Research map / Reinforcement learning / RL from feedback
Expected Return Causes Outcome-Level Mode Collapse in Reinforcement Learning and How to Fix It with Inverse Probability Scaling
ABHIJEET SINHA, Sundari Elango, Dianbo Liu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace GRPO reward with inverse-probability-scaled reward (IPS-GRPO) | GRPO on Gemma-3 4B · HypoSpace causal inference | recovery rate (%): 16.02% -> 43.91% | Improved |
| replace GRPO reward with inverse-probability-scaled reward (IPS-GRPO) | GRPO on Gemma-3 4B · HypoSpace 3D reconstruction | recovery rate (%): 31.61% -> 90.00% | Improved |
| replace GRPO reward with inverse-probability-scaled reward (IPS-GRPO) | GRPO on Gemma-3 4B · HypoSpace Boolean/DNA | recovery rate (%): 9.50% -> 60.74% | Improved |
| replace GRPO with IPS-GRPO | GRPO · Hyper-grid n=2 | L1 distance: 2.8 × 10−2 -> 3.0 × 10−3 | Improved |
| replace GRPO with IPS-GRPO | GRPO · Hyper-grid n=4 | L1 distance: 5.2 × 10−4 -> 3.7 × 10−4 | Improved |
| use IPS-GRPO as policy learner in REINVENT | REINVENT on SYNTH · SYNTH threshold 0.85 | Yield: 418 -> 1580 | Improved |
| use IPS-GRPO as policy learner in REINVENT | REINVENT on ALL-AMIDE · ALL-AMIDE threshold 0.80 | Yield: 1 -> 1478 | Improved |
| use IPS-GRPO as policy learner in REINVENT | REINVENT on ALL-AMIDE · ALL-AMIDE threshold 0.85 | OB100: > 5000 -> > 5000 | No clear effect |
| increase group size G from 4 to 16 (eps 0.2) (ablation) | IPS-GRPO · HypoSpace causal inference | recovery rate (%): 17.63 -> 43.91 | Improved |
| increase group size G from 16 to 32 (eps 0.2) (ablation) | IPS-GRPO · HypoSpace causal inference | recovery rate (%): 43.91 -> 25.62 | Hurt |
| increase clipping epsilon from 0.01 to 0.2 at G=8 (ablation) | IPS-GRPO · HypoSpace causal inference | recovery rate (%): 27.50 -> 41.99 | Improved |
| increase clipping epsilon from 0.01 to 0.2 at G=32 (ablation) | IPS-GRPO · HypoSpace causal inference | recovery rate (%): 32.13 -> 25.62 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- FlowRL — FlowRL: Matching Reward Distributions for LLM Reasoning (ICLR 2026) · better on 4 of 4 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- ContraDiff: Planning Towards High Return States via Contrastive Learning (ICLR 2025 · similarity 0.87)
- Delightful Distributed Policy Gradient (NeurIPS 2026 · similarity 0.87)
- Beyond Expected Return: Accounting for Policy Reproducibility When Evaluating Reinforcement Learning Algorithms (AAAI 2024 · similarity 0.87)
- Intrinsic Benefits of Categorical Distributional Loss: Uncertainty-aware Regularized Exploration in Reinforcement Learning (NeurIPS 2025 · similarity 0.86)
- Overestimation, Overfitting, and Plasticity in Actor-Critic: the Bitter Lesson of Reinforcement Learning (ICML 2024 · similarity 0.86)
- STARE: Surprisal-Guided Token-Level Advantage Reweighting for Policy Entropy Stability (NeurIPS 2026 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
