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

When to Think, When to Speak: Learning Disclosure Policies for LLM Reasoning

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 Side-by-Side (SxS) interleaved reasoning, where an LLM learns when to disclose entailment-supported partial answers while continuing private reasoning, trained with entailment-aligned SFT and GRPO RL on Qwen3 models.

Problem

Single-stream autoregressive LLMs couple private deliberation with public commitment, so longer reasoning delays task-relevant visible content (a "silence tax"), while early streaming risks premature or unsupported disclosure.

Approach

Makes disclosure timing a learnable think/speak action within one token stream, building SFT trajectories by aligning answer blocks to reasoning prefixes that entail them (via an LLM entailment checker), then using GRPO to recover accuracy under the interleaved format.

Key results

• Qwen3-4B SxS RL Final reaches 80.0% on AIME25 vs 73.8% for Standard CoT RL Final, with AIRW 8,519 vs 21,316 tokens.
• Qwen3-4B SxS RL Final scores 49.3% on GPQA-Diamond vs 19.0% for Standard CoT RL Final (base 55.9%).
• Qwen3-30B-A3B SxS RL Final reaches 79.2% on AIME25 vs 80.6% for Standard CoT RL Final, with AIRW 13,829 vs 16,709.

Surprising findings

• Standard CoT RL on math drops GPQA-Diamond accuracy for Qwen3-4B from 55.9% to 19.0%, while the SxS variant retains 49.3%.
• SFT alone on the interleaved format drops AIME25 accuracy sharply (30B-A3B to 50.8%, 4B to 38.3%), requiring RL to recover.

Limitations

• Entailment alignment relies on an expensive 120B-class checker and can unlock segments too early or late.
• Latency is measured via token-index proxies, not real user-perceived wall-clock latency.
• Outcome-only RL gives no explicit incentive to keep interleaving; objective uses simple structural proxies.

Experimental setup

Datasets: DeepMath OpenMathReasoning OpenThoughts DAPO

Benchmarks: AIME25 GPQA-Diamond LiveCodeBench KOR-Bench

Metrics: accuracy pass@1 ARI ABO AIRW

Baselines: Standard CoT (SFT) Standard CoT (RL Final) Base Model

Builds on: GRPO Chain-of-Thought ReAct DAPO

Closest papers by meaning

• THE PATH OF LEAST RESISTANCE: GUIDING LLM REASONING TRAJECTORIES WITH PREFIX CONSENSUS (ICLR 2026 · similarity 0.86)
• Attention Illuminates LLM Reasoning: The Uncovered Preplan-and-Anchor Rhythm Enables Fine-Grained Policy Optimization (ICML 2026 · similarity 0.86)
• The Tell-Tale Norm: $\ell_2$ Magnitude as a Signal for Reasoning Dynamics in Large Language Models (ICML 2026 · similarity 0.86)
• $\textbf{Re}^{2}$: Unlocking LLM Reasoning via Reinforcement Learning with Re-solving (ICLR 2026 · similarity 0.86)
• Self-Distilled Reasoner: On-Policy Self-Distillation for Large Language Models (ICML 2026 · similarity 0.86)
• RFEval: Benchmarking Reasoning Faithfulness under Counterfactual Reasoning Intervention in Large Reasoning Models (ICLR 2026 · similarity 0.86)