Large Language Models (LLMs) as agents often fail to improve in new environments. We identify and characterize a failure mode we call exploration collapse: under reinforcement learning (RL) in environments whose states are unfamiliar to the policy, Pass@k, the probability that at least one of k sampled trajectories succeeds, drops markedly over training even as Pass@1 edges up, revealing increasingly brittle exploration; environments closer to the pretraining distribution show no such decline. We trace this collapse to weak grounding in environment states and dynamics, and study a simple remedy: explicitly teaching the agent to estimate the current state and predict its transitions before optimizing for reward. We instantiate it as SPA, an explore-then-exploit recipe that cold-starts the policy with a Self-Experience supervised finetuning (SFT) stage, collecting the model’s own interaction trajectories and supervising state and next-state prediction, and then runs standard RL. The resulting world model serves as a grounded initialization for RL rather than an inference-time planner. Across unseen environments, SPA consistently and substantially improves over vanilla RL: for example, it raises the Sokoban success rate from 25.6% to 59.8% on Qwen2.5-1.5B-Instruct, letting sub-3B models surpass a 20B baseline on these tasks. Controlled studies indicate that the gains track four factors: grounded state representations, explicit transition modeling, self-experience trajectories from a sufficiently strong exploration policy, and adequate coverage of transition data.

Direct Preference Optimization (DPO) is notoriously sensitive to attribution ambiguity: when chosen and rejected responses differ along multiple entangled dimensions, the model cannot identify which behavior the preference label teaches. We propose Criterion-Activated Preference Synthesis (CAPS), a criterion-first framework that turns explicit quality standards into attribution-verified preference pairs. CAPS synthesizes contexts and responses from a target criterion, then applies dual-signal scoring: a rubric-bound judge measures criterion compliance, while a blind general judge measures holistic quality. Two filters close the attribution loop: the Empirical Activation Test (EAT) keeps criterion-context pairs where satisfying the criterion improves blind quality, and the Attribution Ratio (AR) keeps pairs whose quality gap is dominated by the target criterion. Instantiated with general rubrics, medical rubrics, and Chinese legal statutes, CAPS shows strong data efficiency and domain transfer. With 20k general synthetic pairs, it outperforms 35k human-annotated HelpSteer3 pairs on IFEval, IFBench, AlpacaEval, and Creative Writing v3. In domain settings, CAPS improves HealthBench with 2k medical pairs and improves Qwen3-8B/14B by +4.9 to +6.6 points on legal benchmarks using only public statutes, without human preference labels. We additionally construct the CAPS-Legal preference dataset.