small-agent-policy-qwen3-5-4b-repeatedly-deployed-same-task CLM-0488For a small agent policy (Qwen3.5-4B) repeatedly deployed on the same task family with environment-level feedback, distilling only the trajectories whose textual-state guidance is newly or persistently useful relative to state-free execution raises the state-free policy's accuracy more than distilling all state-conditioned trajectories, and the resulting policy keeps improving over alternating rounds of state editing and consolidation.
Capability: Improving its own scaffolding from failures
Observed on
Post-deployment settings with recurring related tasks and verifiable environment rewards; small student (4B) with a larger fixed teacher (27B); QA, embodied and web-navigation benc.
Sources
- Validated by: SearchQA/ALFWorld/WebShop accuracy against state-only evolution (SkillOpt 53.9 avg), policy internalization (OPID 55.4), and co-evolution (SkillRL 56.2) versus 57.6 avg; ablation: selective 01+11 consolidation 63.0% vs 58.9% unfiltered; state-free accuracy rose 58.1 to 61.3 over accepted rounds; teacher control shows 27B teacher helps only when state-conditioned.. Measured on three benchmarks with one base model and one teacher pair; [truncated]
Disagreeing is the most useful thing you can do here. Nothing on this page waits for a reviewer: the counts above move when evidence lands, and every source stays attached with who brought it.
Related claims
- For LLM agents that score their own actions during exploration in text-based games and then reuse that experience, the accuracy of those self-scores against the environment verifier does not predict later gains on held-out stricter configurations: block-level correlations between judgment agreement or calibration and the next checkpoint's score change sit near zero, so recognizing a good action is not the bottleneck — converting it into a transferable policy is.Improving its own scaffolding from failures · reviewed by AI
- When training 4B-8B Qwen3 agents with group-based RL (GRPO, DAPO, GSPO) on sparse-reward long-horizon tasks, injecting hints into the environment's feedback — procedural action guidance in early episode steps and early training, switching to enriched state observations in later steps and later training — raises success on the unmodified environments by a few points on average, while lowering performance on tasks that require questioning whether the request is answerable at all.Completing tasks that take many steps or many hours · reviewed by AI
- When RL-training a central-agent multi-agent harness with 1.5B-3B Qwen2.5 backbones on multi-hop QA, optimizing only the coordinating central policy and leaving sub-agents frozen scores higher than jointly training central and sub-agent policies (average F1 47.79 vs 45.36, accuracy 42.88 vs 40.49 at 3B).Getting more from several models than from one · reviewed by AI
- When a calculator-using model is trained with reinforcement learning on verifiable final-answer rewards for an arithmetic search task (Countdown), the gain lands almost entirely at low k — pass@1 rises sharply while pass@16 moves little or falls — because the update can only reinforce correct trajectories the starting policy already samples, and prompt groups with no correct sample supply no gradient.Using the tools it is given · reviewed by AI
- When training an LLM with policy-gradient RL on tasks that decompose into many dependent sub-problems, rewarding each trajectory segment by how many reference reasoning points it newly reaches beats a sparse final-answer reward by a margin that widens with task horizon, and still learns on problem sets where the base policy never samples a success and sparse reward gives no gradient at all.Completing tasks that take many steps or many hours · reviewed by AI
Notes
Drafted from the paper by a model and filed unreviewed. Visible here so it can be read, not because anyone has vouched for it: it does not move any technique's standing and does not count toward the internal scorecard. Drafted confidence: low. Falsifier as drafted: Under matched training compute, distilling all state-conditioned successful trajectories matches or beats transition-filtered distillation, or the state-free policy stops improving across accepted rounds on independent task families.