RS-Claw-Evolution: Environment-Feedback-Driven Evolution for Lightweight Remote Sensing Agents in Long-Horizon Tasks
Kai Ouyang, Dongyang Hou, Liangtian Liu, Zeyuan Wang, Ziyu Li, Chengfu Liu, Zichao Tang, Xuezhi Cui, et al.
cs.LG · cs.AI · cs.CV
Quick verdict
Three-stage feedback-driven training lifts Qwen3-4B to 65.9% on Earth-Bench AP, beating 32B baselines; strong gains but single-run, benchmark-coupled evidence.
Summary
The paper proposes RS-Claw-Evolution, a three-stage training framework for lightweight remote sensing (RS) agents. Stage one replaces JSON tool calls with executable Python (CodeAct) to control observations and maintain intermediate state. Stage two constructs SFT data by having an open-source teacher analyze its own failed trajectories, generate hindsight hints, and re-explore; rejection sampling keeps successful trajectories that contain errors plus recovery, and an error-turn masking loss zeroes supervision on faulty actions while conditioning recovery on them. Stage three applies GRPO with a multi-dimensional environment feedback reward (progress ratios for tool, document, and dataflow steps against reference workflows, plus turn and error penalties) and turn-level advantage protection. On Earth-Bench (176 held-out workflow tasks; 12 training tasks covering 28 of 104 tools), Qwen3-4B improves from 31.8% to 65.9% AP accuracy after SFT+RL, exceeding untrained Qwen3-32B (43.8%) and DeepSeek-V3.1 (60.8%), approaching GPT-5 (71.6%); Qwen3-8B reaches 61.4%. Hint guidance raises teacher success rates from 23.5% to 69.3%; masked SFT beats oracle and unmasked variants on both accuracy and error-interaction rate (4B: 41.5%/31.6 vs 33.5%/57.4 and 37.5%/36.9). The Code paradigm cuts token overhead (GPT-5 action tokens down 50.4% in AP mode). Limitations: reliance on benchmark ground-truth reference workflows for hints and rewards, a static tool ecosystem, single-benchmark validation, and no variance estimates or significance testing in the visible text.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 3
● low 1
5 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
- A three-stage pipeline (CodeAct interaction, failure-aware SFT with error-turn masking, GRPO with MEFR rewards and TAP) converts environment feedback into training-time supervision for lightweight RS agents.
- Using only 12 training tasks, the optimized Qwen3-4B reaches 65.9% AP accuracy on 176 held-out Earth-Bench tasks, exceeding untrained Qwen3-32B (43.8%) and DeepSeek-V3.1 (60.8%) and approaching GPT-5 (71.6%).
- Error-turn masking outperforms both error-free Oracle SFT and unmasked SFT on accuracy and error-interaction rate (Qwen3-4B: 41.5%/31.6 vs 33.5%/57.4 and 37.5%/36.9).
- Hindsight hints raise teacher trajectory success rates from 23.5% to 69.3% averaged over the 12 training tasks.
- CodeAct interaction reduces context overhead (e.g., GPT-5 AP action tokens down 50.4%), with the largest benefit for less capable models.
Strengths
Concrete, fully specified mechanisms
Error-turn masking is formalized (Eq. 2-5), MEFR is defined with reward pseudocode (Eq. 8-9, Algorithms 1-2), and all key prompts (CodeAct, JSON baseline, hindsight hint) are printed verbatim in Appendix E.
Strong low-resource empirical gains
With only 12 training tasks covering 28 of 104 tools, Qwen3-4B rises 31.8% to 65.9% AP accuracy and Qwen3-8B 19.3% to 61.4% (Table I), surpassing untrained Qwen3-32B (43.8%).
Careful SFT ablation design
Quantified failure-aware data construction
Transparent limitations section
Weaknesses
No uncertainty quantification
RL objective is benchmark-coupled
Narrowed and prompt-only baselines
Uncharacterized error judge
Paradigm-mixed tool metrics
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
RS-Claw-Evolution (arXiv 2609.22258; Central South University, Hunan University of Science and Technology, UESTC) is a method paper on training lightweight (4B-8B) LLM agents for multi-step remote sensing tool use, built on the RS-Claw skill-tree environment and Earth-Bench. It makes three mechanism claims plus one headline empirical claim: (1) CodeAct-style executable-code interaction controls context growth; (2) failure-aware SFT with error-turn masking extracts more value from imperfect teacher trajectories; (3) GRPO with a multi-dimensional environment feedback reward (MEFR) and turn-level advantage protection (TAP) fixes credit assignment in long horizons. Headline: Qwen3-4B rises from 31.8% to 65.9% AP accuracy over 176 held-out tasks, surpassing untrained Qwen3-32B (43.8%) and DeepSeek-V3.1 (60.8%), approaching GPT-5 (71.6%), all under the same Code paradigm (Table I).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 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 RS-Claw-Evolution?
A three-stage training framework that turns environment feedback into learning signals for small remote sensing agents: executable-code (CodeAct) interaction, failure-aware supervised fine-tuning with error-turn masking, and GRPO reinforcement learning with multi-dimensional environment rewards and turn-level advantage protection.
How accurate is the trained Qwen3-4B agent on Earth-Bench?
65.9% accuracy in Autonomous Planning mode across 176 held-out workflow tasks, versus 43.8% for the untrained Qwen3-32B, 60.8% for DeepSeek-V3.1, and 71.6% for GPT-5 under the same code-interaction protocol (Table I).
How much training data does the method need?
Only 12 Earth-Bench tasks — about 100 sampled teacher trajectories per task (roughly 1,200 SFT trajectories) — covering 28 of 104 tools; the remaining 176 workflow tasks are used purely for evaluation.
Is code available for RS-Claw-Evolution?
The provided text describes prompts, reward pseudocode, and hyperparameter symbols in detail but states no public code or data release; the framework builds on Earth-Bench from the Earth-Agent project and the slime post-training framework.
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 excerpt contains two [...sampling gap...] blocks; the Section III-A/B openings, the TAP subsection body, part of the Section IV setup, and parts of Appendices B-C were skipped - judgments about those components rely on the captions and summaries visible in the provided text.
- Hyperparameter values (MEFR weights, beta, epsilon, group size, learning rates) were not visible; they may appear in the skipped sections.
- No code-availability statement appears in the provided text; one could exist in skipped pages.
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.