Residual Drift Dominates Contradiction in Multi-Turn Constraint Reasoning
The paper discusses how multi-turn reasoning systems often fail not due to logical contradictions, but rather due to a phenomenon called satisfiable drift. This drift occurs when the system's internal state remains consistent while the answers it provides violate previous commitments. The study introduces a benchmark called DRIFT-Bench and evaluates various methods, finding that residual errors predominantly manifest as satisfiable drift rather than contradictions.
- ▪The dominant failure mode in multi-turn reasoning systems is satisfiable drift, not logical contradiction.
- ▪The study introduces DRIFT-Bench, a benchmark with 816 test problems across three constraint domains.
- ▪MUS-Repair is identified as the strongest method, improving performance significantly over non-MUS baselines.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Artificial Intelligence arXiv:2605.23940 (cs) [Submitted on 28 Apr 2026] Title:Residual Drift Dominates Contradiction in Multi-Turn Constraint Reasoning Authors:Sebastien Kawada View a PDF of the paper titled Residual Drift Dominates Contradiction in Multi-Turn Constraint Reasoning, by Sebastien Kawada View PDF HTML (experimental) Abstract:How do multi-turn reasoning systems fail? The expected answer is logical contradiction, in which the system's maintained state becomes unsatisfiable. We show that the dominant mode is instead satisfiable drift, where the internal state stays consistent while the returned answer silently violates prior commitments.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
