Research map / Label-efficient and robust learning / Recognition and long-tail

FALCON: Fine-grained Activation Manipulation by Contrastive Orthogonal Unalignment for Large Language Model

NeurIPS 2025 · Label-efficient and robust learning › Recognition and long-tail · approach: Multilingual language representation pretraining

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes FALCON, a representation-guided LLM unlearning method that picks a low-mutual-information layer, contrastively unaligns activations using SVD-derived principal offset vectors, and resolves forget/retain gradient conflicts by orthogonal projection.

Problem

Machine unlearning of harmful or sensitive knowledge in LLMs with coarse loss combinations, which struggle to separate forget and retain knowledge and to balance removal with utility, and are vulnerable to knowledge recovery attacks.

Approach

Uses KDE-estimated mutual information between forget and retain activations to choose the intervention layer, then applies an InfoNCE-style contrastive loss toward SVD-based principal offset vectors and projects conflicting forget gradients onto the orthogonal complement of retain gradients.

Key results

• On Zephyr-7B, FALCON reaches WMDP-Bio 26.7 and WMDP-Cyber 25.3 with MMLU 57.4 and PPL 1.5, versus RMU at 34.5 / 28.9 with MMLU 57.4.
• On Yi-6B-Chat, FALCON reaches WMDP-Bio 27.7 and WMDP-Cyber 25.3 with MMLU 60.3, versus RMU at 50.8 / 33.5 with MMLU 59.6.
• On MUSE News with Llama-2-7b-hf, FALCON gets forget_knowmem_ROUGE 0.02 and forget_verbmem_ROUGE 0.03 with retain_knowmem_ROUGE 0.54, the best forget scores in the table.

Measured effects reported in the paper

Surprising findings

• Replacing the contrastive loss with RMU's loss makes unlearning ineffective in the ablation, indicating the contrastive term is critical.
• Removing gradient projection or using random vectors instead of POVs has only a minor effect on unlearning but degrades retention and increases jailbreak vulnerability.

Limitations

• Limited to text-based LLMs and relatively small models due to computational constraints.
• Some method details (e.g., the transformation f, epsilon, and the POV construction) are loosely specified and the paper says hyperparameters are dynamically adjusted.
• Experiments on TOFU/MUSE use first-order optimizers for fairness, while the main method uses Sophia.

Experimental setup

Datasets: WMDP WikiText MMLU MUSE TOFU

Benchmarks: WMDP MMLU WikiText perplexity MUSE TOFU

Metrics: WMDP accuracy MMLU accuracy perplexity forget_knowmem_ROUGE forget_verbmem_ROUGE retain_knowmem_ROUGE forget quality model utility training runtime recovery score under GCG

Baselines: LLMU SCRUB SSD RMU GradAscent GradDiff NPO SimNPO IdkDPO

Builds on: RMU LLMU SCRUB SSD Sophia

Closest papers by meaning

• Large Language Model Unlearning via Embedding-Corrupted Prompts (NeurIPS 2024 · similarity 0.88)
• ASRU: Activation Steering Meets Reinforcement Unlearning for Multimodal Large Language Models (ICML 2026 · similarity 0.88)
• On Effects of Steering Latent Representation for Large Language Model Unlearning (AAAI 2025 · similarity 0.87)
• Learning What to Forget: Improving LLM Unlearning via Learned Token-Level Importance (NeurIPS 2026 · similarity 0.87)
• Constrained Entropic Unlearning: A Primal-Dual Framework for Large Language Models (NeurIPS 2025 · similarity 0.87)
• Wisdom is Knowing What not to Say: Hallucination-Free LLMs Unlearning via Attention Shifting (NeurIPS 2025 · similarity 0.87)