Apple & CMU's SSR cuts multimodal agent reasoning latency by 90%
jiqizhixin · x · 2026-10-25
Apple and Carnegie Mellon present SSR (ArXiv 2610.01892), which reformulates agent reasoning as selection from a library of six predefined strategies (image search, text search, crop/zoom, answer from evidence, read from image, proceed) instead of freeform generation. Parallel scoring via shared KV cache evaluates candidates concurrently. On seven multimodal search benchmarks with 2B/4B models, SSR matches leading same-scale search agents' success rates while cutting per-turn reasoning latency by over 90% and total per-question inference latency by 28–54%.
More from coding & agent
- Poll: Only 8.4% of X developers still hand-write over 20% of their code, vs 65% on Mastodon — charliermarsh · 2026-10-25
- $39/mo analytics tool gets open-sourced in 4 hours, then an AI agent deploys it for free — fekdaoui · 2026-10-25
- Open-source userscript brings back a table of contents for long ChatGPT chats — TonyHMeow · 2026-10-25
- Opus 5.5 reviews other models' code: deletes a lot without breaking anything — CtrlAltDwayne · 2026-10-25
- OpenUI: open-source generative UI standard uses a streaming language that saves up to 67% tokens vs JSON — tom_doerr · 2026-10-25
- Chip Huyen's free AI Engineering companion repo packs 10-chapter notes and real-world prompts — mdancho84 · 2026-10-25