Reliable LLM reasoning
Telling correct reasoning from errors using how a model's internal representations change across layers.
Papers: Truth as a Trajectory (ACL 2026, oral, Best Paper Award nominee) · Reasoning Errors Have a Region and a Direction in the Residual-Stream Trajectory of LLMs (preprint, 2026) · Project page
The problem
Language models can produce reasoning that reads well but is wrong. Judging an answer from its text alone tends to pick up superficial lexical cues rather than whether the reasoning is valid.
The approach
Instead of looking at a single layer, we treat a response as a trajectory through the model’s layers and use how its internal representations change across them to distinguish valid reasoning from errors.
The follow-up work combines these changes with coarse regions and directions in the model’s internal state space.
Results
- Truth as a Trajectory reduces reliance on superficial lexical cues and outperforms conventional probes across several benchmarks.
- The follow-up improves answer selection on unseen reasoning benchmarks and generalises to factual tasks.
- Truth as a Trajectory was presented as an oral at ACL 2026 and nominated for the Best Paper Award.