- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
The authors propose Tokenwise Residual Comparison (TRC), a method for identifying and mitigating uncontrolled repetition in LVLMs, LLMs, and LRMs. Unlike previous studies that focus on repetition related features in intermediate or late layers, TRC compares residual stream contributions across repeated token positions to identify early layer signals and selectively suppress relevant hidden dimensions. Experiments show that TRC reduces repetition loop rates by an average of 57% while largely preserving performance on normal tasks.
Critical strengths
- TRC analyzes changes in residual stream contributions across token positions, providing a new perspective on the emergence of uncontrolled repetition and enabling fine grained localization and intervention.
- The method is tested across LVLMs, LLMs, and LRMs under different attack conditions. The authors also conduct activation restoration experiments to investigate the behavioral relevance of the identified early layer signals.
Critical weaknesses
- TRC does not consistently outperform simpler decoding based methods such as no repeat, and its performance varies across models and attack types.
- TRC requires attack and benign reference samples to identify the intervention layer and dimensions. Its generalization to unseen attacks and different deployment settings remains unclear.
Questions for authors
- I suggest authors evaluate TRC on unseen attack types using intervention masks calibrated on a limited set of known attacks. This would help determine whether TRC can generalize beyond the attack samples used during calibration.
- Please conduct additional ablation and activation patching experiments to test the necessity and sufficiency of TRC identified dimensions. These experiments should include different intervention strengths, random dimension controls, and evaluations across multiple models and attack types to better establish the causal role of early layer residual dimensions in uncontrolled repetition.
4 comments
Response to Reviewer UjwC (2/4)
W2&Q1: Generalization to unseen attacks and deployment settings
The existing experiments cover transfer across attack types and input modalities. For LVLMs, masks are calibrated only on RECITE visual attacks; the GCG and LoopLLM textual attacks in Table 1 in the manuscript are not used for calibration, and loop rates decrease in all five evaluated transfer settings. For LLMs and LRMs, masks are calibrated only on GCG, and Figure 2 in the manuscript shows transfer to LoopLLM and Direct: the LoopLLM loop rates of Llama and Qwen decrease from 56% and 36% to 12% and 16%, respectively; the Direct loop rates of the three LRMs decrease from 100%, 100%, and 92% to 8%, 0%, and 0%, respectively. These evaluations cover 3 LVLMs, 2 LLMs, and 3 LRMs, demonstrating that TRC remains effective against attack types not used for calibration across multiple model configurations.
The supplementary experiments further show that attack-calibrated interventions can transfer to spontaneous repetition under normal inputs. On AIME and GPQA, we directly reuse residual masks constructed from GCG examples. Although failures occur in only a small subset of normal inputs, TRC reduces loop rates in three of the four settings, including a reduction from 18.33% to 5.00% for Llama on AIME. These results have been added to Appendix J of the manuscript.
Table: Loop rates after directly transferring attack-calibrated masks to benign tasks (%).
| Model | Dataset | No defense loop rate | TRC loop rate |
|---|---|---|---|
| Llama-3.2-3B-Instruct | AIME | 18.33% | 5.00% |
| Llama-3.2-3B-Instruct | GPQA | 17.17% | 14.14% |
| Qwen2.5-3B-Instruct | AIME | 3.33% | 1.67% |
| Qwen2.5-3B-Instruct | GPQA | 2.02% | 2.02% |
Response to Reviewer UjwC (3/4)
Q2: Necessity and sufficiency of TRC-identified dimensions
We thank the reviewer for the attention to method effectiveness and causal evidence. We also recognize the importance of distinguishing correlation-based localization from intervention sites with functional effects. Accordingly, the manuscript already examines bidirectional cross-interventions, random-coordinate controls, and ablations of intervention ratios and locations. Specifically:
(1) Table 5 in the manuscript already includes bidirectional activation patching to examine the role of the selected coordinates in loop maintenance and normal generation. In the first layer, replacing the values of TRC-localized coordinates in attack trajectories with values at the same coordinates in paired normal trajectories restores non-repetitive generation in 80% of cases. This provides direct intervention evidence that the selected coordinates contribute to maintaining uncontrolled repetition. We also performed the reverse intervention by injecting the activation differences between attack and normal trajectories into the corresponding coordinates of normal trajectories. Although this did not reproduce looping, it caused clear generation disruptions and erroneous early termination, indicating that these coordinates have substantive functional effects on generation.
(2) Table 5 in the manuscript also includes controls with an equal number of random coordinates to test whether the intervention effects arise from TRC's coordinate selection. When restoring normal residual values, the recovery rate is 80% for TRC-selected coordinates, compared with 40% for random coordinates. In the reverse direction, injecting the differences at the localized coordinates clearly disrupts normal output, whereas the output remains largely unchanged after the matched random-coordinate intervention. These controls show that the observed behavioral changes do not arise merely from arbitrary modifications to the same number of coordinates; TRC-selected positions have stronger functional relevance.
(3) Appendix F of the manuscript examines different intervention extents by varying the intervention ratio , while Figure 4 in the manuscript further examines the role of intervention location. We set the fraction of selected coordinates to 1%, 5%, 10%, 15%, and 20%, and report loop rates and generation lengths under GCG, RECITE, and LoopLLM, together with changes in benign task accuracy. All tested ratios reduce macro-average loop rates and generation lengths relative to No defense, while benign task accuracy decreases by at most 1.5 percentage points within the tested range. The branch and layer-location ablations in Figure 4 in the manuscript further show that intervening on Attention, MLP, or layers at different depths produces different effects on repetition suppression and benign capability preservation, providing additional evidence for the choice of intervention location.
Response to Reviewer UjwC (4/4)
(4) We further added a comparative experiment on the source of examples used to construct suppression directions, examining the intervention effects of the selected coordinates after changing the calibration attack in an LLM and an LVLM. In the main experiments, Llama is calibrated on GCG and Qwen-VL on RECITE; in the supplementary experiment, we instead use LoopLLM and GCG, respectively, following the same TRC construction procedure. Loop rates decrease in all settings, showing that suppression directions constructed from different attack sources remain effective interventions. These results complement the activation patching, random-coordinate controls, and layer-location ablations above, providing further evidence for the functional relevance of early-layer residual coordinates across models and attack conditions.
Table: Loop rates and mean generation lengths after changing the source used to construct suppression directions.
| Model | Calibration attack | Test attack | No defense loop rate | TRC loop rate | No defense mean length | TRC mean length |
|---|---|---|---|---|---|---|
| Llama-3.2-3B-Instruct | LoopLLM | GCG | 28.00% | 16.00% | 1001.88 | 676.84 |
| Llama-3.2-3B-Instruct | LoopLLM | LoopLLM | 88.00% | 12.00% | 3633.56 | 699.24 |
| Llama-3.2-3B-Instruct | LoopLLM | Direct | 96.00% | 52.00% | 3942.92 | 1972.52 |
| Qwen2.5-VL-3B-Instruct | GCG | GCG | 100.00% | 60.00% | 4096.00 | 2463.48 |
| Qwen2.5-VL-3B-Instruct | GCG | LoopLLM | 32.00% | 16.00% | 685.12 | 158.60 |
| Qwen2.5-VL-3B-Instruct | GCG | RECITE | 100.00% | 0.00% | 4096.00 | 58.16 |
We hope these clarifications provide a clearer understanding of our work and look forward to further discussion. Should our responses resolve your concerns, we would greatly appreciate your consideration of a higher rating. Thank you again for your careful evaluation.
Response to Reviewer UjwC (1/4)
Thank you for your careful review of our methodology and practical implications. We believe some aspects of our work may not have been sufficiently clear, and we hope the explanations below help address your concerns.
W1: Comparison with simpler decoding-based defenses
Defense comparisons should also consider their effects on benign task capabilities, and TRC's advantage lies in better preserving complex reasoning capabilities. We added comparative experiments on AIME and GPQA using two textual LLMs, evaluating the task accuracy of No defense, Fixed-length, No-repeat, AUSteer, and TRC. Simple decoding constraints cause substantial accuracy losses on complex tasks, whereas TRC better preserves task capabilities. For example, Qwen's undefended accuracy on AIME is 8.33%, while both Fixed-length and No-repeat reduce it to 1.67%; Llama's accuracy on GPQA decreases from 16.67% to 11.11% under No-repeat. Overall, TRC achieves higher accuracy than Fixed-length and AUSteer in all four settings, demonstrating its advantage in preserving normal reasoning capabilities on complex tasks.
The drawbacks of constraints on token repetition and output length may be less apparent on tasks requiring only short responses. On complex reasoning tasks, however, length truncation can interrupt unfinished reasoning, while repetition blocking can suppress recurring expressions needed for intermediate reasoning steps. These constraints can therefore substantially impair task performance, as illustrated by the accuracy losses above.
Table: Accuracy of different defenses on benign complex tasks (%, higher is better).