interwhen: ASU-Microsoft Research framework for test-time verification of reasoning traces accepted at NeurIPS
rao2z · x · 2026-09-25
The interwhen paper, a Microsoft Research India project involving Subbarao Kambhampati's group, has been accepted to NeurIPS. It extends the LLM-Modulo generate-test framework to verify reasoning processes rather than just final answers.
- Single-trajectory verification: periodically polls the reasoning trace and forks model inference to recover intermediate states, avoiding costly multi-branch exploration
- Asynchronous, near-zero overhead: verifiers run alongside generation and intervene only on policy violations
- Automatic verifier synthesis: generates verifiers from natural-language policy documents, addressing verifier scarcity beyond math and code
The core claim: steering reasoning models requires checking the process itself, and interwhen offers a generalizable way to do it at test time.
More from Research
- Neural Spectral Capacity: a training-free architecture score that beats #Params and prunes LLaMA-7B 5900x faster — CityU-HongKong · 2026-09-25
- OpenAI reportedly spent $10M in a week on a Navier-Stokes counterexample — funding 50 mathematicians for a year — RexDouglass · 2026-09-25
- Economists call for zero tolerance on AI-written papers: reject, retract, and ban authors — paulnovosad · 2026-09-25
- HRI 2027 Adds Archival Industry White Paper Track for Real-World Robot Deployments — petitegeek · 2026-09-25
- Study: Reasoning hurts 15.7% of multimodal embeddings; training-free SURE router fixes it — _reachsumit · 2026-09-25
- Google scales learned cross-task relationships in YouTube's production recommender — _reachsumit · 2026-09-25