Search papers, labs, and topics across Lattice.
This paper addresses the challenge of distinguishing sound reasoning from flawed reasoning in language models by proposing a three-stream detector that integrates motion with two restricted views of location. By employing a coarse region reader based on vector quantization and a fine direction reader over normalized multi-layer states, the method effectively balances the trade-off between retaining contextual information and avoiding shortcut-prone data. The proposed approach achieves significant improvements in selection accuracy on reasoning benchmarks, outperforming existing methods by up to 21%, indicating that reasoning validity is more effectively assessed through state-conditioned motion rather than static or decontextualized representations.
Reasoning validity in LLMs can be more accurately assessed through a novel three-stream detector that integrates motion with contextual state information, achieving up to 21% better accuracy than existing methods.
As language models are increasingly used for tasks that require verifiable reasoning, reliably distinguishing sound reasoning from flawed reasoning has become an important practical problem. Recent trajectory-based methods seek this signal in layerwise residual-stream displacements, which capture how representations change while attenuating some stable, token-specific information. However, displacement omits the state from which an update originates, whereas restoring the full state risks reintroducing shortcut-prone information. We identify this trade-off and propose a three-stream detector that combines motion with two restricted views of location. A coarse region reader based on vector quantization and a fine direction reader over normalized multi-layer states. This design restores enough state context to interpret the motion without returning to full-state probing. On reasoning benchmarks unseen during training, our method improves selection accuracy by up to 12% over the displacement-only state of the art and 21% over single-layer probing baselines. Although trained only on reasoning benchmarks, it also reads factual completion and fact verification, ahead of every detector we compare against, which places the signal on correctness rather than on a kind of reasoning. Ablations further show that motion, region, and direction provide complementary signals. These results suggest that reasoning validity is better read from state-conditioned motion than from either static states or decontextualized trajectories alone.
