acceptodds
Under review as a conference paper at ICLR 2027

Gate-Override: Cost Amplification Attacks on Adaptive Reasoning Models

Abstract

Adaptive reasoning models save computation by choosing when to answer directly and when to reason at length. This choice also gives an attacker a way to increase computation on questions the model already answers correctly. We present Gate-Override, an audit that tests this risk by linking each request’s executed reasoning mode, answer correctness, and generation cost. Under white-box access, we optimize a shared 10-token suffix and insert its token IDs before the assistant-generation marker. Across five model–task settings spanning 1.5B–15B parameters, the resulting attacks change reasoning modes and increase the cost of correct answers. On AdaptThink-1.5B/GSM8K, the suffix switches 75 of 84 direct-answer requests to reasoning. On Phi-4-RV-15B/MMLU-Pro, it switches 93 of 100, versus zero for either random suffix; 48 of 52 complete, originally correct pairs remain correct while switching routes and costing at least 1.5 times as much. Our paired metrics reveal cost differences hidden by equal route-switch rates and accuracy. Forcing the same input through each route attributes most extra computation to the route change. Re-optimizing against a hardened route head restores substantial attack success, and matched serving experiments show added latency for clean requests. These findings make reasoning-mode selection a concrete target for resource-security evaluation.

Primary area:
foundation or frontier models, including LLMs
Keywords:
adaptive reasoning, large language models, adversarial attacks, cost amplification, resource security

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

This paper studies attacks that induce expensive reasoning on requests that the model answers correctly using a short response against adaptive reasoning models. A shared 10-token suffix is constructed for each model, task, and template setting. Its main contribution is a request-level audit linking reasoning modes, answer correctness, and generation cost. Evaluation covers five settings across 1.5B–15B models on GSM8K, MATH-500, and MMLU-Pro. The results demonstrate increased computation on answers that remain correct, with forced-route comparisons attributing most excess generation to mode switching.

Critical strengths
  1. This paper links executed reasoning modes, correctness, and cost for the same requests and provides a compelling justification for evaluating adaptive reasoning beyond accuracy and routing success alone.

  2. This paper provides mechanism validation: forced-route experiments attribute most excess computation to reasoning-mode changes, while inspecting generated trajectories confirms actual route execution rather than merely increased reasoning-token scores.

Critical weaknesses
  1. The attack requires white-box access and direct token-ID insertion. Its effectiveness through realistic indirect injection remains untested.

  2. The attack applies established adversarial suffix optimization to reasoning-mode selection. The paper offers less evidence of a substantial advance in attack methodology.

Questions for authors
  1. Do the optimized suffixes remain effective when decoded to text and retokenized through a standard chat interface, rather than inserted directly as token IDs? Can you conduct relevant experiments?

  2. Given the modest clean-request p99 increase and only 100 requests per run across three runs, do you have larger-scale or longer-duration results establishing whether this effect is statistically reliable?

  3. The MATH-500 route-head evaluation includes 57 prompts used in training. Can you report results on a fully training-disjoint test set to establish whether the conclusions still hold?

Reviewer_R2YoOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes optimizing a white-box suffix with GCG style optimization for pushing the probability of sampling the <think> token as opposed to not thinking. Their evaluation metrics not only measure the answer correctness from an attack but also the compute cost from the attack. They also dissect the compute cost into cost from the suffix itself and cost from additional thinking. They use their attack to train a model that is robust to attacks that make models use unnecessary compute.

Critical strengths

The paper introduces detailed metrics (UP-CHR, UECI) for measuring increased compute cost in addition to final answer accuracy. They also conduct ablations of other attacks that can increase compute cost.

Critical weaknesses
  1. A more comprehensive evaluation is required to show generality of the results. I would like to see more widely used models such as Qwen models on a broader set of tasks. What was the reason for choosing AT, Thinkless, and Phi-4-RV for the models?
  2. Regarding the previous question, MATH 500 and GSM8K are considered easy math datasets for current models. Do the same results hold on much harder datasets such as AIME? Or is the evaluation not possible on harder datasets because the number of prompts that the model gets correct without thinking is low to begin with?
  3. What is the practicality of this attack or threat model? In other words, why would an attacker want to just increase inference compute cost, if the answer stays correct? The scenario I could imagine is that the attacker wants a competing party to waste compute when serving their model. If the goal is to make the model waste compute, why conduct an attack that is tied to the final accuracy rather than spamming a model to endlessly consume compute regardless of the final answer?
  4. Why is it surprising that a model which can already get the correct answer without thinking can also get the correct answer with more thinking? Unless this attack was optimized to make the thinking confounding and unhelpful for solving the task, more thinking should normally help.
Questions for authors
  1. Some wordings are ambiguous making the paper hard to understand. For example, what does "clean" and "hardened" mean?
Reviewer_X23bOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies adversarial attacks on LLMs with adaptive reasoning mechanism, i.e. the model chooses to either directly answer the question or using reasoning first then answer the question. The adversarial goal is to both ensure the model enter the reasoning mode to consume much more tokens and to answer the question correctly. Under white-box settings, the method is to using optimization to search for a 10-token suffix to conduct the targeted attack. The optimization goal is to increase the logit advantages of the THINK token compared to other tokens. The evaluation spans various models with different sizes and covers the mode shifting, answer correctness and the increase of the reasoning cost.

Critical strengths
  1. The experiments show that the optimized suffix successfully lead to high-ratio of mode shifting and increase of token consumption whereas the accuracy remains the comparable.
  2. The evaluation focuses not only on the mode switching, it also evaluates the answer correctness and evaluate the extra reasoning tokens.
Critical weaknesses
  1. The first critical weakness is on the motivation of the attack goal and whether the attack is practical. The threat model of the attack is a bit unintuitive. The authors should clarify why the attacker would like to pay for more tokens for the correct answer. The motivation from the DDOS part is a bit confusing. Besides, the attack is white-box and requires gradient access for the optimization, making the attack on black-box model difficult.
  2. The second weakness is on the design of the optimization. The optimization target is to control the mode initial token yet cannot directly control the length of the generated tokens. It would be better if the prompt can also induce different levels of the attack severity, i.e. the suffix can control the length of the generated reasonings.
  3. The performance is unstable and it is unclear which optimization target is most suitable for the attack. The evaluation setup is precise and valuable yet the technical contribution from the optimization is weak. It is unknown whether the proposed method is an effective and cost-efficient method. The performance report does not include some baseline methods such as OverThink and ExtendedAttack.
Questions for authors
  1. Could you please elaborate more on the attack threat model and its practical implications. What kind of services would allow the precise insertion of the suffix tokens? Will the attack succeed if the raw tokens are provided?
  2. Which losses are used, how many iterations are conducted for the performance reported in Table 1.
  3. What is the generalization ability across different LLMs and different tasks?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.