MihailSlutsky/vistr-process-verification-pilot

ViSTR Process-Verification Pilot (14 answer-correct trajectories, multimodal)

Agent trajectories for studying process false positives in multimodal agents: cases where the answer is correct but the visual reasoning that produced it is wrong. Ships the raw perception tool outputs so any claim in a trajectory can be independently re-verified, plus human annotations and an unmodified XSkill critique of the same trajectories.

Why this exists

Harness / skill self-evolution pipelines usually decide whether a rollout is "good" from the final answer or task reward, then distil skills from the good ones. If a rollout is right for the wrong reasons and the critic does not notice, the faulty reasoning enters the skill library and keeps contaminating later evolution. This package is the pilot evidence for that failure mode, and a concrete measurement of how one published method (XSkill, ICML 2026) behaves on it.

Headline: 5 / 14 answer-correct trajectories have broken processes. XSkill flagged 0 of them, and distilled the faulty rubric into reusable "experience" in at least 2 cases. Details and full evidence trails in XSKILL_FINDINGS.md.

Provenance

• Benchmark: ViSTR-Bench public split @ d87a003, skill_val split (seed 42, 408/131/131)
• Solver: qwen3-vl-plus driving a pi-4d-agent harness (fork of pi, branch gaozhe); model sees only read + bash, perception via a 4d CLI inside a bwrap sandbox
• Perception backends (real GPU, AMD MI308X): SAM 3.1, CoTracker3, Depth-Anything-3, VGGT-Omega
• Mode: perception (tool guidance in the system prompt, no distilled skill), max_steps 50, unlimited output tokens, seed 42
• These 14 are the answer-correct subset of a 20-sample skill_val slice; the 6 answer-wrong ones are kept out of this package

Layout

manifest.json                     # per-sample metadata, answers, counts
XSKILL_FINDINGS.md                # the XSkill failure analysis (main writeup)

trajectories/<sample_id>/
  trajectory.json                 # full-fidelity harness trajectory (no base64)
  viewed_images_and_jsonl/
    trajectory.jsonl              # XSkill-format trace (line 1 = initial_prompt/ground_truth)
    tool_image_NNN.jpg            # the frames the solver ACTUALLY read, in order
  tool_outputs/                   # raw perception evidence, re-verifiable
    tracks/tracks.json            # per-frame bbox/centroid/visibility/confidence
    detections*.json              # grounding results
    depth/…, *_masks/mask_*.png   # depth metadata, segmentation masks
    metadata.json                 # inspect-video output

xskill_critique/<sample_id>/
  summary.txt                     # XSkill SINGLE_ROLLOUT_SUMMARY output
  exp_intra_prompt.txt            # exact prompt sent to INTRA_SAMPLE_CRITIQUE
  exp_intra_resp.txt              # full free-text critique + distilled JSON
xskill_critique/results.json      # distilled experience ops per sample

annotations/
  human_gt.json                   # per-case labels, verbatim reviewer notes, confidence
  tool_grounded_critic_v1.json    # claim-wise critic reading raw tool data (naive)
  tool_grounded_critic_v3.json    # same, with severity calibration

Videos are not included — resolve manifest.json → video_relative_path against the ViSTR-Bench public split.

Annotation schema

annotations/human_gt.json labels each case:

label	meaning
process_false_positive	answer correct, reasoning rests on a false/fabricated premise
process_ok	answer correct, cited evidence supports the conclusion
gray_zone	no outright false premise, but the rubric is unverifiable / not generalisable, or annotators disagree

Counts: 5 process_false_positive, 6 process_ok, 3 gray_zone.

Every row carries confidence (high / derived / reviewer_only) so annotation gaps stay visible, false_claim + why_wrong for the failure, and reviewer_crosscheck recording the second pass — including two cases where the reviewer and the annotator initially disagreed.

The two clearest false positives

#23 (billiards, "will the ball go in?", GT Yes) — solver claimed the ball's bbox shrank and it vanished into a corner pocket. tracks/tracks.json says: area went up (0.0007 → 0.00076), final centroid (0.733, 0.862) is not a corner, and after a 4-frame gap the ball is tracked for 49 more frames at visibility=1.0. XSkill called this "a domain-aware heuristic that aligned with physics and ground truth".

#156 (Jenga stability, GT Yes) — solver argued "no wobble in the video, so the tower stays stable", but the video ends before the block is out. XSkill turned this into an experience telling future agents to prioritise tracking stability "over static structural flaws" — i.e. it promoted the invalid signal and demoted the valid one.

Limitations

• n = 14 supports existence and taxonomy claims, not stable rates
• Single primary annotator; inter-annotator agreement not measured
• XSkill's ExperienceLLM defaults to temperature=0.6; repeated runs on the same trajectory produce different experiences, so single-run numbers must not be quoted without variance
• XSkill was run in "option A" (no original_image*, so its critique sees only what the solver saw, textualised). The overview-frames variant is untested
• Only XSkill Phase-I summarize → intra-sample critique was run; hierarchical consolidation and Phase-II injection were not
• The solver's own tool output was sometimes truncated by the bash tool (#23 received a 51KB mid-section of the tracking JSON), so a critic reading tool_outputs/ sees strictly more than the solver did

Related

• Raw pilot trajectories (100 rollouts, images stripped): MihailSlutsky/vistr-4dagent-skillopt-pilot
• XSkill: paper · code