Research map / Large language models / Reasoning and chain of thought

CoRe: Collaborative Reasoning via Cross Teaching

ICML 2026 · Large language models › Reasoning and chain of thought · approach: Efficient and adaptive thinking · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 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 CoRe, a training-time cross-teaching framework where small LLM pairs share correct peer traces as hints for failed collaborators, trained with a combined reward (correctness, DPP-lite diversity, rescue bonus, cross-model complementarity) and evaluated via oracle Team Pass@2.

Problem

Small LLMs make complementary, correlated reasoning errors; inference-time ensembling does not change the policy, and training-time collaboration needs dense credit assignment and must avoid diversity collapse.

Approach

A two-round micro-protocol: cold independent sampling, then a contexted rescue round where failed models receive a capped hint from a successful peer's trace, combined with a rescue bonus, a greedy DPP-lite within-instance diversity reward, and a cross-model complementarity term, optimized with group-normalized policy gradients.

Key results

• On GSM8K with 1,000 training examples, the Qwen2.5-3B + Qwen3-4B team reaches 99.54% oracle Team Pass@2 versus 87.20% for the SD-E^2 oracle pair.
• On MATH, the Phi-4-mini + Ministral-3B team reaches 92.08% oracle Team Pass@2 versus 77.70% for the SD-E^2 oracle pair.
• On AIME, a 3-model non-oracle majority vote reaches 81.20%, versus 78.45% for the best single model; the oracle is 82.60%.

Surprising findings

• Removing the cross-model complementarity reward has a larger effect on MATH/AIME/GPQA than on GSM8K, where performance is near-saturated.
• Scaling from 2 to 3 models on AIME improves oracle Team Pass@2 by only 2.95 points, and the scaling result is preliminary.

Limitations

• Headline metrics are oracle Team Pass@K, which assume perfect answer selection; non-oracle results are only given for a 3-model AIME setting.
• Teacher context is capped at 1536 tokens and can truncate the decisive step; the hybrid distance can mis-score diversity.
• Possible contamination on AIME and the test split is small (AIME 193, GPQA 90 examples), and the paper states that the comparison to MALT/MAPoRL is not controlled.

Experimental setup

Datasets: GSM8K MATH AIME GPQA

Benchmarks: GSM8K MATH AIME GPQA

Metrics: Pass@1 Pass@2 oracle Team Pass@2 majority vote accuracy agreement-gated selection accuracy

Baselines: Base (off-the-shelf) SD-E^2 SD-E^2 + Oracle MALT MAPoRL Qwen2.5-7B-Instruct Phi-3-small-8k-Instruct Ministral-3-8B-Reasoning

Builds on: GRPO GSPO SAPO SD-E^2 DPP LoRA

Closest papers by meaning

• Collaborative Reasoning Distillation via Cross-Feedback and Coherent Curation (NeurIPS 2026 · similarity 0.91)
• The Quest for Efficient Reasoning: A Data-Centric Benchmark to CoT Distillation (ICLR 2026 · similarity 0.88)
• Exploring the Limits of Sub-Billion Language Model Reasoners with Open Training Recipes (ICLR 2026 · similarity 0.88)
• PRISM: Demystifying Retention and Interaction in Mid-Training (ICML 2026 · similarity 0.88)
• Rethinking Fine-Tuning when Scaling Test-Time Compute: Limiting Confidence Improves Mathematical Reasoning (NeurIPS 2025 · similarity 0.87)
• SuperCorrect: Advancing Small LLM Reasoning with Thought Template Distillation and Self-Correction (ICLR 2025 · similarity 0.87)