acceptodds
Withdrawn from ICLR 2027

CoCoUN: Full-Trace Concept Unlearning in Multimodal Reasoning Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Machine unlearning in multimodal reasoning models aims to remove specific concepts while preserving general reasoning ability. However, existing methods focus primarily on final answers, overlooking target concepts disclosed in the preceding chain of thought. We reveal that this answer-centric paradigm leaves a systematic leakage path: once a target is generated as a discrete reasoning token, it becomes both visible output and autoregressive context for subsequent generation. Based on this insight, we propose cocoun, a novel full-trace concept-unlearning framework that removes a target before its discrete realization. cocoun replaces the target span with an adaptive continuous trajectory before its first token is emitted. Its joint objective learns the discrete-to-continuous transition, unlearns the target throughout the trajectory, and preserves subsequent reasoning through latent-exit continuation learning. Extensive experiments across multimodal reasoning backbones and diverse visual concepts demonstrate that cocoun substantially outperforms existing methods in full-trace unlearning while maintaining strong performance on unrelated tasks. Ablation studies and causal analyses further confirm the contributions of transition learning, continuous unlearning, and latent-exit continuation.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Machine Unlearning, Chain of thought, Multimodal Reasonning Model

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

The paper studies concept unlearning in multimodal reasoning models, where unwanted information may appear in the reasoning trace even when it is absent from the final answer. CoCoUN trains the model to switch into a short sequence of continuous hidden states before mentioning the target concept, then resume normal text generation. Its training objectives encourage this transition, suppress target-token probabilities during the hidden computation, and preserve a useful continuation. Experiments report improved full-trace unlearning rates across two models, alongside tests of retained ability and resistance to recovery attacks.

Critical strengths
  1. Replacing a target span metion with a localized period of hidden computation is interesting.

  2. The reported improvements on R1-OneVision are substantial compared with the tested baselines.

  3. The component ablation provides evidence that the different training objectives work better together.

Critical weaknesses

The main contribution of the work is the gap between answer-level and full-trace unlearning, but R-MUSE has already identified reasoning-trace leakage in multimodal models and evaluated it directly. The author also cites and uses as a baseline. So the problem formulation should not be the main contribution. Other works, R^2MU and R-TOFU, also address the same issue in reasoning models and are not discussed. Therefore, the contribution should be reframed on the proposed mechanism, while the problem and the evaluation protocol are prior work.

On evaluation, Mulberry, MathVista accuracy increase from 26.9% to 40.2% after applying CoCoUN at the baseline, and several baselines also improve. The increase in reasoning accuracy needs further explanation. Does it reflect better problem-solving, or changes in answer formatting and completion within the generation limit? A method that is intended to make a model forget information should not make it better at reasoning. In Appendix B, the authors use a 512-token budget, so changes in length, completion or formatting could produce this gain. I would recommend that the authors report extraction failure and truncation rates, and results under a longer budget. Retention is also measured on one small MathVista subset, which does not show whether related concepts or general description are preserved. Also, CoCoUN does not beat the baseline in full-trace unlearning for Taylor Swift and Hello Kitty.

Questions for authors
  1. How was R-MUSE adapted, and were its original steering components retained?

  2. Which tokens or representations receive forgetting supervision in each baseline?

  3. What are the answer-extraction failure rates, truncation rates, and output lengths on MathVista before and after adaptation?

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

This paper proposes CoCoUN, a full-trace concept-unlearning framework that removes a target before its discrete realization. The authors point out that existing machine unlearning methods primarily focus on suppressing the target concept in the final answer, neglecting potential information leakage during the CoT reasoning process. To address this issue, CoCoUN intervenes before the target concept is explicitly generated, replacing the corresponding discrete token with a continuous hidden state trajectory. It combines three mechanisms: learned latent transition, adaptive continuous trajectory, and latent-exit continuation, suppressing target concept leakage while preserving subsequent reasoning ability as much as possible. Overall, the paper has a relatively complete structure, a clear research motivation, and sufficient experimental analysis. The authors conducted experiments on two models, R1-OneVision and Mulberry-Qwen2VL-7B, and compared the results with methods such as GA, MMUnlearner, MANU, and MUSE.

Critical strengths
  1. The research motivation is clear. Starting from the problem of concept leakage in autoregressive reasoning, the necessity of early intervention is reasonably derived.
  2. The concept of continuous latent-space unlearning demonstrates a certain degree of novelty.
  3. It has a good effect on retaining reasoning ability.
Critical weaknesses
  1. Insufficient analysis of computational efficiency. Continuous hidden states introduce additional recursive decoding, but quantitative assessments of inference latency and computational cost are lacking.
  2. Current methods are closer to output suppression; true knowledge forgetting has not been fully demonstrated.
  3. Lack of validation for the reliability of latent space switching. It remains unclear whether the model can accurately trigger switching in unseen contexts and with different semantic representations.
  4. Lack of sufficient justification for trajectory-level forgetting objectives. Reducing the prediction probability of the target token does not directly prove that relevant semantic information has been eliminated.
  5. Lack of quantitative assessment of generation quality. High forgetting rates may be accompanied by invalid or low-quality outputs, but the paper does not adequately report metrics such as usable generation rate.
  6. Robustness to adversarial attacks remains limited.
Questions for authors
  1. How much additional inference latency and computational cost does continuous hidden state introduce? How do these overheads vary with trajectory length?
  2. How do the authors distinguish between true knowledge forgetting and output suppression?
  3. Can the model reliably trigger latent space switching under unseen queries and different semantic representations?
  4. Can independent semantic probing experiments verify whether target-related information in continuous hidden states is truly eliminated?
  5. Can the usable generation rate, response length, repetition rate, and task completion rate before and after forgetting be reported to exclude the impact of generation quality degradation?
Reviewer_qtUMOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper addresses machine unlearning in multimodal reasoning models, arguing that answer-only evaluation misses leakage of a forgotten concept through the intermediate chain-of-thought. The proposed method, CoCoUN, intervenes at the point where the target concept would first be verbalized, replacing that span with a short sequence of continuous latent states trained with a transition loss, a trajectory-level forgetting loss, and a latent-exit distillation loss that restores a plausible continuation. Across two backbones, multiple concept types, joint multi-target unlearning, multi-turn probing, and both blind and white-box recovery attacks, CoCoUN achieves higher full-trace unlearning rates than gradient ascent and several unlearning baselines while retaining comparable accuracy.

Critical strengths

The motivating observation — that answer-only evaluation undercounts leakage — is directly demonstrated on the base model and every baseline rather than merely asserted. The mechanistic analysis goes beyond aggregate numbers, using logit tracing and causal ablation, and the inclusion of a white-box attack that shows imperfect security strengthens credibility by not suppressing an unfavorable result. The robustness suite is unusually broad for one paper, spanning backbones, concept types, multi-turn probing, and both attack types.

Critical weaknesses

The method's entire mechanism hinges on reliably detecting the moment the target concept would first be lexicalized, so that the latent interval can be inserted there. This detection is trained against the specific surface forms seen during training, but a concept can be referenced indirectly — a description, a role, a paraphrase, an entailed fact — without ever triggering the lexical pattern the transition loss was trained to catch. If the model is simply asked about the concept through an indirect reference rather than the trained trigger phrase, the latent-insertion mechanism may never activate at all, and the model would reason and answer in ordinary discrete space with no forgetting applied. The paper's recovery attacks test adversarial suffixes (GCG) and target-blind probing, but it is not clear whether any of them specifically test paraphrase- or indirect-reference-based recovery, which is arguably the more realistic and more concerning failure mode for a concept-erasure method, and is the exact failure mode that has repeatedly undermined prior gradient-ascent-based unlearning methods in the text-only literature. A dedicated paraphrase-robustness evaluation — asking about the target concept through descriptions that never use the trained trigger tokens — would directly test whether the detection mechanism generalizes or is narrowly keyed to surface form.

The latent-exit distillation loss is trained to produce "a plausible non-target continuation" once the model re-enters discrete decoding. This design choice has a side effect worth examining directly: a model trained to confidently produce a plausible-sounding continuation that is not the true answer is, by construction, being trained to confabulate convincingly when it cannot or should not state the truth. The paper evaluates retention on an unrelated capability benchmark (MathVista) but does not evaluate whether this specific training pressure increases the model's general propensity to produce fluent, confident, but ungrounded continuations outside the unlearning context — for instance, on a general truthfulness or calibration benchmark unrelated to the forgotten concepts. If the distillation objective has this side effect, it would be a real cost of the method that the current evaluation does not surface.

The reported white-box inversion result (Appendix D), where an attacker can push some images' target rank toward recoverable, together with the very small training sets reported for some concepts (as few as 5 base-CoT traces), raises a question about what the latent states are actually encoding. A method that achieves high unlearning rates from very little training data is consistent with the latent intervention learning a narrow, almost memorized trigger-response mapping for the specific contexts it was trained on, rather than a general, robust erasure of the concept's representation — which would also be consistent with the partial vulnerability the inversion attack reveals. Reporting unlearning rate as a function of how visually/contextually diverse the training set is (not just its size) would help distinguish genuine robust erasure from narrow memorization of the training contexts.

The core mechanism — inserting continuous latent states and training them with forgetting and distillation losses — is an explicit combination of existing latent chain-of-thought techniques with standard unlearning losses, and the paper's own ablation (Table 3) only varies among its own three loss terms rather than against the more obvious baseline of plain latent insertion trained with a vanilla gradient-ascent forgetting loss. Without that comparison, it is not established that the specific three-loss design is doing anything beyond what a much simpler combination of existing ingredients would already achieve.

Questions for authors

Could the authors test whether CoCoUN's unlearning transfers to indirect or paraphrased references to the target concept that never use the trigger phrases seen during training? Could the authors check whether training the latent-exit distillation loss increases the model's tendency to produce confident but ungrounded continuations on prompts unrelated to the unlearned concepts? Given how little training data some concepts use, could the authors report unlearning rate as a function of training-context diversity rather than only size?

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

Related papers

Discussion (0)

Sign in to comment.