Neura Market
When to Think, When to Speak: Learning Disclosure Policies for LLM Reasoning (May 2026) logo

When to Think, When to Speak: Learning Disclosure Policies for LLM Reasoning (May 2026)

Free

Side-by-Side (SxS) Interleaved Reasoning — makes disclosure timing a controllable decision in autoregressive generation; interleaves partial disclosures with continued private reasoning, releasing content only when supported by reasoning so far; improves accuracy–latency Pareto trade-offs on Qwen3-3

FreeFree tier
Inputs: textOutputs: text
Type
Open Source
Visit Website GitHub (40 stars)

About When to Think, When to Speak: Learning Disclosure Policies for LLM Reasoning (May 2026)

This paper introduces Side-by-Side (SxS) Interleaved Reasoning, a method that makes disclosure timing a controllable decision within standard autoregressive generation for large language models. SxS interleaves partial disclosures with continued private reasoning in the same context, releasing content only when it is supported by the reasoning so far. The approach addresses the 'silence tax' of single-stream autoregressive interfaces, where additional deliberation postpones task-relevant content and naive early streaming risks premature commitments. SxS is trained using supervised fine-tuning (SFT) to acquire dual-action semantics and reinforcement learning (RL) to recover reasoning performance. Evaluated on Qwen3-30B-A3B (MoE) and Qwen3-4B (dense) models, the method improves accuracy–content-latency Pareto trade-offs on in-domain (AIME25) and out-of-domain (GPQA-Diamond) benchmarks. The paper was accepted at ICML 2026.

Key Features

Side-by-Side (SxS) Interleaved Reasoning: makes disclosure timing a controllable decision in autoregressive generation
Interleaves partial disclosures with continued private reasoning in the same context
Releases content only when supported by reasoning so far
Trained with supervised fine-tuning (SFT) for dual-action semantics and reinforcement learning (RL) for reasoning performance
Improves accuracy–latency Pareto trade-offs on Qwen3-30B-A3B and Qwen3-4B models
Evaluated on AIME25 (in-domain) and GPQA-Diamond (out-of-domain) benchmarks

Pros & Cons

Pros
  • Addresses the 'silence tax' by allowing controlled disclosure timing
  • Reduces risk of premature commitments that bias subsequent generations
  • Demonstrates improved accuracy–latency Pareto trade-offs across multiple benchmarks and model scales
  • Method is compatible with standard autoregressive generation and does not require architectural changes
Cons
  • Requires additional training with SFT and RL, increasing computational overhead
  • Complexity of interleaved reasoning may be challenging to implement in production systems
  • Limited evaluation to Qwen3 models; generalizability to other LLMs not yet shown

Best For

Improving LLM reasoning accuracy while controlling latency in autoregressive generationReducing premature commitments in step-by-step reasoning tasksOptimizing the balance between deliberation and disclosure in conversational AI

FAQ

What is Side-by-Side (SxS) Interleaved Reasoning?
SxS is a method that makes disclosure timing a controllable decision in autoregressive LLM generation. It interleaves partial disclosures with continued private reasoning, releasing content only when supported by the reasoning so far.
Which models were used for evaluation?
The method was evaluated on Qwen3-30B-A3B (Mixture-of-Experts) and Qwen3-4B (dense) models.
What benchmarks were used?
AIME25 (in-domain) and GPQA-Diamond (out-of-domain) benchmarks.
How is SxS trained?
SxS is trained using supervised fine-tuning (SFT) to learn dual-action semantics and reinforcement learning (RL) to recover reasoning performance under the new format.