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

interwhen: A Generalizable Framework for Steering Reasoning Models with Test-time Verification

NeurIPS 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 interwhen, a single-trajectory test-time verification framework that forks the reasoning model to extract states, runs code/formal verifiers asynchronously (with verifiers auto-synthesized from policies, including Lean/z3), and steers the model on violations.

Problem

Test-time verification of long reasoning traces: final-answer-only verification misses early errors, while branch-and-verify search multiplies compute; verifiers for process/policy compliance are scarce outside math and code.

Approach

Decouple verification from generation: periodically fork the model to extract verifiable state variables from the partial trace, run verifiers asynchronously, and intervene with feedback only on violations; verifiers are auto-generated from natural-language policies (with Lean/z3 formal checks).

Key results

• On tau2-bench telecom solo mode, Qwen3-30B-A3B-Thinking pass^4 rises from 32.17% (CoT) to 87.70% with interwhen at 96.93% of CoT tokens.
• On non-agentic tasks, interwhen reaches 97.2% on GameOf24 at 61.6% of CoT tokens versus 96.1% for Generate-Test at 116.3% tokens.
• k-Stable early stopping reduces tokens by 32.2% on Maze and 55.1% on Verina-Spec for Qwen3-30B without accuracy loss.

Surprising findings

• interwhen often reduces token usage relative to CoT on agentic tasks, attributed to verifier feedback avoiding unproductive reasoning.
• A 3B SLM extractor matches a frontier GPT-5.4 extractor within 0.4 pp on tau2-bench.

Limitations

• Verifier completeness is not guaranteed; implicit policy rules may be missed.
• Strict verifiers on vague policies can over-block tool calls, lowering helpfulness on fulfillable tasks (AgentSafetyBench).
• Extraction by LLM can err, and soundness holds only modulo extraction errors.

Experimental setup

Datasets: tau2-bench AgentSafetyBench VitaBench Maze SpatialMap GameOf24 ZebraLogic Verina

Benchmarks: tau2-bench AgentSafetyBench VitaBench Maze SpatialMap GameOf24 ZebraLogic Verina

Metrics: accuracy pass^4 avg@4 reward safety score token %

Baselines: CoT Best-of-N critic Generate-Test critic Tree of Thoughts DEER EAT Claude Haiku 4.5 GPT-5.4 Mini GPT-5.4

Builds on: LLM-Modulo Tree of Thoughts Generate-Test with critic Best-of-N

Closest papers by meaning

• MAS-ProVe: Understanding the Process Verification of Multi-Agent Systems (ICML 2026 · similarity 0.88)
• What If We Allocate Test-Time Compute Adaptively? (ICML 2026 · similarity 0.88)
• V1: Unifying Generation and Self-Verification for Parallel Reasoners (ICML 2026 · similarity 0.87)
• Rewarding Progress: Scaling Automated Process Verifiers for LLM Reasoning (ICLR 2025 · similarity 0.87)
• Variation in Verification: Understanding Verification Dynamics in Large Language Models (ICLR 2026 · similarity 0.87)
• Scaling Code-Assisted Chain-of-Thoughts and Instructions for Model Reasoning (NeurIPS 2025 · similarity 0.86)