Skip to main content
Search

# Agentic Reinforcement Learning

22 related articles

Subscribe to updates
  1. arXiv

    Multi-turn RL pushes LLM agents past the base model only with roomy budgets, and the edge fades or reverses when turns tighten

    Multi-turn agentic reinforcement learning is widely used to strengthen the long-horizon behavior of large language model agents, yet trajectory-level success rates hide what it does to individual decisions. Across four interactive tasks, DeepCoder-MT, BFCL, InterCode-SQL and FrozenLake, models trained with GRPO are evaluated along two axes, sampling breadth and interaction depth: with more sampled trajectories the RL model consistently beats its base model, but when the turns allowed per trajectory are tightened that edge shrinks, disappears, and on some tasks reverses. Matched-state comparisons show better next-step actions on only 46.8% of DeepCoder-MT states and 42.8% of FrozenLake states, with the gain concentrated in error recovery and later trajectory stages.

  2. arXiv

    A $20 reward function can strip a model's safety alignment; Bait probes the reward to block the learning

    Agentic reinforcement-learning-as-a-service hands data sampling and reward rating to the user, and a harmful reward function can erase a model's safety alignment at a cost as low as $20. Bait mixes explicit harmful responses from a separately trained Bait model into the policy model's response group during rollout, uses the sign of the reward gap between Bait and non-Bait responses (reward divergence) to flag a harmful reward function, and flips that group's reward sign to penalize learning when it detects one. On the authors' testbed, Bait cut the average harmful score across three harmful-instruction datasets from 20.4 to 6.8 while average agentic accuracy rose from 93.1 to 96.0.

  3. arXiv

    Tool calls get worse with training: TACA supervises the reasoning-to-call transition and lifts six-benchmark accuracy by 4.61 points

    In multi-turn tool-use reinforcement learning, reasoning accuracy can rise early and then fall below the non-agentic baseline; TACA traces this to negative aggregate credit on tool-call tokens and instead supervises the step where reasoning turns into a tool call. Against MT-GRPO it raises mean accuracy over six reasoning benchmarks by 4.61 and 3.61 percentage points for Qwen2.5-1.5B and 3B, while tool execution success during training stays at 86.28%.

  4. arXiv

    Long-horizon agents should bank credit per turn: MPO splits turn-level advantage from token refinement, lifting search pass@1 by up to 4.33 points

    Long-horizon LLM agents interact with their environment through complete turns but generate each action token by token, and existing reinforcement learning either gives a whole trajectory one shared advantage or treats every token as an action. MPO splits credit assignment into two scales: turn-level generalized advantage estimation carries credit across interactions, a mean-preserving local correction refines it within each turn, and the policy update uses complete-turn importance ratios. Across search, tool-use and automation tasks, MPO reports the highest score on all nine metrics at both Qwen3-4B-Thinking and Qwen3-30B-A3B-Instruct scales, including a 4.33-point BrowseComp-Plus pass@1 gain over the strongest non-MPO baseline at 4B.

  5. arXiv

    A 35B verifier stops guessing verdicts: it writes its own counterexamples and lifts judging accuracy from 71% to 87%

    A 35B agentic verifier that sees only the problem statement and a candidate program, never the official tests, uses self-generated counterexamples as executable evidence to raise binary verdict accuracy on USACO-Judge from 71.41% to 87.41%.

  6. arXiv

    Finer process rewards are not automatically better: decaying the weight lifts long-horizon trailer editing by 45.4%, while a fixed weight truncates 44.1% of rollouts

    On four trailer-editing task types built from 49 real films and their released trailers, decaying an LLM-judged turn-level process reward to zero within the same GRPO setup gives a 45.4% relative gain over outcome-only GRPO on Fine Cut, while the fixed-weight variant scores below outcome-only GRPO on the three shorter task types and truncates 44.1% of rollouts. Process rewards do encourage visual inspection and reasoning early, but continued optimization drives redundant tool calls without a committed result; token-level on-policy self-distillation proves highly sensitive to the teacher's privileged information.

  7. arXiv

    Professional motion editing stops guessing from one instruction: an agent finds a demo, self-checks it, and Stoppie no longer becomes Wheelie

    Professional motion video editing now has an agent retrieve and self-check a demonstration clip, then hand it with a grounded prompt to a frozen editor; across 314 professional motion tasks, Seedance 2.0 rises from 55.1 to 65.6 and MiniMax-H3 from 51.2 to 64.5.

  8. A 24B reasoning model that activates only 2B parameters per token leads or matches on five of eight tool-use benchmarks

    Liquid AI released LFM2.5-24B-A2B, a reasoning agent model with 24B total parameters but only 2B active per token, built to run locally on device. It keeps the predecessor LFM2-24B-A2B architecture while growing context from 32K to 131K tokens and vocabulary from 65K to 128K, and it now emits an explicit chain of thought before answering. Across 16 benchmarks it competes with Qwen3.6-35B-A3B, which has roughly 50% more total and active parameters: 48.09 versus 38.19 on LiveClaw and 27.69 versus 8.72 on AutomationBench, while decoding 184 tokens/s on an M5 Max.

  9. arXiv

    Agent RL stops relying on its own luck: failed rollouts get corrected mid-sampling, lifting navigation pass rate from 35% to 57%

    Alibaba researchers introduce Rollout-Time Intervention, which corrects the decisive error step using reward-side rubrics whenever an entire rollout group fails, letting agent RL learn from successful trajectories the policy could not explore on its own and raising Navigation Pass@1 from GRPO's 35.33 to 57.33.

  10. arXiv

    Skill libraries stop only growing: SkillForge retires and rewrites skills as the policy evolves, lifting WebShop success by 7.8% relative

    SkillForge turns an agent's skill library from an append-only store into a fitness-driven lifecycle of trial, active, stable, and retired states, so skills and policy co-evolve during reinforcement learning and reach the highest aggregate success rate across ALFWorld, WebShop, and Search-Augmented QA, up to 7.8% relative over the strongest baseline.