- Rating
- 1
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
The paper presents STARS, a training-free method for steering large reasoning models during decoding. The authors group overthinking and rigid adherence to prior assumptions under the term cognitive inertia. STARS monitors the size of the hidden-state change between consecutive tokens at a selected layer. When a spike occurs, it checks whether the update direction reverses and whether the current state resembles states recorded at earlier reversals, and on that basis appends one of two short sentences to the generation, either prompting the model to verify its new direction or telling it that its reasoning is looping. On DeepSeek-R1-Distill-Qwen 1.5B and 7B and Qwen3-4B-Thinking, across several mathematics benchmarks, PuzzleTrivial and GPQA-Diamond, STARS improves average accuracy by roughly two to three points over standard decoding and runs faster than the DEER and ConCISE baselines.
Critical strengths
- The timing experiments are well conceived. The same sentence is inserted at occurrences of the token wait, at fixed positions, or whenever an LLM judge flags a problem, and spike-triggered insertion performs best in most configurations. This is precisely the control needed to show that the benefit does not come from the wording of the suffix alone.
- The method is inexpensive and straightforward to deploy. The trigger relies on hidden states that are already computed, requires no additional forward pass, and yields lower latency than DEER, which makes it easy to adopt for open-weight models.
Critical weaknesses
- The evidence is modest relative to the strength of the claims. After filtering, ConditionedMath contains 52 problems, PuzzleTrivial 33 and each AIME year 30. The 6.5-point gain on ConditionedMath therefore corresponds to three or four problems and the 6.7-point gain on AIME25 to two, and many differences fall within one standard deviation, with essentially no change on GPQA-Diamond. Although the thresholds are described as calibrated on held-out data, Appendix E.4 selects the final values by inspecting accuracy on ConditionedMath and AIME24, both of which are evaluation sets, and the chosen pair happens to be optimal on both. The calibration pool, moreover, consists of the filtered-out portion of the same benchmarks, including near-duplicates of test items. The prefix-continuation analysis in Table 4 raises a further concern. From the intervention points, the base model without any suffix solves most of the problems on average, which suggests that much of the recovery comes from resampling at that point rather than from the suffix, and treating the remaining nine to sixteen percent as causal evidence overstates what the analysis shows.
- Important comparisons are missing, and the mechanistic interpretation is not fully convincing. The prompt-hinting baseline in Appendix I reaches 67.7 on ConditionedMath with Qwen3-4B, well above the 58.9 obtained by STARS, yet it is absent from the main results. Closely related latent-steering methods such as SEAL and manifold steering, as well as NoWait, are discussed but not evaluated, and the two baselines that are included are early-exit methods designed for efficiency rather than accuracy. All models have at most seven billion parameters, so it remains unclear whether cognitive inertia, and the corresponding gains, persist in larger reasoning models. As for the mechanism, the layer selected for Qwen3-4B is layer 2, which is very shallow and largely lexical, and the examples in Figure 3 place spikes directly on discourse tokens such as Wait and Oh, nice. The detector may therefore be tracking particular tokens rather than internal reasoning transitions. The notion of cognitive inertia itself adds little beyond a new label for overthinking and rigidity.
Questions for authors
- Could the thresholds be selected without reference to any evaluation benchmark, and could the results be accompanied by a paired significance test over problems? I would like to know which gains survive both changes.
- On the problems in Table 4, how does STARS compare with suffix-free continuation from the same intervention points? A small difference would indicate that the suffix contributes little.
- How often do spikes coincide with specific tokens such as Wait or line breaks? A control that triggers on those tokens at the same frequency as STARS would help separate the two explanations.
- Why is prompt hinting excluded from the main table, and how does STARS compare with SEAL and NoWait?
- Do the gains hold for a reasoning model in the 30B range?
2 comments
W2-1 & Q4. Prompt hinting
Prompt hinting outperforms STARS in only one of the nine settings in Table 13 of the original submission (now Table 23 in the revised manuscript), underperforming in the other eight and on average. It also falls below vanilla on AIME24 for all three models. The hint explicitly signals an unusual condition, a defining property of ConditionedMath, thereby providing benchmark-specific information at the input, whereas STARS responds dynamically as reasoning unfolds. We therefore report prompt hinting separately in Appendix I.
W2-2 & Q4. Reflection suppression
SEAL and Manifold Steering suppress reflection through direct hidden-state modification, whereas STARS monitors hidden states and redirects reasoning via language cues. As a representative reflection-suppression baseline, we added NoWait (Appendix J).
| Method | CM 1.5B | CM 7B | CM 4B | AIME24 1.5B | AIME24 7B | AIME24 4B | AIME25 1.5B | AIME25 7B | AIME25 4B | Avg |
|---|---|---|---|---|---|---|---|---|---|---|
| Vanilla | 46.54 ± 4.28 | 48.08 ± 5.01 | 56.54 ± 1.54 | 28.00 ± 4.52 | 53.33 ± 2.11 | 76.00 ± 3.89 | 22.67 ± 3.27 | 39.33 ± 1.33 | 70.67 ± 5.33 | 49.02 |
| NoWait | 43.08 ± 5.78 | 46.54 ± 4.28 | 56.15 ± 0.77 | 26.00 ± 2.49 | 44.00 ± 6.46 | 70.00 ± 2.11 | 18.67 ± 5.42 | 29.33 ± 4.42 | 72.67 ± 2.49 | 45.16 |
| STARS | 53.08 ± 2.61 | 54.62 ± 1.96 | 58.85 ± 3.12 | 28.00 ± 4.00 | 55.33 ± 3.40 | 76.67 ± 2.98 | 27.33 ± 3.89 | 43.33 ± 2.98 | 77.33 ± 3.89 | 52.73 |
Table C: Accuracy (%, mean ± std over five trials). NoWait suppresses the reflection keywords of Wang et al. (2025) by setting their logits to .
NoWait shortens generation to 8,958 tokens (vanilla: 14,987; STARS: 14,104) but falls below vanilla in eight of nine settings. STARS is best or tied in all nine settings, supporting selective redirection over uniform reflection suppression.
W2-3 & Q5. Model scale
Evaluating a 30B-scale model with 32K-token generation and per-token hidden-state access exceeds our single 48GB GPU capacity without quantization, which would alter the monitored states. Reasoning rigidity has been documented in Qwen3-32B (Jang et al., 2025), and our gains persist across 1.5B–7B models. STARS requires only hidden-state access, making it applicable in principle to larger open-weight models.
W2-4 & Q3. Token tracking
Across the main experiments, only 0.73%, 2.04%, and 0.00% of spikes coincide with discourse tokens (e.g., Wait, Oh, Hmm, But, Alternatively) for DeepSeek-1.5B, DeepSeek-7B, and Qwen3-4B, respectively, below their prevalence in generated text (2.6–2.8%). At most 0.10% of paragraph line-break positions coincide with a spike. In Figure 3, the highlighted tokens are "style" and "numer"; "Wait" and "Oh, nice" occur in the preceding context.
Table 6 already provides the suggested Token-Triggered control, which injects the same suffix at "wait" tokens, with its intervention limit matched to the average flip frequency. STARS outperforms this control in all six settings, supporting latent detection beyond these surface cues. We have included the token overlap analysis in Appendix F.2 of the revised manuscript.
W2-5. Cognitive inertia
The unification is operational: both failure modes involve self-reinforcing trajectories, enabling a common detector but requiring different corrective actions. Flip-only and STARS share the same intervention trigger, differing only in the suffix applied at recurrent flips. STARS consistently outperforms Flip-only (Tables 5 and 12 in the original submission, now Tables 5 and 20 in the revised manuscript), supporting the value of this distinction.
We thank the reviewer for the thoughtful feedback and for recognizing the value of our timing controls and deployment efficiency. The revised PDF reflects all changes below.
W1-1 & Q1. Problem-level significance
Five-run standard deviations measure seed variability rather than uncertainty over problems. We therefore compute paired per-problem accuracy differences averaged over trials, with problem-level bootstrap CIs and paired permutation tests across five target benchmarks.
Table A: Problem-level paired comparison with vanilla across MathPerturb, ConditionedMath, PuzzleTrivial, AIME24, and AIME25.
The pooled gain and the gains for both DeepSeek models are statistically significant. Qwen3-4B also shows a positive gain (+1.26 pp), although it does not reach statistical significance, with identical outcomes on 76% of problems. MATH500 and GPQA-Diamond assess generalization rather than targeted improvements (Section 4.2).
W1-2 & Q1. Threshold selection and calibration pool
The default thresholds were frozen before the results in Table 10 of the original submission were examined. Table 10 (now Table 14 in the revised manuscript) was a subsequent sensitivity analysis, not a selection step (clarified in Appendix E.4). Although the calibration pool includes near-duplicates, threshold selection relies exclusively on label-free statistics (SPR, spike density, and cosine and recurrence distributions), never ground-truth answers or evaluation accuracy.
Calibration uses only excluded ConditionedMath and MathPerturb instances. The frozen configuration transfers unchanged to AIME24/25, PuzzleTrivial, and GPQA-Diamond, none of which appears in calibration, with some of the largest gains on AIME25 and PuzzleTrivial. Table A establishes the pooled significance of the gains without accuracy-based threshold selection.
W1-3 & Q2. Outcome-independent effect of the suffix
Table 4 in the original submission examined whether reaching the intervention point alone accounts for STARS's recoveries. To directly quantify the suffix effect, we conducted an outcome-independent paired comparison. Since STARS follows vanilla until its first intervention, we compare STARS and suffix-free continuations from every first-intervention prefix, regardless of final outcome, with identical prefixes and budgets.
Table B: Paired continuation accuracy (%). CM denotes ConditionedMath.
Across 1,058 prefixes, STARS improves accuracy by 6.99 pp (95% problem-level bootstrap CI [+4.58, +9.46]), with 69 wins versus 23 losses (sign test, p<10−5). STARS improves five settings and ties in one.
Restricting to STARS successes as in Table 4 of the original submission yields suffix-free accuracies of 63.6%, 69.4%, and 86.9% on ConditionedMath, consistent with the original results. Including all eligible prefixes demonstrates the benefit of suffix-based steering beyond resampling. We have replaced Table 4 in the revised manuscript with this analysis.