TL;DR
Multi-turn reasoning systems often fail not due to logical contradictions, but because of 'satisfiable drift', where answers violate prior commitments without inconsistency. To address this, DRIFT-Bench was developed, a benchmark with 816 test problems that evaluates various methods.
✦ Why It Matters
Engineers should implement validation mechanisms to ensure multi-turn reasoning systems respect prior commitments.
Key Takeaways
Full Summary
Multi-turn reasoning systems are expected to fail through logical contradictions, where their internal state becomes unsatisfiable. However, this research reveals that a more common failure mode is 'satisfiable drift', where the system's answers remain logically consistent but violate previous commitments.
To investigate this, DRIFT-Bench (Decomposing Reasoning Into Failure Types) was created, consisting of 816 test problems across three constraint domains. Four methods were evaluated, with MUS-Repair, which utilizes minimal unsatisfiable subsets to guide the generator, outperforming others by 1.8 to 15.0 percentage points.
Notably, after structured feedback, contradictions were nearly eliminated, but residual errors remained at 98-100% satisfiable drift. This highlights the importance of validating that returned answers align with the maintained state, suggesting that future systems should incorporate mechanisms to check for this drift.
Related