Skip to content
TheChatArchiveThe Chat Archive
gpt-4.1-mini-2025-04-14· OpenAIHugging Face LMArena Human Preference 140K

Please paraphrase the following passage into a third person voice with a strong…

From a Chatbot Arena battleThe same person sent this prompt to two anonymous models at once. Compare it with o4-mini-2025-04-16’s answer.
  1. User

    Please paraphrase the following passage into a third person voice with a strong scientific research bent: For graph unlearning, we propose D2DGN, a knowledge distillation-based approach designed to delete elements of graph while preserving the integrity of the retained graph structure. By partitioning graph knowledge into retain and forget sets and using response-based and feature-based distillation, D2DGN achieves efficient unlearning without repetitive retraining. Experimental results on real-world datasets demonstrate significant improvements in unlearning performance, with up to 43.1% higher AUC in edge and node unlearning tasks compared to existing methods. In the domain of multi-modal recommendation, we introduce MMRECUN, the first method to unlearn item data in MMRS. Leveraging a novel Reverse Bayesian Personalized Ranking (BPR) objective, MMRECUN diminishes the influence of user-item interactions within the forget set while reinforcing interactions in the retain set. This enables effective unlearning of outdated preferences, revoked licenses, or legally mandated removals. Experiments on benchmark datasets show recall improvements of up to 49.85% over baseline methods and computational speed gains of up to 1.3× compared to retraining. For unlearning in LLMs, we propose UNSTAR, a framework that introduces the novel concept of anti-samples. These are generated using misleading rationales to reverse learned associations, enabling fine-grained targeted unlearning. Unlike vii prior approaches, UNSTAR selectively removes specific knowledge without degrading related information, offering a precise and efficient unlearning strategy. This approach has significant implications for privacy-preserving AI and adaptive model modification. To complement these methods, we introduce ECLIPSE, a black-box evaluation framework designed to rigorously assess unlearning vulnerabilities. Our framework combines adversarial prompt generation, targeted attacks to probe unlearned content. Initial results reveal that current unlearning techniques often fail to ensure true knowledge erasure, instead suppressing information that remains accessible through adversarial probing.

  2. Assistant

    The study introduces several novel methodologies aimed at advancing the field of unlearning across diverse machine learning domains. For graph unlearning, the proposed method, D2DGN, employs a knowledge distillation-based framework to selectively remove graph elements while maintaining the structural integrity of the remaining graph. By partitioning graph knowledge into retain and forget subsets and applying both response-based and feature-based distillation techniques, D2DGN achieves efficient unlearning without necessitating repeated retraining. Empirical evaluations on real-world datasets demonstrate that D2DGN significantly outperforms existing approaches, achieving up to a 43.1% increase in AUC for edge and node unlearning tasks.

    In the context of multi-modal recommendation systems (MMRS), the research presents MMRECUN, the first approach designed to unlearn item data within MMRS. MMRECUN utilizes a novel Reverse Bayesian Personalized Ranking (BPR) objective to attenuate the influence of user-item interactions associated with the forget set while reinforcing those in the retain set. This mechanism facilitates the effective removal of outdated user preferences, revoked licenses, or legally mandated data deletions. Benchmark experiments reveal that MMRECUN improves recall by up to 49.85% relative to baseline methods and achieves computational speedups of up to 1.3 times compared to full retraining.

    For large language models (LLMs), the framework UNSTAR is introduced, which innovatively incorporates the concept of anti-samples—data points generated with misleading rationales designed to invert learned associations. This enables fine-grained, targeted unlearning of specific knowledge components without compromising related information, distinguishing UNSTAR from prior methods that often degrade overall model performance. The approach holds significant potential for privacy-preserving AI applications and adaptive model updates.

    Complementing these unlearning techniques, the study also proposes ECLIPSE, a black-box evaluation framework aimed at rigorously assessing unlearning vulnerabilities. ECLIPSE integrates adversarial prompt generation and targeted attacks to probe for residual unlearned content. Initial findings indicate that many current unlearning methods fail to achieve complete knowledge erasure, instead merely suppressing information that remains recoverable through adversarial probing. This highlights critical challenges in verifying the efficacy of unlearning protocols.

Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.