Gyubeum/AndroidFlux_RM_train

AndroidFlux RM train

Training data for a reward model of mobile GUI agents on AndroidWorld: 1,803 groups. Each group is one screen of a task on which an agent has just made an error. It holds six candidate next actions, and the likelihood each local agent gives to the candidates it did not write.

Groups

• Context. Every group comes from an AndroidWorld trajectory in which an agent made a mistake. The context is the state right after the first erroneous step: the task goal, the step history so far and up to five screenshots, the current screen last. An error detector labelled the type of the error and the step where it happened.

• Candidates. Six per group:

  • five from local agents: Qwen3-VL-2B, Qwen3-VL-8B, GUI-Owl-7B, GUI-Owl-1.5-8B and MAI-UI-8B. Each agent was replayed through the same trajectory up to the context, in its own prompt format, and asked for its next action;
  • one from Gemini 3.1 Pro, generated the same way with the Qwen3-VL prompt.

  Every action is written as a mobile_use tool call in one common format.

• Peer log-probabilities. Each local agent scored every candidate it did not write, as the assistant turn after its own request for that context. The score is the log-probability of the tool call, summed over its tokens, together with the token count. An agent never scores its own candidate, so each group has 5 × 5 = 25 scores.

• Selection. A Gemini judge reviewed 2,154 candidate groups. For each it decided whether the context is usable and whether each local agent's candidate is good, bad or uncertain.

  • A group was kept when its context is usable and at least one local candidate is good and valid under the tool specification.
  • 1,803 groups passed.
  • The judge only selected groups: no row, candidate or score was changed.

Files

file	content
train.jsonl	one group per line: sample_id, task, error_type, first_error_step, messages (system prompt and user turn, with an <image> marker per screenshot), images, image_ids, n_images, screen_width, screen_height, history_entries, candidates (candidate_id, source_model, response, parsed_action)
peer_logprobs.jsonl	45,075 scores: sample_id, evaluator, candidate_id, logprob_sum, token_count
images/	6,380 screenshots, each named by the sha256 of its content

Distribution

Contexts

• Tasks. 88 AndroidWorld tasks, a median of 16 groups per task and at most 74. The most frequent:

  task	groups
  BrowserDraw	74
  BrowserMultiply	63
  AudioRecorderRecordAudioWithFileName	60
  MarkorCreateFolder	54
  CameraTakePhoto	54
  RecipeAddMultipleRecipes	53
  SimpleSmsSend	52
  SimpleCalendarAddOneEvent	48
  MarkorCreateNote	46
  RetroPlayingQueue	44

• Error types.

  error type	groups
  stagnation	832
  goal_regression	479
  incomplete_execution	295
  goal_irrelevant	181
  goal_direction	12
  system_error	4

• First error step. From 1 to 30, with a median of 5.

• Screenshots per group. 5 in 1,492 groups, 4 in 213 and 3 in 98.

• Screen size. 1080 × 2400 in 1,283 groups and 1092 × 2408 in 520.

• Distinct local candidates per group. 5 in 1,368 groups, 4 in 305, 3 in 93 and 2 in 37.

Candidate actions

action	Qwen3-VL-2B	Qwen3-VL-8B	GUI-Owl-7B	GUI-Owl-1.5-8B	MAI-UI-8B	Gemini 3.1 Pro	all
click	1,232	1,168	1,140	1,348	1,413	1,185	7,486
swipe	182	235	253	208	91	194	1,163
system_button	67	136	108	103	70	122	606
terminate	118	110	188	5	5	59	485
type	79	55	48	69	35	32	318
answer	56	64	47	25	57	28	277
open	25	24	3	12	51	150	265
long_press	9	8	14	11	71	28	141
wait	35	3	2	22	10	5	77

In source_model, the agents are qwen3vl_2b, qwen3vl_8b, gui_owl_7b, guiowl15_8b, mai_ui_8b and gemini_3_1_pro_qwen_prompt.