acceptodds
Under review as a conference paper at ICLR 2027

Signal from Noise: Activation Redistribution Improves Reasoning in Large Language Models

Abstract

Motivated by the puzzling observation that prepending long sequences of meaningless tokens can consistently enhance LLM reasoning performance, we investigate the underlying mechanism and propose a more principled alternative. We find that these improvements arise from a redistribution of activations in the MLP layers, where near-zero activations become less frequent while large-magnitude activations increase, effectively suppressing weak signals and promoting more informative representations. Building on this insight, we propose the Activation Redistribution Module (ARM), a lightweight inference-time technique that directly modifies post-nonlinearity activations without altering the input sequence. ARM adaptively shifts near-zero activations outward, implicitly reproducing the benefits of meaningless tokens in a controlled manner. Extensive experiments across diverse benchmarks and model architectures demonstrate that ARM consistently improves reasoning performance while requiring only minimal implementation overhead.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Counter-Intuitive Phenomenon

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_mYrzOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies improvements in LLM reasoning obtained by inserting repeated, semantically uninformative tokens between the system prompt and the question. It approximates their first-layer attention effect using scaling and additive contributions and associates this intervention with fewer near-zero MLP activations. Motivated by these observations, ARM directly applies bounded, sign-preserving outward shifts to selected post-nonlinearity activations in the first-layer MLP. The paper reports improvements across six models and mathematical, scientific-question-answering, and coding benchmarks.

Critical strengths
  1. Interesting research question. Using training-free methods like inserting uninformative tokens / ARM to redistribute activation / attention is a interesting trial.
  2. Extensive experiments. Table 2, Table 13 and Table 15 have demonstrate the effectiveness of method across different models, benchmarks and pass@k.
  3. Well-motivated method design. The observed activation redistribution motivates ARM as a direct internal intervention, providing a useful connection between the token-insertion phenomenon and training-free reasoning improvement.
Critical weaknesses
  1. The performance degradation with overly long meaningless-token sequences remains insufficiently explained. The authors attribute it to accumulated perturbations in Section F.1 and by Eq 3, but provide no comparison of attention outputs across insertion lengths to support this hypothesis.
  2. The evaluation of ARM focuses on relatively older model families, leaving its effectiveness on more recent, stronger reasoning models unclear.
Questions for authors
  1. While ARM improves performance, how it changes generation behavior and how these changes relate to the gains remain unclear. Although the authors have discussed mless's influence on termination rate in F.1 and ARM's influence on 2-gram diversity in Fig 13, I am still curious about the influence on output length.
3 comments
AuthorsOpenReview ↗

Reply(1/1)

We thank the reviewer for their insightful questions and constructive feedback.

1.The performance degradation with overly long meaningless-token sequences remains insufficiently explained. The authors attribute it to accumulated perturbations in Section F.1 and by Eq 3, but provide no comparison of attention outputs across insertion lengths to support this hypothesis.

As the number of inserted tokens increases, these tokens receive a larger share of attention, while the original input receives a smaller share. With a moderate number of inserted tokens, this change can produce the beneficial activation redistribution discussed in Section 3. However, with too many inserted tokens, the attention output may deviate significantly from its original representation, making information from the original question less influential. Consequently, excessive insertion may interfere with the model's reasoning process rather than improve it.

2.The evaluation of ARM focuses on relatively older model families, leaving its effectiveness on more recent, stronger reasoning models unclear.

We evaluated ARM on Qwen3.5-4B using MATH-500 and AIME 2025, with a maximum output length of 32,768 tokens. For each benchmark, we conducted three runs with different random seeds and report the mean accuracy and standard deviation. The results show that ARM also improves performance on Qwen3.5-4B, extending our evaluation to a more recent model.

ModelBenchmarkBaselineARM
Qwen3.5-4BMath50092.07 +- 0.3893.67 +- 0.90
Qwen3.5-4BAIME202543.33 +- 0.0044.82 +- 1.05
Qwen3.5-9BMath50095.00 +- 0.0096.60 +- 0.43
Qwen3.5-9BAIME202553.33 +- 0.0064.43 +- 4.16

1.While ARM improves performance, how it changes generation behavior and how these changes relate to the gains remain unclear. Although the authors have discussed mless's influence on termination rate in F.1 and ARM's influence on 2-gram diversity in Fig 13, I am still curious about the influence on output length.

We report the average output length for Qwen2.5-Math-1.5B and Qwen2.5-7B-Instruct on MATH-500 and AIME 2025. ARM produces outputs of similar average length to the vanilla models, suggesting that its performance gains are not primarily driven by generating more tokens.

ModelBenchmarkVanilla ModelARM
Qwen2.5-Math-1.5BMath500646.39651.37
Qwen2.5-Math-1.5BAIMR20251055.731096.40
Qwen2.5-7B-InstructMath500618.99626.43
Qwen2.5-7B-InstructAIME2025932.57920.40

Thank you for your response. I will keep my score.

Dear Reviewer mYrz,

Thank you for your positive feedback and for acknowledging our efforts to address your concerns. We are glad that our responses have adequately resolved your major concerns about generalizability of new model family, output length and further mechanism explaination. Given that your main concerns have been addressed, we would sincerely appreciate it if you could consider updating your rating accordingly.

Thank you again for your time and constructive feedback.

Reviewer_niDuOpenReview ↗
Rating
3
Confidence
3(+1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates how inserting meaningless tokens can improve LLM reasoning and identifies activation redistribution in the first-layer MLP as a potential underlying mechanism. Experiments across six LLMs and multiple reasoning benchmarks demonstrate generally consistent performance improvements with minimal implementation overhead.

Critical strengths

1- The paper connects the counterintuitive meaningless-token effect to activation redistribution and it offers a meaningful perspective on how LLMs reason. 2- ARM is lightweight and training-free, achieving generally consistent improvements across six LLMs and diverse reasoning benchmarks.

Critical weaknesses

1- Although activation redistribution correlates with improved reasoning, the theoretical analysis relies on simplifying assumptions, and the experiments do not fully establish causality. 2- Some performance gains are relatively modest, and the stochastic activation modification may introduce variability (e.g. Table 2).

Questions for authors

1- What evidence establishes that activation redistribution directly contributes to improved reasoning, rather than being a secondary effect of token insertion? 2- How well do the simplifying assumptions in Section 3 reflect the actual activation distributions observed in LLMs? 3- How sensitive are ARM's improvements to stochastic activation modifications and random seeds? Are the reported gains statistically significant? 4- What explains the substantial variation in improvement across models and benchmarks (Table 2)? Are there identifiable conditions under which ARM provides little benefit?

4 comments
AuthorsOpenReview ↗

Reply(1/2)

We sincerely thank you for your thoughtful comments and insightful questions.

W1. Although activation redistribution correlates with improved reasoning, the theoretical analysis relies on simplifying assumptions, and the experiments do not fully establish causality.

Q1.What evidence establishes that activation redistribution directly contributes to improved reasoning, rather than being a secondary effect of token insertion?

Our paper provides three complementary pieces of evidence that activation redistribution is one of a reason of improving reasoning.

  • Meaningless tokens can improve performance. Section 2 shows that inserting meaningless tokens improves performance, and our subsequent analysis identifies the associated activation redistribution.
  • ARM use similar mechanism to improve the performance. After analysis, ARM directly induces a similar redistribution mechanism without inserting additional tokens into the model and also improves performance.
  • Removing redistribution leads performance drop. Section 3.3 reports a restoration intervention: while retaining the inserted meaningless tokens, we replace the first-layer activations with the corresponding activations from the original input, which reduces performance from 63.8 to 62.6.

To assess the robustness of this finding, we repeated the evaluation across three random seeds on MATH-500 using Qwen2.5-Math-1.5B on removing redistribution. Restoring the original activations eliminates the observed benefit of token insertion and reduces mean accuracy below the baseline. Together with ARM’s gains without token insertion, these complementary interventions support activation redistribution as a contributing mechanism in reasoning improvement, rather than merely a correlated by-product.

BaselineMeaningless-token insertionARMToken insertion + activation restoration
63.80 ± 0.2765.90 ± 0.7166.47 ± 0.4161.27 ± 0.68

Q2.How well do the simplifying assumptions in Section 3 reflect the actual activation distributions observed in LLMs?

We assess its empirical relevance in several ways:

  • Figure 4 presents post-activation distributions collected from actual model runs with and without meaningless-token insertion, rather than distributions constructed by manually applying the theoretical affine transformation. Across the three analyzed models, insertion reduces the proportion of near-zero activations and increases the frequency of larger-magnitude activations. The accompanying measurements show decreased relative sparsity and increased L1 and L2 norms. These observations agree with the redistribution pattern predicted by our analysis.

  • Appendix F.2 directly checks the sufficient local condition underlying the scaling-factor analysis. Using 100 randomly selected MATH-500 examples with Qwen2.5-Math-1.5B, we evaluate the induced local scaling perturbations and find that 96.4% of the evaluated perturbation instances satisfy this condition. This check concerns the local change around the operating point. It therefore supports the relevance of the local variance analysis to the observed intervention.

  • Appendices F.4–F.5 examine whether the predicted variance changes occur across models and activation functions. Table 5 reports post-activation variance ratios of 31.66, 1.30, and 40.00 for Qwen2.5-Math-7B, Llama3.1-8B-Instruct, and Gemma3-4B-it, respectively. Although the magnitude varies substantially, all three show an increase. Appendix F.5 further compares SiLU and GeLU, including activation-function replacement experiments, and observes variance increases under both functions.

W2.Some performance gains are relatively modest, and the stochastic activation modification may introduce variability (e.g. Table 2).

Q3.How sensitive are ARM's improvements to stochastic activation modifications and random seeds? Are the reported gains statistically significant?

Thank you for highlighting the need to assess variability across random seeds. We evaluated both ARM using three random seeds on MATH-500 and AIME 2025 with two models.

ModelBenchmarkBaselineARM, mean ± SD
Qwen2.5-Math-1.5BMath50063.866.47 +- 0.41
Qwen2.5-Math-1.5BAIME20256.711.10 +- 1.56
Qwen2.5-7B-InstructMath50073.275.27 ± 0.41
Qwen2.5-7B-InstructAIME202510.021.10 +- 1.56
AuthorsOpenReview ↗

Reply(2/2)

Q4.What explains the substantial variation in improvement across models and benchmarks (Table 2)? Are there identifiable conditions under which ARM provides little benefit?

Thank you for this question.

One possible explanation is that the original activation distributions vary across models and tasks. Consequently, different tasks may require different magnitudes of redistribution. However, our paper primarily aims to identify this phenomenon and propose a general method, demonstrating that such redistribution does not degrade model performance and can yield improvements of varying magnitudes across tasks. Our work analyzes the underlying mechanism and demonstrates its effectiveness through a general intervention. But we do not perform fine-grained, task-specific selection. We leave the development of fine-grained adaptive ARM methods for future work.

AuthorsOpenReview ↗

Follow-up on Our Rebuttal

Dear Reviewer niDu,

I am writing to gently follow up on our rebuttal. During the rebuttal period, we give the further clarification of causality of our mechanisms and provide more context of the analysis and our method. We also reported the mean and standard deviation over three runs with different random seeds.

We hope these additional explanations and results have addressed your concerns. We would greatly appreciate it if you could take a moment to review our responses. If your concerns have been resolved, we would also sincerely appreciate it if you could consider updating your rating accordingly.

Thank you very much for your time and constructive feedback.

Best regards,

Authors

Thank you for your responses. Based on the discussions, I would like to keep my score and raise my confidence score to 3. Thanks.

Reviewer_r31zOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

It seems that this paper is studying the effect of prepending uninformative tokens to a prompt and why that often helps the LLM get better and more accurate responses to prompts. As far as I can tell their contributions are two things:

  1. Mechanistically, at least part of the reason this occurs is because near-zero activations on the normal prompt get shifted upwards in magnitude and affinely translated.

  2. They propose an inference-time mechanism and intervention that effectively implements this mechanism, but without having to prepend extra tokens.

Critical strengths

If this is true I think this is a pretty cool contribution. This is quite an interesting phenomenon, and finding the mechanism and then showing that you can take advantage of that mechanism post hoc in this way is quite powerful.

Importantly, I found the writing and the way the sections were organized very clear and well thought out. I thought the analysis and the figures made it easy to understand, and this is no small feat.

Critical weaknesses
  1. It seems that there's a contradiction in the way that you present previous work. In particular, in the intro and contributions, you say "we present a surprisingly counterintuitive finding named meaningless-token effect" and "We uncover a meaningless-token effect in LLMs", but then you also say that other work has found this observation before. I already knew about this phenomenon, for instance, from reading other work. I don't think you want to say that you claim this. In other parts of the paper and the related work, for instance, you say that your contribution is not the observation. There needs to be some change to the writing to make it all consistent.

In Contribution 1 you say "This runs counter to the common assumption that such tokens only add noise." It's not obvious to me that it is a common assumption. I did not have that assumption. In fact I had the assumption of the phenomenon you're studying in this paper. I think it's still really cool that you're studying this phenomenon mechanistically, but that undercuts your first contribution.

  1. In the ARM experiments, the improvements are quite small and it's hard to tell how much of them is noise. ARM samples random values, but Table 2 reports one number per cell with no error bars, and many gains amount to one or a few problems on the eval. Also, Gaussian noise recovers about half the gain (65.6 vs 67.0 over a 63.8 baseline, Table 21), so it's not clear the specific redistribution matters beyond adding noise.
Questions for authors

Please report ARM results over multiple random seeds with error bars. This control is important, so that we can see if the gain by ARM is larger than the seed-to-seed variation?

4 comments
AuthorsOpenReview ↗

Reply(1/2)

W1.It seems that there's a contradiction in the way that you present previous work. In particular, in the intro and contributions, you say "we present a surprisingly counterintuitive finding named meaningless-token effect" and "We uncover a meaningless-token effect in LLMs", but then you also say that other work has found this observation before. I already knew about this phenomenon, for instance, from reading other work. I don't think you want to say that you claim this. In other parts of the paper and the related work, for instance, you say that your contribution is not the observation. There needs to be some change to the writing to make it all consistent.

We sincerely thank the reviewer for this valuable suggestion. We agree with the reviewer that our current wording may have caused confusion and that our main contribution lies in understanding the underlying mechanism behind this phenomenon

Nevertheless, we would like to highlight an empirical distinction between our findings and those of prior work. As discussed in our Related Work section, previous studies have shown that inserting additional tokens at the end of the input can improve model performance [3]. Other studies have explored training models to utilize pause tokens in the reasoning process [5, 6, 7]. These approaches generally introduce additional tokens without disrupting the original input structure. In contrast, we find that inserting a long sequence of repeated, meaningless characters between the system prompt and the question can also improve reasoning performance, even though such insertion interrupts the continuity of the input. We believe this specific observation provides an interesting perspective on the phenomenon and further motivates our investigation into its underlying mechanism.

We have revised the corresponding words in the rebuttal revision to highlight the different findings of us.

W2.In Contribution 1 you say "This runs counter to the common assumption that such tokens only add noise." It's not obvious to me that it is a common assumption. I did not have that assumption. In fact I had the assumption of the phenomenon you're studying in this paper. I think it's still really cool that you're studying this phenomenon mechanistically, but that undercuts your first contribution.

Thank you for this comment. As discussed in our response to W1, our observation concerns a specific setting: inserting repeated meaningless characters between the system prompt and the question. Prior work reports different outcomes under related interventions. For example, [1] inserts repeated instances of “dog” between two questions and observes degraded behavior, while [2] tests filler tokens before the question without observing performance gains. Other studies primarily insert meaningless tokens after the question [3, 4] or introduce learnable special tokens through training [5].

We have revised corresponding words in the rebuttal revisions and remove the words like "This runs counter to the common assumption that such tokens only add noise."

[1] https://dropbox.tech/machine-learning/bye-bye-bye-evolution-of-repeated-token-attacks-on-chatgpt-models

[2] https://arxiv.org/abs/2607.22925

[3] https://arxiv.org/abs/2509.24884v1

[4] https://arxiv.org/abs/2607.03502v1

[5] https://arxiv.org/abs/2310.02226v1

[6] https://arxiv.org/abs/2404.15758

[7] https://arxiv.org/abs/2505.21024

AuthorsOpenReview ↗

Reply(2/2)

W3.In the ARM experiments, the improvements are quite small and it's hard to tell how much of them is noise. ARM samples random values, but Table 2 reports one number per cell with no error bars, and many gains amount to one or a few problems on the eval. Also, Gaussian noise recovers about half the gain (65.6 vs 67.0 over a 63.8 baseline, Table 21), so it's not clear the specific redistribution matters beyond adding noise.

Q1.Please report ARM results over multiple random seeds with error bars. This control is important, so that we can see if the gain by ARM is larger than the seed-to-seed variation?

Thank you for highlighting the need to assess variability across random seeds. We evaluated both ARM and the Gaussian control using three random seeds on MATH-500 and AIME 2025 with two models. The tables below report mean accuracy ± standard deviation, in percent. Gaussian perturbations can occasionally improve performance, as illustrated by the original single-run results in Table 21, but these benefits are not consistently maintained across random seeds. In contrast, ARM achieves higher mean accuracy than both the baseline and the Gaussian control in all four settings, with lower or equal variability. These results suggest that ARM provides more reliable improvements than the Gaussian control in the evaluated settings.

ModelBenchmarkBaselineGaussian, mean +- SDARM, mean ± SD
Qwen2.5-Math-1.5BMath50063.863.40 ± 1.6666.47 +- 0.41
Qwen2.5-Math-1.5BAIME20256.77.78 +- 1.5611.10 +- 1.56
Qwen2.5-7B-InstructMath50073.271.67 ± 2.2375.27 ± 0.41
Qwen2.5-7B-InstructAIME202510.07.77 ± 4.1421.10 +- 1.56
AuthorsOpenReview ↗

Follow-up on Our Rebuttal

Dear Reviewer r31z,

I am writing to gently follow up on our rebuttal. During the rebuttal period, we give the further clarification of our findings and core motivation and revise our text in the paper. Also we also reported the mean and standard deviation over three runs with different random seeds.

We hope these additional explanations and results have addressed your concerns. We would greatly appreciate it if you could take a moment to review our responses. If your concerns have been resolved, we would also sincerely appreciate it if you could consider updating your rating accordingly.

Thank you very much for your time and constructive feedback.

Best regards,

Authors

Thanks for the response, and for running the seeds. I also appreciate the change to the wording around the meaningless-token effect. I'll keep my score at weak accept.

Comments on the paper

1 comment

Dear Reviewers and AC,

Thank you for your valuable questions and feedback. During the rebuttal period, we have addressed the following key concerns:

  1. Evaluation on Recent Models: We conducted additional experiments on more recent model families, such as Qwen3.5, and observed consistent results that further support our findings.

  2. Stability of Our Method: We conducted multiple experimental runs and reported the mean and standard deviation to demonstrate the stability and consistency of our method.

  3. Further Clarifications and Manuscript Revisions: We provided additional explanations of several aspects of our work, including our analysis, comparisons with related work, and further discussion of the experimental results. We also revised the manuscript accordingly to improve its clarity.

We sincerely appreciate your time and constructive feedback. We kindly invite you to review our responses and would greatly appreciate any further comments or feedback.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.