DENSE: Distilling Agent Trajectories into Evidence-Grounded Shortcut Trees for Self-Refinement
Siyuan Liu, Fan Yu, Dongyu Ru, Yizhu Liu, Yifan Yang, Xuezhi Cao, Xunliang Cai, Yixin Cao
cs.AI
Quick verdict
DENSE distills unlabeled agent trajectories into shortcut-tree feedback; large pass-rate gains and token savings on Terminal-Bench 2.1, though untested significance and single-benchmark scope temper the claims.
Summary
DENSE converts an unlabeled agent trajectory into nested shortcut trees: an analysis model segments action-observation pairs into subtasks bottom-up, scores actions locally, compresses sibling attempts into shortcuts that retain dead ends and evidence, and closes inherited issues only when later recovery evidence is cited. REFIT, the paired evaluation protocol, hands every feedback method the same initial run0 under post-hoc outcome blindness, then reruns the agent from a reset environment with fresh context; strict pass (verifier reward exactly 1) changes are compared on Terminal-Bench 2.1 across four recipients (MiniMax-M2.7, DeepSeek V4 Pro, GPT-5.5, Kimi K2.6; 89 tasks, three repetitions). DENSE posts the highest strict pass rate among non-privileged baselines (+7.12 to +15.64 pp over run0) while cutting observed recipient tokens 19.0-43.6%, and beats privileged Verifier Feedback on MiniMax by 7.49 pp. GPT-5.5 ablations support hierarchy and shortcut/reconciliation components; iterative use on 30 hard tasks beats Self-reflection by 3.3-10.3 pp. Limitations: one benchmark, no significance testing, and no content-matched control isolating hierarchy from feedback content.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Trajectories contain reusable evidence that can be distilled into feedback without post-execution rewards, verifier outputs, or correctness labels.
- On Terminal-Bench 2.1, DENSE achieves the highest strict pass rate among non-privileged feedback methods across four recipient models, improving over initial attempts by 7.12-15.64 pp.
- DENSE reduces observed recipient tokens on retries by 19.0-43.6%, and on MiniMax-M2.7 organized evidence outperforms raw history augmented with privileged verifier signals by 7.49 pp.
- GPT-5.5 ablations show both nested hierarchy and shortcut/reconciliation components contribute (+3.00 and +3.37 pp over ablations).
- An exploratory three-iteration extension on 30 hard tasks shows DENSE exceeding Self-reflection in cumulative pass rate by 3.3-10.3 pp on all four models.
Strengths
Outcome-blind feedback protocol (REFIT) is the paper's strongest methodological asset
Section 3 gives all generators the identical run0, resets environment and model context, and keeps verifier outputs out of feedback generation; only the explicitly privileged Verifier Feedback condition sees them, directly addressing the resampling/privileged-information confounds the self-correction literature (Kamoi et al. 2024) warns about.
Consistent multi-recipient wins with honest privileged reference
DENSE gains +7.12 to +15.64 pp over run0 across four recipients (Table 2/Figure 9: MiniMax +7.87, DeepSeek +15.64, GPT +7.12, Kimi +11.61) while cutting observed tokens 19.0-43.6% (Table 3); the paper openly reports Verifier Feedback beating DENSE on GPT-5.5 (83.9 vs 75.66) and Kimi (53.9 vs 50.9) instead of hiding it.
Rare cost transparency including the feedback generator
Component ablation and candid case analyses
Weaknesses
No inferential statistics anywhere
Content/structure confound acknowledged but unresolved
Single-benchmark evidence
Token savings are lower bounds under differential censoring risk
Generator dependence unexamined
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
DENSE (Fudan University + Meituan Longcat Team; arXiv:2609.21423v3) targets feedback extraction from unlabeled agent trajectories: no rewards, no verifier outputs, no annotations. The method converts a finished trajectory into nested shortcut trees (Section 2), and REFIT (Section 3) evaluates any feedback by rerunning the same agent from a reset environment with fresh context, comparing strict pass (verifier reward exactly 1) against a shared run0. Claims: (1) highest strict pass rate among non-privileged methods on Terminal-Bench 2.1 across four recipients (MiniMax-M2.7, DeepSeek V4 Pro, GPT-5.5, Kimi K2.6); (2) +7.12 to +15.64 pp over run0 with 19.0-43.6% fewer observed tokens; (3) on MiniMax, DENSE beats even privileged Verifier Feedback by 7.49 pp; (4) ablations support both core components; (5) iterative use on hard tasks beats Self-reflection by 3.3-10.3 pp over three rounds.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the DENSE method in one sentence?
DENSE is a training-free pipeline that distills a finished AI-agent execution trace into hierarchical 'shortcut tree' feedback, retaining what worked, failed attempts, evidence, and unfinished requirements, so the same agent can retry the same task more successfully without any reward or correctness labels.
What is REFIT and why does it matter?
REFIT is an evaluation protocol: all feedback methods are built from the identical initial trajectory, the environment and model context are reset, and verifier results are hidden from feedback generation, so measured success-rate changes can be attributed to the feedback rather than to fresh sampling luck or leaked outcome information.
How well does DENSE perform and at what cost?
On Terminal-Bench 2.1 with four frontier models, DENSE improves strict pass rate by 7.12-15.64 percentage points over the first attempt while using 19.0-43.6% fewer observed recipient tokens; for GPT-5.5, feedback generation costs about $0.14 per task and the feedback-plus-rerun workflow totals $2.96 versus $1.81 for one baseline run, with the rerun alone 36.6% cheaper and 7.12 points more accurate.
Does DENSE use any ground-truth labels or verifier signals?
No. The distillation model (Gemini 3 Flash Preview) sees only the task description and the recorded actions and observations; rewards and verifier outputs are used solely to score outcomes afterward. The one exception is the 'Verifier Feedback' condition, an explicitly privileged reference baseline, which underperforms DENSE on two of four models.
What are the main limitations?
Everything is evaluated on a single benchmark (Terminal-Bench 2.1, 89 tasks); the paper reports no statistical significance tests; and no control separates the value of the tree structure from the value of the feedback content itself.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- The extracted text is truncated: portions of Table 2, Section 4.3, part of Related Work, Appendix B prose, and Appendix E prompts P7-P8 were omitted; headline numbers were cross-checked against Figures 8-9, Table 5, and appendix text.
- Figures are only partially parseable from the text dump (heatmap figures degrade to number streams), so per-task claims rely on the appendix prose.
- The supplementary package (code, prompts, consolidated statistics) is described but not itself reviewable here.
- v3 dated 24 Sep 2026; no prior-version diff available to assess claim evolution.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.