CMU's SSR turns agent reasoning into selection, cutting per-turn latency 90%+
CarnegieMellonU · hf · 2026-10-07
CMU released Selection-based Structured Reasoning (SSR). For small models whose free-form reasoning is long, costly, and barely guides actions, SSR reformulates reasoning as selection: recurring high-level reasoning is pre-specified as reusable natural-language candidates, and the model picks one per turn based on context likelihoods, with no auxiliary task head. Pre-specified traces enable parallel scoring via teacher-forced prefilling with a shared KV cache.
On 7 multimodal search benchmarks with 2B/4B models, SSR yields large efficiency gains across RL objectives and SFT: success rates match leading same-scale search agents while cutting per-turn reasoning latency by over 90% and total per-question latency by 28-54%.
More from coding & agent
- Survey maps in-parameter memory methods for LLMs by placement and acquisition time — _reachsumit · 2026-10-07
- Ill-specified real work means harnesses still matter, argues Paras Chopra — paraschopra · 2026-10-07
- Agentic AutoRAG: LLM Agents Diagnose Retrieval vs Generation Failures to Tune RAG Pipelines — _reachsumit · 2026-10-07
- Cursor adds Cloud Agents API endpoints for environment builds with status and error codes — tetsuoai · 2026-10-07
- Figure CEO: filling Vietnam's brutal visa form was our AGI test — now an agent passed it — adcock_brett · 2026-10-07
- Vite+ 1.1 released: 24% faster vp dev startup, 30% less memory, clearer prompts — irvinebroque · 2026-10-07