It is suggested that reasoning validity is better read from state-conditioned motion than from either static states or decontextualized trajectories alone, andlations show that motion, region, and direction provide complementary signals.
Abstract
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.
Vision–Language–Action (VLA) policies remain brittle under modest distribution shift. On LIBERO-Plus, contemporary models that solve clean tasks at high rates can fall below 30% success when the camera’s viewpoint or the robot’s initial pose is perturbed. Most training-free test-time remedies address this problem through the image stream, for example, by augmenting, purifying, or selecting visual observations. In our controlled evaluation, this family of methods improves mean success by only about three points and leaves the robot-initial-state failure largely unresolved. This paper studies the failure at the level of input streams. A VLA receives visual tokens, a proprioceptive state token, and language tokens; different perturbations can move different streams away from their training manifold. In particular, the robot-initial-state perturbation directly shifts the proprioceptive token; therefore, image-space interventions have limited leverage. We introduce Gated Per-Stream Manifold Restoration (G-PSMR), a training-free wrapper for a frozen policy. For each stream, a lightweight gate detects off-manifold inputs and applies a stream-specific restoration before the policy forward pass. We instantiate the framework with entropy-gated visual consensus and gated relative-orientation debiasing, which preserves the within-episode orientation trajectory. In the original 280-episode paired evaluation, the joint method improves total success by +5.3 points compared with a +3.2 image-only gain and raises the most fragile factor from 20% to 38%. On 1120 previously unevaluated, manifest-disjoint task instances, the state restoration improves robot-initial-state success from 23.8% to 28.7%; a separate prospectively specified confirmation on 600 new gate-active instances yields 26.3%→32.2% (+5.8 points; 95% CI [+3.2,+8.5]; p<0.001). Together, the original cross-stream results and two independent state-stream evaluations support the central principle of matching the restoration to the input stream carrying the shift.
The results suggest that VLAs have rich, linearly readable internal representations of semantic quantities like task progress, and that learning to read these signals offers a lightweight, interpretable path toward monitoring deployed visuomotor policies.
Atiksh Bhardwaj, E. W. Duan, Prithwish Dan et al.· 1 citation
This work instantiates AutoCRAT, a decoder-side controller for frozen backbones that operates over a discrete action space and updates control decisions only at semantic boundaries, improving stability while remaining responsive to the evolving reasoning process.
Han-Jun Luo, Qiu-Shi Liu, Jing-Yang Zhang et al.· 0 citations
This survey systematizes recent progress in tree-search-based reasoning, viewing inference as instance-specific optimization rather than decoding, and introduces a Unified Design Space spanning search topology, evaluation signals, and control dynamics to unify a fragmented literature.
Jiaqi Wei, Xiang Zhang, Yue-Jin Yang et al.· 0 citations
Chain-of-Thought (CoT) improves the reasoning ability of Large Language Models (LLMs) but incurs substantial computation and context costs. Existing methods either lose intermediate information through hard pruning or lack a principled criterion for continuous compression. We present A*-Thought-V2, a geometric dynamics of LLM guided framework that models CoT as a hidden-state trajectory and replaces hard deletion with an explicit-implicit interleaved latent architecture. After projecting question, step, and solution representations into a 3D PCA space, it measures alignment between each local transition and global question-to-solution direction. Aligned steps remain explicit text, whereas deviating steps are compressed into continuous latent tokens. Directional angles capture both local semantics and reasoning dynamics: small angles indicate direct execution and answer formation, while large angles more frequently involve checking, correction, and branch exploration; their temporal variation reveals exploration, convergence, and refinement stages. To train this architecture, we introduce stepwise embedding forcing, which pools each redundant step into a single latent embedding, and label forcing, which supervises that latent token with a soft multi-modal vocabulary distribution instead of a hard one-hot label. Experiments on Qwen3.5-9B and Qwen3.6-27B across six in-domain and out-of-domain benchmarks show that A*-Thought-V2 improves average accuracy by up to 2.6% while reducing response length by up to half, increasing Accuracy per Computation Unit by 2.29$\times$, and reducing preprocessing and training time by 94.6% and up to 80.3%, respectively. Representation analyses suggest that latent states form a compact region distinct from textual states, while higher entropy at latent-token positions reflects broader soft targets that encourage richer step-level feature learning.
Xiao-An Xu, Si-Yuan Liu, Shuo Wang et al.· 0 citations
This work formulate multi-turn reasoning as a hidden-state trajectory of the underlying LLM that is characterized via two complementary signals: temporal curvature that captures the directional consistency of turn-to-turn updates, and variance slope which measures the expansion or contraction of the exploration space.
Jie Liang, Zhengxin Yu, H. Nasiri et al.· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
