- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper proposes a multiscale predictive account of language-model reasoning using SchemaBench and residual-stream analyses. Procedurally generated narratives provide annotations for studying role binding, event structure, and reasoning progress. Clustering and low-dimensional readouts characterize local operations, schema variables, and control states, while fitted fields predict subsequent representations. A direction extracted near spontaneous termination is injected throughout continued generation to test its effect on stopping. Experiments on SchemaBench and traveling-salesman problems show decodable task states and shorter outputs under STOP steering, with no statistically significant accuracy change reported.
Critical strengths
- SchemaBench varies role disclosure, narrative order, and role binding while retaining structured annotations, enabling more targeted representational analyses than final-answer accuracy alone.
- The paper is well-written and easy to follow. The figures and tables are visually appealing and clearly presented.
Critical weaknesses
- The long-horizon results rely on separately fitted successor or horizon-specific predictors; iterating the one-step field reaches only about four primitives. This establishes useful predictability, but provides weaker support for a shared field coordinating computation across scales. The STOP direction is also extracted through a separate contrast of attention-head activations, rather than derived from the fitted field. Its causal effect therefore does not validate the field's predicted dynamics. Even allowing the stated approximation to Transformer computation, the link between the predictive fits and the proposed coordinating mechanism remains incomplete.
- The intervention timing uses information obtained from completed baseline traces. Appendix D selects 17 capped responses that have already stated the required schema-role assignments and resumes them at that retrospectively identified point. TSP injection similarly begins at fractions of the baseline trace length. The experiments therefore do not test whether the proposed latent monitor can decide when to intervene using only currently available information. This is central to the claimed control application, since stopping safely depends on recognizing completion rather than merely shortening a trace after an externally supplied trigger.
Questions for authors
- Are PCA, k-means, transition-based spectral merging, and all predictor choices fitted exclusively within each training fold? Since temporal transitions help define the primitive labels, how much future-primitive predictability remains against lexical and position-based controls under a fully held-out extraction pipeline?
- At the same intervention prefixes, how does STOP steering compare with directly requesting final-answer production or forcing the model's end-of-thinking marker? Could the authors also report confidence intervals for the paired accuracy differences, since nonsignificant tests alone do not establish that accuracy is preserved?
3 comments
Part 2
re: How does STOP steering compare with directly requesting final-answer production or forcing the model's end-of-thinking marker?
Thank you for the suggested additional baselines. We have included direct comparisons to “</think>” and “Final Answer:”, see Table 1 below. On the same 84 passages (Phi-4-mini-reasoning), the full reasoning plus answer ran to a median of 2,404 tokens without intervention. The STOP vector shortened it to 2,210 tokens at alpha .25 (9% reduction; significant only at offset 30, Wilcoxon p = .02) and to 1,961 at alpha .5 (6%, p < 1e-4), with all person-role pairs right in 976 of passages at alpha .25 and .976 at alpha .5, .976 for control (the only significant accuracy loss is alpha .5 at the solved run, McNemar p = .016). Inserting "Final answer:" gave 1,837 tokens (16% shorter than control, p < 1e-4) but at the solved run 24 of the 84 completions stopped without writing any answer, so its usable accuracy was .786. Inserting </think> was the shortest at 1,208 tokens (44% shorter than control, p < 1e-12; 34% shorter than "Final answer:", p < 1e-4), with accuracy .893. All pairwise length differences were significant (p ≤ .009). In total length the ordering is therefore </think> < "Final answer:" ≤ STOP alpha .5 < STOP alpha .25 ≤ control, and in accuracy </think> is 8 points below both (p = .039 against alpha .25).
Table 1 Additional Stop injection experiments at offset=30. Strict: both schema and role have to be correct.
| Condition | Total tokens | Tokens after | Role accuracy | All pairs right | Strict |
|---|---|---|---|---|---|
| Control | 2,404 | 1,148 | .991 | .976 | .905 |
| STOP alpha .25 | 2,210 | 990 | .995 | .976 | .917 |
| STOP alpha .5 | 1,961 | 742 | .995 | .976 | .929 |
| "Final answer:" | 1,837 | 662 | .985 | .940 | .786 |
</think> | 1,208 | 194 | .967 | .893 | .869 |
Direct prompt interventions are thus an effective way to stop reasoning and output a correct answer. This is consistent with prior findings (e.g., Wu et al., 2025): direct prompting commonly outperforms state-of-the-art steering methods. As the goal of our paper is not to identify the optimal steering approach, but demonstrate that it can reliably and causally terminate an ongoing reasoning trace, as shown in Figure 5b. Crucially, prompting and other standard latent interventions do not provide the same insight into the model's internal dynamics or the coordination between primitives and schemas that our framework offered.
As a final note: We would like to clarify that an earlier dataset of SchemaBench (version A) had 5-role prompts for Airport schema and 4-role prompts for restaurant, missing the matched traces for the two schemas. We corrected this (version B) for generating Fig. 3B, where the role decoding was crucial. Other figures used version A. We are regenerating all figures using the extended SchemaBench task prompts, and the camera-ready and future arxiv version will update all figures according to SchemaBench (version B). Crucially, because the number of roles per schema is orthogonal to our main findings, we strongly expect that none of this will change the main findings that a field captures latent reasoning, the predictive structure of the field and operations, the low dimensional control signals, the causal injection results, or our main behavioral findings (Fig 2). Nonetheless we wanted to report corrected additions to the datasets.
References [Wum25] Wu, et al. "Thought calibration: Efficient and confident test-time scaling." EMNLP 2025.
Thank you for your rebuttal
Thank you for your rebuttal. However, some of my concerns remain unresolved. Therefore, I decide to maintain my score.
Part 1
We sincerely thank the reviewer for their careful reading and helpful comments. First, we would like to highlight that while experiments with the global STOP vector are an important practical implication of our work, our central contribution is conceptual and methodological, providing a framework for extracting and analyzing multiscale dynamical structure that underlies these interventions. We do not claim to identify the optimal injection point, but rather show that, guided by our multiscale dynamics framework, a global STOP intervention can reliably terminate an ongoing reasoning trace.
Regarding your concern that intervention timing relies on completed baseline traces, we clarify that TSP injections were performed at 25%, 50%, and 75% of the baseline trace length for all trials, including those that terminated naturally (see Appendix D, “Transfer of stop vector injection to traveling-salesman problems (TSP)”).
Re long-horizon results: Long-horizon prediction was measured in two different ways, iterated one-step prediction and direct discounted long-horizon prediction. We show that the one-step field predicts approximately four primitives, while our successor analyses separately show predictability over longer horizons of eight discounted primitives. This grounds our multiscale predictive account, which parallels similar claims about brains (Brunec JNeuro). We find shared representational geometry given a shared vectorfield predicts operations across scales, though it does not establish that a single fitted transition operator generates all those predictions. STOP intervention shows causal control of termination through an internal direction, which also transfers to TSP. But because STOP was extracted independently, this establishes causal controllability of the model's reasoning, not causal validation of the fitted vector field itself.
re Question 1:
K-means and PCA were fitted on randomly sampled 15k tokens per layer and task family (e.g., identify, spot-error, and next-token prediction), without using temporal order or transition information during fitting. The resulting primitives and projections may still capture temporal patterns in the activations, but this does not introduce circularity, as these patterns are properties of the learned representations rather than information used to define the primitives. For PCA, we extracted the top 1–96 principal components of layer 16 for each model and selected the number of dimensions based on role-decoding accuracy on a held-out balanced dataset (see Figure 13 that is now added in Appendix G.3). Adding more than 48 dimensions gave no additional improvement, resulting in the selected 48-dimensional subspace out of the full residual dimensionality (3072 for phi4-mini-reasoning and 2560 total dimensions for qwen3-4B) for further analysis. As PC identifies dimensions of maximal variance, no explicit temporal information was used here. For spectral clustering, we used transition probabilities among the 50 clusters to identify 8 meta-clusters with primitive-motifs, i.e., within-meta-cluster temporal relationships. However, this does not extract, fit, or impose a sequential relationship among the 8 different meta-clusters. The meta-clustering step does not learn a transition model between meta-clusters.Therefore, predictability of transitions between meta-clusters is not built into their definitions, only their boundaries are, by our design, informed by the original transition matrix. Lexical and position-based controls: We compared decoding from residual activations against lexical bag-of-words and position-based baselines. Residual activations contain substantially more information (Supplementary Figure 9). We also examined different temporal orderings, including reversed order, to disentangle temporal structure from positional information and assess whether predictability persists (Figure 2C). As reported in Section 2 of the main paper, model performance was only mildly affected by event order (e.g., 50-52% for canonical and reversed orders with Phi4-mini-reasoning, compared to 45% for shuffled order).
re Question 2: We have added the accuracies and 95% confidence intervals for the paired accuracy differences for the two tasks. We calculated paired differences in final-answer accuracy for matched items, defined as the stop-vector outcome minus the uninjected outcome. For Schema, final-answer accuracy was 0.588 with the stop vector and 0.412 without injection, with a paired difference of +0.176 (95% CI: −0.277 to +0.630, 17 paired items). For TSP, accuracy was 0.361 with or without injection, with a paired difference of 0.0 (95% CI −0.109 to +0.109, 54 paired items).
Part1/2