acceptodds
Desk rejected at ICLR 2027

VIGIL: Candidate-Conditioned Pool Depletion for Conversational Passage Selection

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Committing a passage changes the choices that remain: a candidate useful now can exhaust every nonredundant, entity-consistent, and topic-coherent alternative in the current retrieval pool. Conventional ranking utility and reactive rule filtering do not score this candidate-conditioned depletion. VIGIL does so by branching on every possible first commitment, evolving a typed-graph state, and scoring a three-step greedy trajectory that removes candidates only from the observed pool. Constraint traps separately mark real turns with no rule-compatible candidate, keeping the commitment-time score distinct from prediction of a future query or pool. With all learned heads fit on CAsT 2021, VIGIL raises CAsT 2022 rank-one commitment utility from 0.441 to 0.479 and lowers information-need-macro commitment violation rate (Q-CVR) from 22.3% to 8.1%; against SlackPreserve-H3, matched on pool, rules, horizon, and latency accounting, it reduces Q-CVR by 1.3 pp [0.2,2.4] () at comparable utility ( ). OR-QuAC reproduces changes of +0.009 [0.001,0.017] in utility and -2.1 pp [-3.8,-0.4] in Q-CVR, while an independent blinded census records bad continuations in 12.0% (9/75) of VIGIL outputs versus 40.0% (30/75) for E5-large. Residual pool quality is therefore a measurable and optimizable part of conversational retrieval, even when retrieval exposure is held constant.

Primary area:
reinforcement learning
Keywords:
world models and rl, VIGIL

Related papers

Discussion (0)

Sign in to comment.