acceptodds
Under review as a conference paper at ICLR 2027

Multiscale Predictive Dynamics Make Model Reasoning Observable and Controllable

Abstract

Language models routinely perform long multi-step reasoning, yet they sometimes loop without stopping, their chains of thought are not a reliable record of what they compute, and we lack an account of how local computations are coordinated into reasoning trajectories. Here we procedurally generate SchemaBench, multi-phase schema tasks that vary roles, temporal structure, and complexity and allow ground-truth parsing of reasoning traces. We evaluate 10 LLMs' behavior and the residual-stream activations of Phi4-mini-R and Qwen3-4B. Their latent computation is organized at three levels: a) task-general operations conduct local computations with a shallow grammar, b) schema-level representations of roles, phases, and progress bias which operations follow, c) slower control states track the switching and completion of reasoning. Across levels, reasoning is organized in a low-dimensional subspace of the residual stream, where a state-dependent vector field predicts the next latent transition and reasoning 5-10 steps ahead. Causally injecting a STOP control-vector into Phi4-mini-R ends traces that could otherwise loop to the token cap, shortens response length by 4000+ tokens without sacrificing accuracy, and transfers to an unseen mathematical task. LLM reasoning is thus organized as multiscale predictive dynamics within a shared low-dimensional field, which makes hidden reasoning observable and controllable.

Primary area:
interpretability and explainable AI
Keywords:
LLM, reasoning, language models, multiscale, successor representation, vector field, dynamics

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_TU8QOpenReview ↗
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
  1. SchemaBench varies role disclosure, narrative order, and role binding while retaining structured annotations, enabling more targeted representational analyses than final-answer accuracy alone.
  2. The paper is well-written and easy to follow. The figures and tables are visually appealing and clearly presented.
Critical weaknesses
  1. 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.
  2. 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
  1. 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?
  2. 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

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

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.

ConditionTotal tokensTokens afterRole accuracyAll pairs rightStrict
Control2,4041,148.991.976.905
STOP alpha .252,210990.995.976.917
STOP alpha .51,961742.995.976.929
"Final answer:"1,837662.985.940.786
</think>1,208194.967.893.869

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.

Reviewer_TU8QOpenReview ↗

Thank you for your rebuttal

Thank you for your rebuttal. However, some of my concerns remain unresolved. Therefore, I decide to maintain my score.

Reviewer_FbicOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies LLM reasoning on schema-based tasks motivated by neuropsychology. It introduces SchemaBench, procedurally generated restaurant and airport narratives with ground-truth role and event annotation. Ten LLMs are evaluated on it.

For Phi-4-mini-reasoning and Qwen3-4B, the paper:

  • reuses existing reasoning analysis methodology to extract a fixed set of primitives;
  • decodes schema-level states (role, phase) and control states (switch, solved, STOP);
  • fits an affine vector field on 48 PCs that predicts latent transitions between the primitives several steps ahead.

Injecting a STOP vector into Phi-4-mini-reasoning ends looping traces and is reported to transfer to TSP.

Critical strengths
  1. SchemaBench. Schema tasks are a good choice here. They require general knowledge, relationship tracking and abstraction, but still follow an expected template, so traces can be parsed against ground truth. The annotations also make timing analyses possible.
  2. Broad decoding analysis of the reasoning states. The paper maps in detail what the latent states and primitive sequences carry: roles, props, events and schema; role biding; predicting next actions; predicting control states.
Critical weaknesses
  1. The steering evidence is thin.
    • Only STOP is steered, on one model, one task and 17 traces.
  2. Limited novelty, partly uncited.
    • The primitive pipeline and the TSP transfer task come from AlgoTrace. AlgoTrace already injected a "plan final answer" primitive into TSP.
    • Steering a progress or length direction to shorten reasoning appears in Overclocking (arXiv 2506.07240) and ThinkEdit (arXiv 2503.22048).
    • None of these papers is cited.
  3. Analysis transfer outside of SchemaBench is not guaranteed. Primitives, the vector field and the control-state readouts are all extracted and evaluated on SchemaBench (two schemas). The abstract describes the primitives as "task-general operations", but most of the labels are task-specific ("person entity identification", "person list completeness", "event sequence reconstruction"). App. C.3 describes cross-task primitive merging, but no cross-task result is reported. Only the STOP vector is tested outside the benchmark.
Questions for authors
  1. What is the delta to AlgoTrace beyond SchemaBench? How does vSTOP differ from its "plan final answer" primitive?
  2. What happens when STOP is injected before the task is solved? What is the N for TSP? Does the effect hold on Qwen3-4B?
  3. How many generations per trace were run for Fig. 5a–c? Does vSTOP beat no injection on more traces, with several samples per trace? Does it beat a random vector at a strength where random has no effect?
  4. Do the primitives generalise beyond SchemaBench? App. C.3 describes cross-task merging but reports no result. How many SchemaBench primitives merge with AlgoTrace's TSP/AIME primitives? Does the vector field fitted on SchemaBench predict transitions on TSP traces? If most primitives are shared and the field transfers, the "task-general operations" claim (abstract) holds.
3 comments

We thank the reviewer for suggesting the relevant references, both now included and discussed. Our understanding is that Overclocking models reasoning progress and steers later latent layers along a progress direction to accelerate reasoning, while ThinkEdit edits attention heads associated with overly short reasoning to encourage longer reasoning. Our work complements these approaches by characterizing the underlying computational dynamics and transitions among specific algorithmic operations that sequentially implement the reasoning, as well as their multi-step predictive structure, and their coordination across computational scales. The reviewer is correct that AlgoTrace serves as a methodological starting point for extracting reasoning primitives. However, our contributions extend beyond their identification and "plan final answer" steering in several important ways. Your comment prompted us to revisit and more clearly articulate these contributions:

  • (a) Our primary aim was to investigate computational dynamics, moving beyond characterizing individual primitive-operation representations to modeling how computational operations transition the latent space over time, finding a learned vector field. While AlgoTrace identifies a graph of static primitive activations, we investigate the dynamics governing transitions among primitive-operations beyond simple transition statistics. This is important because understanding reasoning requires not only identifying which computations occur, but also understanding mathematical objects that lead to the selection, sequencing, and organization of primitives over time. The vector field (F) and its map (A) offer such mathematical objects.
  • (b) We demonstrate that latent primitive representations have a predictive structure, i.e., contain information about future computations, with a primitive at time step (t) containing predictive information up to 16–32 primitives ahead (Fig. 4A). This predictive structure was not investigated in AlgoTrace, but is important as it suggests that internal representations carry information beyond the computation currently being performed, providing insights into how models organize future computations.
  • (c) Our work goes beyond AlgoTrace by showing multi-scale coordination of latent dynamics, from local operations (e.g., entity identification and role binding) to longer-range schema and control processes (schema phase, prepare to STOP), are represented and coordinated during reasoning. While AlgoTrace focuses on primitive-level analysis, we also examine how representations at different computational and temporal scales interact. This relates to broader perspectives on the temporal organization of reasoning and the interplay between fast and slow processes [Ben19, Zia25], providing a mechanistic account of how such coordination can be implemented in language models.
  • (d) We introduce the SchemaBench benchmark specifically designed to study schema-based reasoning, with explicit roles, entities, events, and ground-truth annotations. Unlike the common mathematical task benchmarks considered in AlgoTrace, this setting enables temporally grounded analyses of how internal computations relate to structured task progression, including the evolution of roles, events, and schema states.
  • (e) Finally, we show how learned dynamics can be used to model and control computational transitions, including through a GRU-based meta-cognitive controller and targeted STOP interventions. Namely, we trained a gated recurrent unit GRU, 32 and 64 hidden units) on the latent patterns of the 48 PCs of layer 16, one step per primitive, with two heads: ready-to-quit and falling-into-a-loop. The loop label is at or after the onset of the first repeated primitive pair that keeps oscillating without exit. Median onset for the loop was at primitive 1007, about 2900 tokens into an 8192-token trace. On held-out passages the GRU's loop head reaches AUC .994 (.992 position-matched) and flags all 48 capped traces with an onset. This extends the analysis of internal representations towards understanding how computational dynamics can be exploited to influence model behavior, improving efficiency and interpretable model control.

References

  • [Ben19] Bengio Y. "From system 1 deep learning to system 2 deep learning." Neurips 2019.
  • [Zia25] Ziabari AS, et al. "Reasoning on a spectrum: Aligning llms to system 1 and system 2 thinking." Neurips 2025.

part 1/2

  • Re analysis transfer:

Great question. To test whether the primitives extend beyond SchemaBench, we compared the AlgoTrace clusters fitted separately on the Schema300 data subset, a 300-passage, counterbalanced dataset with 150 restaurant and 150 airport samples each. At the metacluster level (8 per dataset/model), 1 of 8 metaclusters matched a TSP metacluster in both models at layer 16, at layer 8 it was 1 and 2 of 8. At the fine level (50 clusters per dataset and model), 8 (Phi) and 11 (Qwen) of 50 Schema300 clusters matched a TSP cluster at layer 16; at layer 8, 8 and 12 of 50 with TSP. In summary, we find moderate overlap of up to 25% across investigated tasks and leave more detailed investigation for future work. Table 1 Metacluster and primitive cluster overlap on Schema300 subset (a 300-passage, counterbalanced dataset with 150 restaurant and 150 airport samples each.

Table 1 Metacluster and primitive cluster overlap on Schema300 subset (a 300-passage, counterbalanced dataset with 150 restaurant and 150 airport samples each.

LevelReadingSchema300–TSP PhiSchema300–TSP Qwen
8 metaclusters, L16Token evidence, same criterion1/8 (12.5%)1/8 (12.5%)
8 metaclusters, L8Token evidence1/8 (12.5%)2/8 (25%)
50 fine clusters, L16Token evidence8/50 (16%)11/50 (22%)
50 fine clusters, L8Token evidence8/50 (16%)12/50 (24%)

We would highlight that our main contribution is characterizing multiscale reasoning dynamics of computational operations in depth on SchemaBench, while providing initial evidence that some of the identified operations and control signals extend to more distinct tasks. Our TSP results provide complementary first evidence for task-general sharing, and we agree that while we’ve shown proof of principle results, future studies are required to scale them.

  • Re: How does vSTOP differ from its "plan final answer" primitive?

The key distinction is between local primitive steering and longer-range computational control. AlgoTrace's "plan final answer" is a static primitive injected at a specific layer to induce a local computational operation. In contrast, we interpret vSTOP as a higher-level control state that influences computation over a longer horizon, steering the overall reasoning process towards termination. This distinction is part of our contribution, highlighting how multi-scale coordination between local operations and higher-level control states can shape reasoning, and opening new possibilities for understanding and controlling model behavior, as presented in Section 6 and Fig. 5.

  • STOP injection before task solved: We have tested this and would like to clarify that TSP injections were performed at 25%, 50%, and 75% of the baseline trace length for all trials, before the solution has been arrived at (see Appendix D, “Transfer of stop vector injection to traveling-salesman problems (TSP)”).

  • Re: How many generations per trace were run for Fig. 5a–c?/does STOP beat random: 4 generations per trace. As seen in the figure, random vector injection never decreases the token length, across the strength levels and injection layers that we tested.

  • Re: Does the vector field fitted on SchemaBench predict transitions on TSP traces? If most primitives are shared and the field transfers, the "task-general operations" claim (abstract) holds.

We tested whether the field transfers from a task to another by comparing the similarity between their A (vector field map) in relation to the similarity of task decoding from the latent space. This was possible given that the tasks within SchemaBench are widely different. Across the 276 task pairs per model, the agreement of two tasks' remainder maps fell monotonically with how separable their layer-16 states were (Spearman −.92 for Phi, −.95 for Qwen; Mantel p < .001). Separability was determined using logistic regression (c=1). Field similarity was measured between two A maps as the normalised Frobenius inner product: the two 48 × 48 remainder maps are treated as vectors of 2,304 entries, and the cosine is their dot product divided by the product of their Frobenius norms. This is computed on the remainder maps A_G, after the per-coordinate relaxation has been removed.

Pairs within a task family (less decodable or separable in the latent space) had separability AUC .64 ± .12 / .63 ± .11 (mean ± SD, Phi / Qwen) and map cosine .87 ± .03 / .87 ± .04, i.e. .96 / .95 of the split-half ceiling; pairs across the identify and spot-error families had AUC .91 ± .03 / .91 ± .03 and cosine .56 ± .03 / .59 ± .02 (.62 / .64 of the ceiling). The relation held inside a family as well (Spearman −.70 / −.90; cosine .88 → .84 / .90 → .83 as the number of differing design factors rose from one to three) and inside the across-family pairs alone (−.63 / −.76), so map sharing is graded with task similarity rather than stepping at the family boundary.

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.

Reviewer_UEwEOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

Reasoning models often keep going well after reaching the answer, and sometimes loop without stopping. This paper studies the internal state behind such reasoning. Prior work like AlgoTrace identifies primitive operations, but according to the authors, it does not explain how they combine or when the model stops. The authors build SchemaBench, a set of multi-step schema tasks with known ground-truth states. Using it, they argue that the hidden state of two small reasoning models is organized at three timescales: primitive, schema, and control. This state is also partly predictable by a low-dimensional vector field. Injecting a stop direction into one of the models makes it end responses that had hit the token limit much earlier, and the same direction also shortens responses on TSP. They take this as evidence of an internal stopping state that causally affects termination.

Critical strengths
  • Great Benchmark as Contribution. SchemaBench gives ground-truth states at each step, so probing results can be checked directly. The controlled variants (naming, detail, distractors, shuffled order) are useful on their own.
  • Novel Perspective on Reasoning. Most interpretability work asks static questions, such as which direction encodes a concept. This paper treats reasoning as a dynamical process at several timescales, drawing on neuroscience and schema theory. The tools mostly exist already, but combining them under this view is new to me, and it makes questions like “when does the model decide it is done” testable.
Critical weaknesses
  • No Comparison to Trivial Stopping. Injection starts at the point where the model has already stated the schema and every (person, role) pair. At that point, simply cutting the trace and appending </think> or "Final answer:" would also end the response, and could give similar accuracy. The vSOLVED and vSWITCH controls show that v_STOP is more specific than other learned directions, but not that it does better than such simple interventions. The authors should report length, termination rate, and accuracy for forced stopping at the same position T, and compare with existing overthinking steering methods (e.g., Huang et al., 2025).
  • Robustness of the STOP Results. The injection setting (L16, strength 0.25) was selected on the same 17 responses used to report the effect, so the reported reduction is likely optimistic. Other settings in Fig 5E even lengthen responses, much like the random vector. On TSP, accuracy is preserved only at strength 0.25. At 0.5, accuracy drops to near zero (Fig 5G), which the text does not mention. Also, 59% of the selected responses stop on their own without injection, so they are not traces that would otherwise loop to the limit. The authors should select settings on held-out items and report results for all settings.
Questions for authors
  • Could the STOP readout or the field’s forecast predict a loop before it happens, and be used to trigger injection only when needed instead of at a hand-chosen position? This would connect the observability and controllability parts of the paper.
  • vSWITCH was used only as a control and lengthened responses. Could switch or phase directions be used to redirect reasoning on demand, for example moving the model to the next person or the next schema phase? That would show controllability that simple forced stopping cannot achieve.
2 comments
AuthorsOpenReview ↗

Thank you, please see new analysis

Thank you for the careful reply and helpful suggestions, and 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.

ConditionTotal tokensTokens afterRole accuracyAll pairs rightStrict
Control2,4041,148.991.976.905
STOP alpha .252,210990.995.976.917
STOP alpha .51,961742.995.976.929
"Final answer:"1,837662.985.940.786
</think>1,208194.967.893.869

Direct prompt interventions are thus an effective way to stop reasoning and output an 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 does not provide the same insight into the model's internal dynamics or the coordination between primitives and schemas that our framework offered.

To more clearly position our study on global stop signals, we will add clarification and the suggested prompting baseline to Section 6, alongside related works on overthinking [Hua26, Li26] and online detection of plateaued reasoning [Wum25]. Complementary to these works, we provide evidence of how stopping signals coordinate across scales, from global control signals through schema-level states to local primitive transitions, enabled by a novel schema-based benchmark, SchemaBench, designed to expose and study these multiscale dynamics.

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, before the solution has been arrived at (see Appendix D, “Transfer of stop vector injection to traveling-salesman problems (TSP)”). We will add clarification.

Re Robustness of STOP results. Section 5 investigates whether a global STOP intervention can reliably terminate ongoing reasoning traces, rather than to optimize intervention location and strength. The sensitivity of steering outcomes to these parameters is an important consideration, but optimizing them is a separate problem that is not the focus of our study.

As highlighted in your comment, our TSP experiments reveal a trade-off between shortening reasoning traces and preserving task performance. Our goal in these experiments, however, is not to identify an optimal STOP intervention, but to demonstrate the potential of global control states as a novel finding enabled by our investigation. This finding complements our other contributions: the shared field, successor vector fields, and predictive dynamics revealed through SchemaBench. More broadly, it illustrates our central conceptual contribution: understanding steering as an intervention that modulates multiscale dynamics.

We agree that identifying optimal intervention points and dosages, and refining them across tasks and models, remains an important direction for future work. In particular, adaptive steering [Hed25] offers a promising approach to calibrating intervention strength and avoiding oversteering.

Re: V-Switch/control to redirect reasoning on demand:

Great question. In Fig 5D we plot changes in primitive distribution compared to baseline after we injected v_SWITCH to layer 16 of phi4mini-r. After injecting vSWITCH (Fig 5D, purple bars), the reasoning traces elicits more primitives corresponding to person entity identification (which is a metacluster containing primitives like: “Enumerate mentioned people,” “Person-role tracking,” and “Entity action mapping” with prominent tokens including person, man, woman, description, and list) and fewer to person list completeness (a check of who belongs on the final list of people in the response) compared to baseline. In contrast, injecting v_STOP (Fig 5D, red bars) decreases person entity identification and increases person list completeness primitives in the trace. Moreover, of all the vector injections (random, switch, STOP, solved) only STOP vector injection increases person-role answer assembly. Together, these findings suggest that v_STOP induces different operations than others in terms of person list completeness and answer assembly. We hope this addresses the reviewer’s questions, showing controllability beyond STOP. An interesting future direction inspired by this comment is to also report on strategy changes before and after a switch and the final accuracy. We believe being able to measure these changes is a promising step toward future extensions.

Re: Could STOP readout/field forecast predict a loop before it happens: We thank the reviewer for their question. We have pursued a related direction where we trained a gated recurrent unit GRU, 32 and 64 hidden units) on the latent patterns of the 48 PCs of layer 16, one step per primitive, with two heads: ready-to-quit and falling-into-a-loop. The loop label is at or after the onset of the first repeated primitive pair that keeps oscillating without exit. Median onset for the loop was at primitive 1007, about 2900 tokens into an 8192-token trace. On held-out passages the GRU's loop head reaches AUC .994 (.992 position-matched) and flags all 48 capped traces with an onset. The STOP-approach readout (P(STOP within 10 primitives) from the 48 PCs does not predict loops, held-out AUC .80/.79 for Phi/Qwen) rises only over the last 8 primitives (16–32 tokens) of a terminating trace.

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.
  • [Wu25] Wu, et al. "Axbench: Steering LLMs? Even simple baselines outperform sparse autoencoders." ICML 2025
  • [Hed25] Hedström, et al. "To steer or not to steer? mechanistic error reduction with abstention for language models." ICML 2025.
  • [Hua26] Huang, et al. "Mitigating overthinking in large reasoning models via manifold steering." Neurips 2026
  • [Li26] Li, et al. "Stop When Further Reasoning Won't Help: Attention-State Adaptive Generation in Reasoning Models." ICML 2026

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.