it; accepted, in the top rating decile of their primary area, with an average contribution score >= 3.0.
validated-lower-novelty (Human-Only) — the mirror image: a majority of reviewers faulted the originality and none praised it; rejected, in the bottom rating decile of their primary area, with an average contribution score <= 2.0.
weakly-labeled-lower-novelty (Human+Generated) — ideas from Claude Sonnet 4.5, prompted simply to propose a novel idea with no scaffold, literature access, or tools.

Formats

Each setup is instantiated in two evaluation formats:

pointwise — one idea per row, binary label: is this idea novel?
pairwise — one D_+ idea against one D_- idea; which of the two is more novel?

and in two idea formats:

abstract (default) — the idea as a free-form abstract.
plan (*_plan configs) — the same ideas normalized into a two-field purpose / mechanism research plan.

The *_backbone-* configs regenerate D_- with a different ideation backbone and are used in the negatives-source experiment.

Configs at a glance

Config	Rows	Contents
`human-only_pointwise`, `human-only_pointwise_plan`	299	154 novel, 145 lower-novelty
`human-only_pairwise`, `human-only_pairwise_plan`	154	Novel idea is `idea_a` in 84 pairs, `idea_b` in 70
`human-plus-generated_pointwise` (default), `human-plus-generated_pointwise_plan`	308	154 novel, 154 lower-novelty
`human-plus-generated_pairwise`, `human-plus-generated_pairwise_plan`	154	Novel idea is `idea_a` in 80 pairs, `idea_b` in 74
`human-plus-generated_pointwise_backbone-{gpt-5,gpt-5.1,gpt-5.4,opus-4-5}`	308 each	As above, with `D_-` from GPT-5, GPT-5.1, GPT-5.4, or Claude Opus 4.5
`human-plus-generated_pairwise_backbone-{gpt-5,gpt-5.1,gpt-5.4,opus-4-5}`	154 each	As above, with `D_-` from GPT-5, GPT-5.1, GPT-5.4, or Claude Opus 4.5

All configs have a single test split.

Loading

from datasets import load_dataset

ds = load_dataset("allenai/novelty-judge-bench", "human-plus-generated_pointwise", split="test")
print(ds[0]["idea"], ds[0]["label"])

Fields

Pointwise configs

Field	Type	Description
`id`	int32	Instance id, unique within the config
`iclr_area`	string	The idea's ICLR primary area
`idea`	string	The idea to judge (abstract or plan form, depending on the config)
`label`	string	`POSITIVE` (`D_+`, novel) or `NEGATIVE` (`D_-`, lower-novelty) — the ground truth
`idea_source`	string	`human` (an ICLR submission) or `generated` (LLM ideator output)
`title`	string	Submission title, or a synthetic identifier for generated ideas
`forum_id`	string	The submission's OpenReview forum ID (`https://openreview.net/forum?id=<forum_id>`); `null` for generated ideas
`rating`	float64	Mean reviewer rating; `null` for generated ideas
`contribution`	float64	Mean reviewer contribution score; `null` for generated ideas
`positive_signals`	list[string]	Reviewer excerpts praising originality
`negative_signals`	list[string]	Reviewer excerpts faulting originality

Pairwise configs

Field	Type	Description
`id`	int32	Instance id, unique within the config
`iclr_area`	string	The ICLR primary area shared by both ideas in the pair
`idea_a`, `idea_b`	string	The two ideas to compare
`expected_winner`	int8	`0` if `idea_a` is the more novel one, `1` if `idea_b`
`label_a`, `label_b`	string	`POSITIVE` / `NEGATIVE` for each side
`idea_source_a`, `idea_source_b`	string	`human` or `generated`
`title_a`, `title_b`	string	Per-side titles
`forum_id_a`, `forum_id_b`	string	Per-side OpenReview forum IDs; `null` if generated
`rating_a`, `rating_b`	float64	Per-side mean reviewer rating; `null` if generated
`contribution_a`, `contribution_b`	float64	Per-side mean contribution score; `null` if generated
`positive_signals_a/_b`	list[string]	Per-side reviewer excerpts praising originality
`negative_signals_a/_b`	list[string]	Per-side reviewer excerpts faulting originality

Idea order within each pair is randomized, so always picking the same position scores close to chance (see the A/B split in Configs at a glance).

Intended use

This dataset is for evaluating novelty judges: measuring how well an automatic judge (e.g., a prompted LLM) tells more novel research ideas from less novel ones. It is a small test set with a single test split. It is not training data and should not be used to train or fine-tune models.

Model-generated content

Parts of this dataset were produced by proprietary models from Anthropic (Claude) and OpenAI (GPT).

Models and their roles

Model	Provider	Role in the paper	Output in this release	Volume
Claude Sonnet 4.5	Anthropic	Idea generator; judge	`D_-` ideas of `human-plus-generated_*`	154 ideas
Claude Opus 4.5	Anthropic	Idea generator; judge	`D_-` ideas of `*_backbone-opus-4-5`	154 ideas
GPT-5	OpenAI	Idea generator	`D_-` ideas of `*_backbone-gpt-5`	154 ideas
GPT-5.1	OpenAI	Idea generator; judge	`D_-` ideas of `*_backbone-gpt-5.1`	154 ideas
GPT-5.4	OpenAI	Idea generator; judge	`D_-` ideas of `*_backbone-gpt-5.4`	154 ideas
Claude Opus 4.6	Anthropic	Data processing and labeling; judge	Edited human abstracts; plan-form rewrites; selection of reviewer excerpts	299 abstracts; 453 plans; 1,192 excerpts
GPT-5.2	OpenAI	Judge	None	—
GPT-5.6	OpenAI	Judge (web-search retrieval experiment)	None	—

No judge outputs are part of this release: no verdicts, prompt/response logs, reasoning traces, or retrieval transcripts.

Authorization status: none required; evaluation use.

Collection method

All model outputs were collected through the providers' APIs (Anthropic API, OpenAI API), using the prompt templates published in the code repository.

Generated ideas — the generator receives only an ICLR primary area and is asked to propose a novel research idea as an abstract, with no scaffold, literature access, or tools. Each generated idea is one model response.
Edited abstracts — each human abstract is passed once through a prompt that removes results, baselines, and evaluation datasets and otherwise keeps the original wording.
Plan-form ideas (*_plan configs) — each abstract-form idea, human or generated, is rewritten into a two-field purpose / mechanism plan.
Reviewer excerpts — the model is given the Strengths and Weaknesses sections of each official review and copies out, verbatim, the sentences that judge the paper's novelty. The model selects the excerpts; the text is the reviewer's.
Labels — POSITIVE / NEGATIVE are assigned by fixed rules over reviewer scores, decisions, and the selected excerpts (see Setups); generated ideas are NEGATIVE by construction. No model assigns labels directly.

Which fields are AI-generated

Field	Origin	Model
`idea` / `idea_a` / `idea_b`, where `idea_source` is `generated`	Model-generated	Claude Sonnet 4.5, or the model named in the `_backbone-*` config
`idea` / `idea_a` / `idea_b`, where `idea_source` is `human` (abstract configs)	Human abstract, edited by a model	Claude Opus 4.6
`idea` / `idea_a` / `idea_b` in `*_plan` configs	Model rewrite of the abstract-form idea	Claude Opus 4.6
`positive_signals`, `negative_signals`	Verbatim reviewer text, selected by a model	Claude Opus 4.6 (selection only)
`title` of human ideas, `forum_id`, `iclr_area`, `rating`, `contribution`	OpenReview	—
`title` of generated ideas, `label`, `expected_winner`, `idea_source*`, `id`	Assigned by our pipeline	—

OpenReview attribution

The positive_signals* and negative_signals* fields are excerpts from official reviews of ICLR 2026 submissions, obtained from OpenReview and licensed under CC BY 4.0. They are excerpts, not full reviews: each is a sentence or short passage copied verbatim from a review's Strengths or Weaknesses section.

Reviews on OpenReview are anonymous. This dataset contains no reviewer names or other reviewer metadata.

Papers and removal requests

Papers are identified by their OpenReview forum ID (forum_id*). The dataset stores no author names.

If you are an author of a paper in this dataset and would like it removed, email the first author (see Authors) with the paper's forum ID. We will remove every instance that contains the paper from this dataset and from the code repository, and add it to the pipeline's blocklist so that rebuilding the benchmark does not bring it back.

License

This dataset is licensed under ODC-BY 1.0. It is intended for research and educational use in accordance with Ai2's Responsible Use Guidelines.

The reviewer excerpts come from OpenReview under CC BY 4.0 (see OpenReview attribution); keep that attribution when redistributing them.

Some fields are outputs of Anthropic and OpenAI models (see Model-generated content) and may also be subject to Anthropic's and OpenAI's terms.

Authors

Noy Sternlicht — Hebrew University of Jerusalem, Allen Institute for AI
Simra Shahid — Microsoft
Peter Jansen — Allen Institute for AI, University of Arizona
Daniel S. Weld — Allen Institute for AI, University of Washington
Pao Siangliulue — Allen Institute for AI
Tom Hope — Hebrew University of Jerusalem, Allen Institute for AI

Cite us!

@misc{sternlicht2026oldideasnovelproblems,
      title={Old Ideas, Novel Problems: The Instability of LLM-Based Novelty Evaluation}, 
      author={Noy Sternlicht and Simra Shahid and Peter Jansen and Daniel S. Weld and Pao Siangliulue and Tom Hope},
      year={2026},
      eprint={2610.02022},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2610.02022}, 
}