Skip to main content
Search

# Outcome-based Reinforcement Learning

3 related articles

Subscribe to updates
  1. arXiv

    Reinforcement learning teaches a language model to search for its own evidence, beating Claude Opus 4.5 on 265 questions at about 5% inference cost

    The work builds an agentic forecasting environment, dataset, and harness from 2,100+ resolved Polymarket questions with layered leak filtering, letting the agent acquire its own context at rollout time via web search, page reading, and financial time series, and trains Qwen3.5-35B-A3B (3B active parameters) with single-epoch GRPO under a Brier-score reward; training improves calibration by 30-40% and cuts search attempts from 3.8 to 2.25 per rollout, and in an identical harness the trained policy finishes ahead of every frontier model tested at evidence-based forecasting, including Claude Opus 4.5 (soft-Brier 0.254 vs. 0.256, n=265), at about 5% of the inference cost, with the widest margin on the hardest questions the crowd itself had not decided.

  2. arXiv

    SWiM distills working memory so multimodal LLMs improve reasoning segmentation even without memory at inference, reaching state-of-the-art on benchmarks

    The work finds that multimodal large language models benefit from using their own self-generated reasoning traces and localization proposals as working memory when revisiting the same image and query, and proposes SWiM: rollouts are selected by segmentation quality to build working memory, the memory-conditioned model serves as a teacher giving token-level distributional supervision along student-generated trajectories while the student receives only the original image and query, and on-policy self-distillation is jointly optimized with outcome-based reinforcement learning, achieving state-of-the-art performance on reasoning segmentation benchmarks.

  3. arXiv

    SUSTech and CityU propose Box2-Bench: frontier models use good workflows but lose up to 43.3 points under bad ones

    The authors introduce Box2-Bench, which holds the model and task fixed while varying workflow availability and reliability across five matched conditions (none, good, bad, partial, mixed), finding that frontier models such as Gemini 3.7 Flash, DeepSeek V4 Flash, and GLM 5.2 benefit from good workflows in 8 of 12 model-task pairs but are hurt by bad workflows in all 12 (drops of 3.3 to 43.3 points), and that mixed workflows underperform their matched partial workflows in 10 of 12 pairs; training two open-weight models on bad workflows only, counterfactual SFT cuts the bad-workflow penalty from 20.0 to 6.7 points on AIME and from 9.8 to 0.6 points on WebShop while weakening use of held-out good workflows, and outcome-based RL restores AIME utilization from -3.3 to +6.

You’re all caught up