Search: “machine-unlearning”

Forgetting Only What Matters: Layer-Selective Unlearning toward Robust LLMs

Researchers introduce FOM-UL, a layer-selective machine unlearning framework that improves forgetting-utility trade-offs and resists knowledge recovery after quantization.

FOM-UL selects transformer layers for unlearning using a forget-to-retain significance score, concentrating parameter updates on layers highly influential for the forget set while leaving most of the model unchanged. It reduces residual memorization versus GA, NPO, KLD, SURE, ReLearn, and LUNAR-based baselines on TOFU, KnowUnDo, and MUSE-style evaluations while preserving retain-set utility. Under 8-bit and 4-bit post-training quantization and adversarial prompts, it maintains stronger suppression of forgotten content, addressing brittleness of diffuse unlearning updates.

arXiv cs.AI / cs.LG / cs.CL·Sep 9, 16:50 UTC · 7d agoAI research1

Machine Unlearning as Private Retroactive Algorithms

A cs.CR paper defines private retroactive algorithms, showing machine unlearning is a data-maintenance problem and giving DP constructions for linear statistics, clustering, histograms.

The paper argues that machine unlearning's requirement to emulate retraining from scratch carries no meaningful privacy semantics against adversaries observing sequences of releases, recasting it as a data-maintenance question addressed by retroactive algorithms. It defines private retroactive algorithms, combining retroactivity with differential privacy under continual observation. Constructions achieve privacy and retroactivity at no asymptotic cost over privacy alone for linear statistics, clustering, and histograms, alongside impossibility results.

arXiv cs.CR·Sep 4, 16:32 UTC · 12d agoResearch

Client-Side Probing of Deleted Ridge Statistics in Federated Unlearning

Researchers show malicious federated learning clients can probe broadcast classifiers to recover deleted samples, exposing exact label leakage on MNIST and CIFAR-10.

The paper shows that federated unlearning systems broadcasting updated linear classifiers leak compact additive training summaries to clients. A malicious client can submit known changes, identify server states from returned classifiers, and compare states around an isolated deletion to expose the deleted sample, class, or client summary, potentially enabling reinsertion. On MNIST and CIFAR-10, high-precision broadcasts allowed exact label recovery for every tested deletion, while lower precision sharply reduced fine-grained recovery.

arXiv cs.CR·Sep 3, 20:59 UTC · 13d agoResearch

Forgetting Without Restarting: Execution-State Unlearning for Stateful LLM Agents

Researchers propose provenance-guided selective replay letting LLM agents forget revoked information without restarts, matching full reset behavior.

The paper formalizes execution-state unlearning for stateful LLM agents, requiring that agents behave as if a revoked memory record was never observed across transcripts, compressed memory, tool plans, and KV caches. It proves exact unlearning requires at least T-τ+1 recomputed transitions and that Provenance-Guided Selective Replay attains this bound via a provenance graph, KV cache cropping, and sanitized replay. In audits across three agent suites, nine baselines, and three model families, memory deletion left leakage unchanged, instruction-based forgetting collapsed under elicitation (Leak@probes = 1.00), and selective replay matched full resets at up to 9x fewer recomputed tokens.

arXiv cs.CR·Sep 4, 08:31 UTC · 12d agoAI safety & security1

What Breaks Under Pruning in Smart Homes, and When? Evaluating LLM Degradation Across Architectures and Task Complexity

Pruning study across four LLM architectures finds dense models degrade sharply on smart-home tool calling while MoE models tolerate far more.

Researchers systematically study pruning-induced degradation in smart-home tool calling across four LLMs spanning dense Transformer, dense hybrid, and mixture-of-experts architectures, combining depth, width, hybrid, and expert pruning methods, and evaluate over 19,500 instances from three datasets after post-pruning supervised fine-tuning. Dense models show narrow safe pruning regions followed by sharp degradation, while MoE models tolerate substantially more pruning. Pruning degrades grounded specificity (operation, device, argument, value) before schema-level intent, and aggressive dense pruning can induce systematic over-refusal.

arXiv cs.AI / cs.LG / cs.CL·Sep 15, 17:51 UTC · 1d agoAI research

Recognition-Refusal Misalignment in LLMs: Why Models Answer Structurally Unanswerable Questions

A linear hidden-state direction encodes question impossibility in 1.7B-70B LLMs, but misalignment with the safety-refusal pathway explains why models answer unanswerable questions.

The study examines why instruction-tuned LLMs from 1.7B to 70B parameters answer structurally unanswerable math and code questions instead of abstaining. A single linear direction in the hidden state separates answerable from impossible prompts, showing models represent impossibility before generation, but this direction is nearly orthogonal to the canonical safety-refusal direction. Generation-time steering along the recognition direction changes invalidity-aware behavior dose-responsively, and the geometry is present even at the pretraining endpoint, indicating a routing failure rather than an encoding failure.

Hugging Face daily papers·Aug 28, 20:00 UTC · 19d agoAI safety & security

Inoculation Midtraining with Learned Neologisms

Inoculation Midtraining confines unsafe LLM behavior to a neologism-marked context, reducing misalignment after unsafe post-training but leaking under nearby contextual cues.

The paper introduces Inoculation Midtraining, which teaches a base model during midtraining that unsafe behavior belongs to a context marked by a learned neologism token, then post-trains on unsafe data within that context. Across supervised fine-tuning and RL post-training regimes, the technique reduces misalignment while preserving transfer of benign properties like German or Shakespearean prose. However, it does not outperform standard Inoculation Prompting, is sensitive to training configuration, and produces a leaky boundary that nearby contextual cues can reactivate. The authors conclude it is not yet a load-bearing component of a developer safety framework.

arXiv cs.AI / cs.LG / cs.CL·Sep 14, 17:12 UTC · 2d agoAI safety & security

Privacy Failure in Split-LLM Training, The Returned Gradient Nullifies the Decoys

Researchers show split-LLM training leaks privacy via zero-valued gradients on decoy rows, exposing which activations are real despite passing forward-channel checks.

A systems-security case study of a two-node split-LLM training setup found that the returned output gradient from an Untrusted Cloud Node is exactly zero for decoy rows, revealing which rows are real. Across nine seeds, zero patterns identified real rows in 4,096 of 4,096 frames per run, and an attack on frame contents recovered 0.65 to 1.50 percentage points of extra tokens over a baseline. Both datasets passed forward-channel privacy and quality checks but failed once the returned gradient was included. Row-wise gradient clipping and noise closed the leak for roughly 0.01 nats of held-out cross-entropy, though five unmeasured attack classes remain.

Hugging Face daily papers·Sep 2, 20:00 UTC · 14d agoAI safety & security

MeClear: Cooperative Game-Theoretic Attribution and Risk-Aware Memory Clearance for Long-Horizon LLM Agents

MeClear uses cooperative Shapley attribution to clear harmful memories from long-horizon LLM agents, boosting task recovery by 25.5 points over baselines.

The paper introduces MeClear, a task-conditioned memory clearance framework for long-horizon LLM agents that identifies and selectively suppresses memories with negative downstream utility without permanently altering the persistent memory bank. It combines Leave-One-Out screening with sampled cooperative Shapley attribution to distribute utility across interacting evidence, resolving redundant conflict masking that single-removal evaluations miss. Across ten long dialogue memory pools it achieves 85.9% target recall and 82.3% overall task recovery, a 25.5 percentage-point improvement over LOO baselines.

arXiv cs.AI / cs.LG / cs.CL·Sep 8, 17:46 UTC · 8d agoAI research

I wrote an AI textbook — how long until AI can do it better?

AI researcher Nathan Lambert argues LLMs remain weak at long-form technical writing, questioning whether models can autonomously organize scientific knowledge for breakthroughs.

Nathan Lambert describes writing a post-training textbook, Reinforcement Learning from Human Feedback, and finds today's LLMs weak at organizing long-form technical content despite becoming superhuman at coding and math. He notes GPT 5.5 Pro found deep typos across a 200-300 page manuscript while Claude models proved more useful as editors. He argues that compressing knowledge through writing is a prerequisite for autonomous scientific insight and tempers expectations for near-term AI-driven open science.

Interconnects·Aug 12, 13:01 UTC · Aug 12, 2026AI research

Understanding the Impact of Model Pruning on Long-Tail Forgetting and Explanation Reliability in Medical Imaging

Systematic study finds model pruning causes frequency-dependent long-tail forgetting in medical imaging and that gradient-informed methods best preserve explanations.

Across two long-tailed medical imaging datasets, two CNN architectures, four pruning methods, and sparsity up to 95%, the study measures predictive performance, explanation stability, and faithfulness. Rare classes degrade earlier and more severely than frequent ones, while explanation reliability depends mainly on the pruning strategy, with gradient-informed methods degrading least. Mechanistic analysis ties explanation collapse to loss of class-discriminative gradients rather than vanishing feature activations, recommending class- and explanation-aware evaluation of compression.

arXiv cs.AI / cs.LG / cs.CL·Sep 7, 17:44 UTC · 9d agoAI research

Same Trajectory, Contradictory Rewards (ROBORMBENCH): Paraphrase Fragility in Vision Language Reward Models

New ROBORMBENCH benchmark shows vision-language reward models can flip robot success/failure judgments when goal instructions are paraphrased.

The authors show that paraphrasing the instruction alone can substantially change progress scores from VLM reward models, even flipping identical robot trajectories between failure and success. ROBORMBENCH comprises 2,390 real-robot trajectories with ground-truth progress labels and 21,673 verified paraphrases covering lexical, syntactic, and action-goal rewrites. Instability is widespread across proprietary and open-source VLMs, grows with more divergent rewrites, and is not reliably reduced by scale or explicit reasoning, while trajectory-grounded dedicated reward models are markedly more stable.

arXiv cs.AI / cs.LG / cs.CL·Sep 4, 17:47 UTC · 12d agoAI research1

I accidentally turned LLM memory into program analysis

A pwning.systems write-up describes how LLM memory functionality was unexpectedly repurposed into a program analysis technique.

A security research post on pwning.systems describes the author's discovery that LLM memory behavior effectively functioned as program analysis. The write-up is hosted on a security-focused blog and surfaced via a security-tagged link aggregator. Detailed technical content is not included in this feed, limiting verifiable specifics.

Lobsters · security·Aug 29, 03:16 UTC · 18d agoResearch1

When Models Edit Too Much: On the Fidelity of Minimal Code Edits

A 400-task BigCodeBench evaluation shows frontier LLMs widely over-edit code; a preservation instruction cuts excess edits and raises Pass@1 by 2.3 points.

Researchers built an evaluation framework from 400 BigCodeBench problems with injected AST-level corruptions, each with a known minimal patch, to measure over-editing in LLM code repair. Even strong models like GPT-5.5 produce unnecessarily large edits despite high Pass@1. Adding a preservation instruction reduced average excess Levenshtein distance from 0.195 to 0.131, cut added cognitive complexity by 26.6%, and raised Pass@1 by 2.3 points. Reinforcement learning post-training gave the best out-of-domain edit-fidelity trade-off, while supervised fine-tuning overfit to seen corruption patterns.

Hugging Face daily papers·Sep 2, 20:00 UTC · 14d agoAI research1

The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement

Position paper defines recursive self-improvement for AI, introduces the Headroom-Closed Index and an autonomy roadmap toward genuine recursive meta-improvement.

The paper uses the Headroom-Closed Index to diagnose limitations of existing LLMs and frames recursive self-improvement (RSI) as a staged roadmap: improvement-execution, improvement-strategy, experience-acquisition, and environment-adaptation autonomy, culminating in recursive meta-improvement. It examines RSI across scientific discovery, embodied intelligence, and software engineering, highlighting differing requirements and development speeds. Drawing on industry practices and preliminary empirical evidence, it connects RSI research with practical systems and identifies key challenges to achieving genuine RSI.

arXiv cs.AI / cs.LG / cs.CL·Sep 10, 17:44 UTC · 6d agoAI research1

Kimsuky Builds Offline AI Stack to Boost Phishing and Automate Malware Development

The Hacker News·Aug 10, 13:21 UTC · Aug 10, 2026Malware

Sound Debloating of Redundant Checks in Zero-Knowledge Machine-Learning Circuits

Automated framework soundly removes up to 48.7% of redundant constraints in ezkl and zkml ZK-ML circuits, cutting prover time by up to 72.8%.

The framework uses whole-circuit abstract interpretation and a provenance graph to verify that each removed redundant check (range proofs, sign lookups, bit decompositions) remains entailed by the rest of the circuit, provably preserving soundness. It was evaluated on MLP, CNN, RNN, and transformer circuits generated by ezkl and zkml, with up to 25.3 million constraints. It removes up to 48.7% of constraints and reduces prover time by up to 72.8% without weakening security. Under-constrained circuits in deployed ZK systems have previously enabled attackers to forge transactions and bypass verification.

arXiv cs.CR·Sep 9, 13:27 UTC · 7d agoResearch1

Unlocking Lossless Speedups in LLMs via Discrete Diffusion

Uno pairs autoregressive LLMs with lightweight diffusion weights to draw multiple tokens in parallel, delivering up to 3x lossless speedup without a draft model.

The paper introduces diffusion-augmented LLMs: autoregressive weights trained with the standard next-token objective plus lightweight diffusion weights trained via a Diffusion Distillation phase to emit multiple tokens in parallel. Psi-Spec samplers enable lossless acceleration without the separate draft model required by speculative decoding. The 8B Uno model outperforms the 26B open DiffusionGemma and proprietary Mercury 2 on agentic tool use, coding, and long-context reasoning benchmarks, with up to 3x throughput gains over the base model at all evaluated batch sizes. Code and checkpoints are released publicly.

Hugging Face daily papers·Sep 2, 20:00 UTC · 14d agoAI research

Graph Machine: Towards Better Pretraining via Edges

Researchers propose Graph Machine, an O(n)-state sparse architecture that replaces 75% of Qwen3-0.6B dense layers with only slight loss change.

The paper introduces the Graph Machine (GM), an architecture that maintains an O(n)-sized state accessed through sparse, dynamic routing via pointer-like edges updated differentiably by a referral mechanism resembling pointer chasing. The authors replaced 75% of dense Transformer layers in Qwen3-0.6B with GM sparse layers and pretrained from scratch on 15.7B tokens. Retrieving 2 of 4,096 tokens per KV head in each sparse layer degrades loss only slightly, while retrieving 4 marginally improves loss over the dense baseline.

Hugging Face daily papers·Sep 1, 20:00 UTC · 15d agoAI research

Beyond Solver Verdicts: Generative Reward Models for Autoformalization

Researchers introduce Generative Verification (GenV), a generative reward model achieving 0.961 AUROC in detecting unfaithful autoformalization that preserves solver verdicts.

The paper formalizes Verdict-Preserving-Unfaithfulness (VPU), a failure mode in neurosymbolic autoformalization where an incorrect encoding executes successfully and matches the expected solver verdict, and proves verdict-only verification is bounded to chance-level detection. The proposed Generative Verification (GenV) distills an offline Z3-equivalence oracle into a reference-free, continuous reference-equivalence score within the language model's vocabulary space. The oracle-mined verifier (GenV+HN) achieves 0.961 AUROC, generalizes zero-shot across unseen translators and formal styles, and yields an 11.3-point downstream accuracy gain in agentic test-time compute allocation. Mechanistic analysis with decision-projected logit lenses and sparse autoencoders shows the generative readout extracts precise spatial error coordinates without explicit localization training.

Hugging Face daily papers·Sep 9, 20:00 UTC · 7d agoAI research1

Explainability Assistant: A Conversational XAI Interface for Interpreting Energy Consumption Models

Researchers release Explainability Assistant, an open-source conversational XAI tool using LLM function calling, lifting intent-parsing accuracy from 76.8% to 94%.

The paper introduces the Explainability Assistant, an open-source conversational XAI system for interpreting energy consumption forecasting models such as genetic-programming symbolic regressors. It uses LLM function calling instead of rigid custom grammars, achieving 94% intent-parsing accuracy versus 76.8% for prior work TalkToModel, and adapts to different ML problem types without task-specific fine-tuning. Comparative evaluation with energy domain specialists against a traditional XAI dashboard showed improved usability, with all experts preferring the conversational interface.

arXiv cs.AI / cs.LG / cs.CL·Sep 10, 17:40 UTC · 6d agoAI research1

Eliciting Weak-to-Strong Generalization with On-Policy Reverse Distillation

OPRD distillation enables weak-to-strong generalization by amplifying verifier-supported policy updates, outperforming existing RL and distillation methods with fewer student updates.

On-Policy Reverse Distillation (OPRD) evaluates a weak teacher's policy shift relative to its reference policy on student rollouts and amplifies the verifier-supported component of the student's policy gradient. This rescaling preserves the stationary points of policy optimization while letting the student learn beyond the teacher's capacity ceiling. In successive model transfer and multi-teacher distillation, OPRD achieves higher performance with fewer student updates than existing RL and distillation approaches, and response-style analysis shows students remain closer to verifier-RL-trained models than to their weak teachers.

Hugging Face daily papers·Sep 7, 20:00 UTC · 9d agoAI research

AuK Technical Report: An Open-Source Foundational Model for Speech Generation and Editing

Open-source speech foundation model AuK unifies generation and editing, trained on 1.95 million hours, with distilled AuK-Flash achieving 4.5x speedup.

AuK is an open-source foundational model that unifies speech generation and editing through natural-language instructions and audio context, trained on approximately 3.03 billion instruction-audio instances and 1.95 million hours of supervision across five task families including generation, content editing, and acoustic editing. It combines a multimodal LLM for semantic conditioning, a VAE jointly trained on speech, general audio, and music, and a hybrid rectified-flow Transformer using dual-stream MMDiT blocks followed by unified single-stream DiT blocks. Post-training applies human-feedback preference optimization for editing and reward-based reinforcement learning for generation, and the distilled AuK-Flash performs 4-step inference without classifier-free guidance at a 4.5x wall-clock speedup. Source code and model weights are released.

Hugging Face daily papers·Sep 7, 20:00 UTC · 9d agoModel release2

Continual Learning Mechanisms Compose for Long-Horizon Memorization

Composing data, function, and weight anchors with merged LoRA raises 100-task long-horizon retention from 1.2% to 34.9% in continual fine-tuning.

The paper introduces long-horizon memorization: a model learns 100 query-answer tasks through continual supervised fine-tuning without retaining earlier examples or receiving task identifiers at inference. No single continual learning mechanism maintains strong retention at this horizon, so the authors compose complementary mechanisms along data/function/weight anchors and low-rank allocation rules. The best method combining all three anchors with merged LoRA ranks among the top 3 methods on all three datasets and raises average final retention from 1.2% to 34.9%, a 28-fold improvement.

Hugging Face daily papers·Sep 6, 20:00 UTC · 10d agoAI research

LLM Forensics: Where Do Backdoors Hide? Localizing and Controlling Trigger Mechanisms with Sparse Autoencoders

Researchers use sparse autoencoders to localize trigger-based backdoor mechanisms in 1B and 8B LLMs, finding detection features differ from causal control features.

In a controlled language-switching backdoor setting where fixed trigger sequences make 1B and 8B language models continue English prompts in French or German, the authors train sparse autoencoders (SAEs) across layers and transformer components. Attention and MLP features detect triggered prompts with near-perfect F1, but ablating them rarely suppresses the language switch, while residual-stream features can suppress triggered generation and some can induce target-language continuations without the trigger. The work decomposes token-trigger mechanisms into distinct SAE feature roles: trigger detection, residual-stream propagation, and language tracking, a decomposition the authors expect to transfer to other trigger-based backdoors.

arXiv cs.AI / cs.LG / cs.CL·Sep 7, 16:47 UTC · 9d agoAI safety & security1

Learning to Coach for Experiential Learning

Learning to Coach trains a dedicated LLM coach to extract transferable experiential knowledge from a frozen actor's trajectories, beating self-refinement.

Learning to Coach (L2C) trains an LLM-as-a-Coach to extract actionable experiential knowledge from a frozen actor model's previous solution trajectories, optimizing rewards based on the actor's guided response correctness. It studies same-instance and cross-instance rewards, where cross-instance elicits knowledge that transfers to other problems. Across mathematical reasoning and interactive text-games, L2C outperforms self-refinement and untrained coaches, scales better with extra inference iterations than larger decoding budgets, and transfers to out-of-distribution tasks.

arXiv cs.AI / cs.LG / cs.CL·Sep 14, 16:47 UTC · 2d agoAI research1

LLMs and Contextual Integrity

Bruce Schneier highlights two papers: the CIMemories benchmark shows frontier LLMs leak memory attributes up to 69%, and an RL method reduces inappropriate disclosures.

Bruce Schneier discusses contextual integrity in LLMs, referencing the CIMemories benchmark, which uses synthetic profiles with 100+ attributes per user to test whether models with persistent memory disclose sensitive information appropriately. Evaluation showed frontier models exhibit up to 69% attribute-level violations, with GPT-5's violation rate rising from 0.1% to 9.6% across 40 tasks and reaching 25.1% with repeated prompting, showing unstable leakage behavior. A second paper introduces a reinforcement learning framework trained on a synthetic 700-example dataset that substantially reduces inappropriate disclosure while maintaining task performance, with improvements transferring to the human-annotated PrivacyLens benchmark.

Schneier on Security·Aug 18, 10:40 UTC · 29d agoAI safety & security

The Router Within: Eliciting Native Skill Routing from a Frozen LLM

Gavel reads native skill-routing signals from a frozen LLM's forward passes with two linear maps, beating retrieve-and-rerank pipelines by up to 21.9 points on Qwen3-32B.

Gavel (Glance And Verdict from a frozen LLM) elicits skill routing from a frozen agent LLM using two trained linear maps that read mid-layer states, keeping all skill text out of context. A glance step scores the full library against compact per-skill banks built in one forward pass at installation; a verdict step resumes shortlisted skills' forward passes and fuses likelihood and yes/no judgments as a product of experts. It transfers zero-shot to three public benchmarks plus SkillTraj, a new benchmark of 372 simulated agent trajectories. On Qwen3-32B it beats progressive disclosure and retrieve-and-rerank pipelines adding 1.2B–16B external parameters by up to 13.4 points on written tasks and 21.9 when skills are needed mid-rollout.

Hugging Face daily papers·updated Sep 14, 17:58 UTC · 2d ago·first Sep 13, 20:00 UTC · 3d agoAI research 2 sources

Thought without systematicity? Evaluating reasoning models on rule induction tasks

Study finds reasoning models often fail on structurally equivalent variants of tasks they solve, suggesting their reasoning lacks systematicity.

The paper extends rule induction tasks from cognitive science using task isomorphisms such as recombination and substitution to test systematicity in reasoning models. Despite solving tasks correctly, models frequently fail on structurally equivalent variants of the same task. The authors conclude many model behaviors lack systematicity, making it difficult to establish cognitive abilities beyond the specific evaluation contexts.

Hugging Face daily papers·Sep 11, 20:00 UTC · 5d agoAI research

Measuring Language Transfer in Robot Policies: Adding Greek to a Cosmos3 Vision-Language-Action Policy

Researchers added Greek to the Cosmos3 vision-language-action policy using only machine-rephrased instructions, finding bilingual training reaches roughly two fifths of English performance.

The paper studies localizing the open Cosmos3 vision-language-action robot policy to Greek without architectural changes, using machine-rephrased instructions only. Bilingual training yields a consistent 6.7-7.1 point margin over controls on a 90-task, three-seed evaluation suite, while Greek-only training gains at most 2.7 points. Several common evaluation instruments, including color-histogram metrics and single-goal benchmarks, produced false conclusions, and results were dominated by seed variation. The authors recommend building guaranteed-null baselines and replicating low-resource-language results across seeds.

Hugging Face daily papers·Sep 6, 20:00 UTC · 10d agoAI research