Research map / Large language models / Reasoning and chain of thought
Batched Contextual Reinforcement
Bangji Yang, Hongbo Ma, Jiajun Fan, Ge Liu
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Batched Contextual Reinforcement (BCR), a minimalist training approach aimed at reducing chain-of-thought token consumption in LLMs without the quality loss or pipeline complexity of prior efficiency methods.
Problem
Chain-of-Thought reasoning in LLMs consumes excessive tokens, inflating inference cost; existing efficiency methods (explicit length penalties, difficulty estimators, multi-stage curricula) either degrade reasoning quality or require complex training pipelines.
Approach
A minimalist reinforcement approach called Batched Contextual Reinforcement that is described as avoiding explicit length penalties, difficulty estimators, and multi-stage curricula. Details beyond the abstract are not available.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train with batched multi-problem contexts (BCR) vs JustRL, token usage | JustRL-deepseek-1.5B · AIME25 (N=1) | tokens per problem: 8,482 -> 3,173 (-62.6%) | Improved |
| BCR on JustRL accuracy AIME25 | JustRL-deepseek-1.5B · AIME25 (N=1) | accuracy (%): 40.0 -> 33.3 | Hurt |
| BCR on JustRL accuracy MATH-500 | JustRL-deepseek-1.5B · MATH-500 (N=1) | accuracy (%): 91.4 -> 87.6 | Hurt |
| BCR on JustRL accuracy AMC23 | JustRL-deepseek-1.5B · AMC23 (N=1) | accuracy (%): 85.0 -> 87.5 | Improved |
| BCR on Qwen3-4B-Thinking accuracy AIME25 | Qwen3-4B-Thinking-2507 · AIME25 (N=1) | accuracy (%): 70.0 -> 83.3 | Improved |
| BCR on Qwen3-4B-Thinking tokens AIME25 | Qwen3-4B-Thinking-2507 · AIME25 (N=1) | tokens per problem: 20,773 -> 17,498 (-15.8%) | Improved |
| concurrent problems N=1 to N=5 at inference (baseline) | JustRL-1.5B · AMC23 | accuracy (%): 85.0 -> 22.5 | Hurt |
| concurrent problems N=1 to N=5 at inference (BCR) | BCR · AMC23 | accuracy (%): 87.5 -> 50.0 | Hurt |
| increase concurrent problems N from 1 to 5 | BCR · AMC23 | tokens per problem: 2,637 -> 1,013 | Improved |
| train with group size N=3 instead of N=4 (ablation) | BCR · AMC23 (N=1 eval) | accuracy (%): 77.5 -> 87.5 | Improved |
| train with group size N=4 instead of N=3 (ablation) | BCR · AIME25 (N=1 eval) | accuracy (%): 30.0 -> 33.3 | Improved |
| train with group size N=3 instead of N=5 (ablation) | BCR · Minerva (N=1 eval) | accuracy (%): 41.9 -> 46.3 | Improved |
Closest papers by meaning
- T1: Advancing Language Model Reasoning through Reinforcement Learning and Inference Scaling (ICML 2025 · similarity 0.91)
- When More is Less: Understanding Chain-of-Thought Length in LLMs (ICLR 2026 · similarity 0.91)
- Resource-Efficient Reinforcement for Reasoning Large Language Models via Dynamic One-Shot Policy Refinement (ICML 2026 · similarity 0.91)
- Attention Illuminates LLM Reasoning: The Uncovered Preplan-and-Anchor Rhythm Enables Fine-Grained Policy Optimization (ICML 2026 · similarity 0.91)
- Rethinking Thinking Tokens: LLMs as Improvement Operators (ICML 2026 · similarity 0.91)
- The Overthinker's DIET: Cutting Token Calories with DIfficulty-AwarE Training (NeurIPS 2025 · similarity 0.91)
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.
