acceptodds
Under review as a conference paper at ICLR 2027

Rethinking Activation Steering for Mathematical Reasoning: What Actually Drives Its Reasoning Gains?

Abstract

Activation steering can improve mathematical reasoning and is commonly interpreted as controlling semantically meaningful internal reasoning states, such as reflection, reasoning depth, or over- and underthinking. We question whether such internal-state-control explanations are necessary to account for the resulting accuracy gains. Across three steering constructions, two model scales, and two mathematical reasoning benchmarks, we find that methods with different internal-state interpretations converge on a common downstream effect: they alter local decoder preferences and redirect generation among reasoning trajectories already supported by the model. Correct solutions and related reasoning strategies are frequently accessible under unsteered sampling. More directly, ordinary stochastic decoding whose parameters are selected to match steering's aggregate departure from greedy decoding achieves comparable accuracy while substantially reducing the remaining marginal benefit of steering; whole-thought exploration reproduces the same pattern. These results support a unified greedy-relaxation account of reasoning-oriented activation steering. Under this view, its accuracy gains need not be explained by a method's ability to identify and structurally control a particular internal reasoning state. The experiments instead localize a shared accuracy-relevant bottleneck at trajectory selection: steering changes which already-supported trajectory is realized, and ordinary stochastic decoding can exploit this bottleneck without reproducing the steering-specific activation intervention. This separates the internal representation targeted by a steering method from the mechanism explaining its performance gain.

Primary area:
interpretability and explainable AI
Keywords:
Activation Steering, Mathematical Reasoning, Reasoning Trajectories

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_TAanOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper concentrates on unifying activation steering methods for improving mathematical reasoning. Plenty of experiments are conducted on DeepSeek-R1-Distill-Qwen-1.5B and DeepSeek-R1-Distill-Qwen-7B. The eval sets are MATH500 and GSM8K.

This work first show that three steering methods--SEAL, REBALANCE and CPS--generate correct responses already in sampled results. Then decoder changes are compared to show that removing steering cache can keep most of the gains. Finally, the improvement from stochastic decoding can improve accuracy to the results of steering.

Critical strengths

The results is quite interesting. A quite clean explanation is given to show that steering methods mainly relax greedy decoding to improve the accuracy. The clean-cache experiment shows that the change of decoding gives most of the improvement. I think this experiment gives steering methods some insight.

Critical weaknesses

My main concern for this work is that the conclusion is not solid enough.

  • The experiments are limited to 2 models that are only SFT trained. The results on models that are trained with RL are not clear. Since RL is known to turn pass@k to pass@1, I think results on RL models might be necessary.
  • Only results on math are shown. Math-500 and GSM-8k are famous math test sets and might not be representative enough. Harder problems like AIME, new problems, and problems from other domains might give us more observations.
  • Further, I am willing to see more formal explanation for the reasoning mechanism. I am not sure whether the greedy-relaxation is the only way for explaining steering methods. The greedy-relaxation itself is not quite clear. Are all tokens equally important? Or only some tokens work? I think a deeper analysis might provide more insights.

Besides, the presentation is hard for me. The references for tables are not clear and made me confusing. Tables and figures are not well organized and many results are directly given in text.

Questions for authors

See weaknesses above.

Reviewer_HzZ1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper claims that accuracy gains from reasoning-oriented activation steering, e.g., SEAL, ReBalance, and the paper's own CPS, don't require the usual explanation that steering controls an internal reasoning state; it is rather the case that steering perturbs decoding decisions and shifts generation onto correct trajectories the model already supports, which the paper calls "greedy relaxation." Experiments were done on DeepSeek-R1-Distill-Qwen-1.5B and 7B, using MATH500 and GSM8K.

Critical strengths
  1. The separation of decoder and cache is a well designed causal experiment. The result that a clean-cache keeps most of the gain when a steered cache without the decoder doesn't (67.25% vs. 73.25% baseline) is a concrete, testable mechanistic claim.

  2. The paper rethinks how steering methods should be evaluated. Steering papers usually compare against greedy decoding, a weak baseline for these models. If the results are correct, future steering work would need stochastic-decoding baselines and interaction tests. That is a methodological contribution.

Critical weaknesses
  1. Inconsistent numbers weaken the central matched test:

(a) Matched sampling doesn't look specific to each method. Table 6 shows that the matched sampling condition (S) has the same accuracy across SEAL, ReBalance, and CPS within each model and dataset pair (e.g., 82.44±0.9482.44 \pm 0.94 for all three on 1.5B MATH500). But Table 8 reports different sampling parameters per method. Therefore, either S was not actually matched for each method (which weakens the paper's test) or there is a reporting error.

(b) Baselines disagree within Table 2: Unsteered greedy accuracy should be same across methods but differs (31.67% vs. 46.50%), and some values imply 120 problems rather than the stated 200.

(c) Figure 6 is not compatible with Table 7: Several native steering gains don't match (e.g., 9.00 vs. 4.75), making the evaluation population for Sec. 6.2 not clear. Since section 6.1 is the paper's main evidence, these mismatches directly affect its main claim.

  1. The main tests don't disqualify the internal state explanation:

(a) Steering with random directions of matched strength would show whether the specific direction matters. But this not tested. Sub-additivity isn’t specific. Adding steering on top of sampling would show little extra gain under almost any explanation, since both act on the same uncertain positions and accuracy has a ceiling.

(b) Counterpart analyses has no baseline. The 93% rate for the same mathematical strategy doesn't say much without knowing how often two random correct solutions share a strategy, since many MATH problems have one standard approach.

(c) Efficiency looks to be ignored. SEAL and ReBalance mainly claim shorter reasoning, but length, truncation, and looping are never compared between steering and sampling. Same accuracy at higher cost would not make them equivalent.

(d) The 7B evidence is weak. Five of six 7B interaction confidence intervals include zero, so “negative in all 12” overstates the result.

Questions for authors

Please see the weaknesses above

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

In this paper, they know that activation steering improves mathematical reasoning but they want to find out that this improvement is from controlling specific internal reasoning states or from changing the trajectory that the model is generating. They do it with two R1 models on math500 and gsm8k with three different methods of steering. They do it by seeing that can they get the same correct answer and reasoning strategies with ordinary sampling and steering. They also see that does steering change decoder decisions and also does the accuracy stay similar without steered hidden states and kv cache. They also compare steering and stochastic decoding. Their results show that with unsteered sampling the correct reasoning trajectory is there already and also when they use stochastic decoding, there is no need for steering because there is a very small improvement. So the thing that we all know that steering improves reasoning is not necessarily in that way and the improvements can come from changing the selection of current trajectories.

Critical strengths

Their motivation to see what is the reason that others are saying steering improves the performance is great and their experiments in 4 to see that does steering make the model to learn or it is already there with sampling is solid. The way they separate the effect of the generated text from the effect of steering hidden states and kv cache in 5 with clean-cache replay is solid. Also the thing that they apply both stochastic decoding and steering beside comparing them in 6 is also solid. Their use of three different steering method is also good for generalize the findings.

Critical weaknesses

It is not convincing that greedy relaxation is the reason for the improvements that comes from steering although in 4.1 they show that steering does not have any advantage to the ordinary stochastic decoding and also in 6.1 they match the greedy departure of stochastic decoding with steering and show that steering does not change it a lot and there is only small improvements. But all of these are kind of saying that ordinary sampling can already have enough performance and there is no need for steering and somehow the same thing when they combine and apply both steering and sampling and there is just a little improvement. So this can be because that both methods correct overlapping sets of problems or the sampling already gets what it should get because it is high accuracy and there is no room for more improvements. Also something else is that the matched sampling is applied at where each method of steering can intervene.

It is kind of hard to compare part of the results in 5 and 6, in table 2, they mention that the clean-cache replay is evaluated on a fixed 200 examples from math500 but, the unsteered greedy accuracy for the 1.5B is 31.67 for seal and rebalance, and it is 46.5 for cps. So it is not clear why the unsteered baseline is different across methods because these should be the same thing. Also some of the accuracy reported like 31.67, 60.83 can not come from 200 examples. Also in figure 2, in the figure the matched stochastic decoding shows the same accuracy in all of the steering methods with each model and dataset thought in 6.1 different sampling policies calibrated for different steering methods. So it is unclear that these differences are from method specific evaluation conditions, number of examples or some other part.

Questions for authors

Can you give more direct control to show that matching the greedy departure distribution is important for the improvements instead of the greedy departure distribution in 6.1? For example you can do it with another sampling policy that has the same number of sampled intervention but its depth is different from the greedy departure.

As mentioned in the weaknesses about figure 2 and 6.1. In figure 2 the accuracy of matched stochastic decoding is same for all of the steering methods with each model and dataset but in 6.1 the parameters are selected independently for each of the greedy departure distributions of each steering method. So for this can you clarify that why these different policies give the same accuracy and also is each sampling result independently evaluated for each method or is it based on prior works or something common?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.