Xiaohongshu proposes CASI: Shapley-based credit assignment for agentic retrieval
_reachsumit · x · 2026-10-15
Xiaohongshu researchers published CASI on arXiv, tackling credit assignment when LLM agents jointly generate multiple retrieval actions but receive only one set-level reward.
- Builds counterfactual signal via snapshot-grounded set interventions, then aggregates marginal contributions with Shapley values
- Shows that for m>=4, singleton/leave-one-out/full-set endpoints don't determine the Shapley allocation; under additive set values, extra coalition evals are redundant
- Exhaustive auction replay over 1,200 requests in a five-action sponsored-search setting reveals strong coalition dependence; CASI achieves the highest median cosine similarity to a replay-defined Shapley reference
- Caveat: under a sensitivity backend, CASI doesn't outperform Singleton
More from coding & agent
- SWEeper-Bench: Best of 15 Frontier Agents Solves Only 59% of Interactive Bug-Finding Tasks — zlab-princeton · 2026-10-15
- User hands Google login and SMS 2FA to AI agent, reigniting credential-handoff debate — jonstephens85 · 2026-10-15
- Grok Bot writes its own Apple Desk skill to control Mac apps, Musk amplifies — elonmusk · 2026-10-15
- Thorsten Ball: latest models make context-management tricks like /btw and forking obsolete — andreisavu · 2026-10-15
- User says Grok Bot finished 10 days of work in under 2 hours on a $30 plan — elonmusk · 2026-10-15
- PersonalAgentEval launches: 7 personal agents tested on real 13-month histories — HennyJieCC · 2026-10-15