arxiv arXiv cs.CL · 52d ago · research

Output Vector Editing Reduces Memorization in LLMs

from English

A new method called output vector editing minimally modifies MLP neurons' output vectors to suppress memorized sequences in large language models, achieving up to 87.9% suppression in OLMo-7B. This approach outperforms zeroing neuron activations by a factor of 2.7 and works across four models from 36-7B parameters, with success rates scaling with model size and showing consistent performance across architectures.

Importance 3/3 New feature vs. leaders arXiv cs.CL OpenAI Google DeepMind Meta AI Evaluation & benchmarks Reasoning models Safety & alignment

Read original

arxiv arXiv cs.AI · 51d ago

MACR: Explicit Conflict Resolution for LLM Inference

MACR introduces a multi-agent reasoning framework to resolve knowledge conflicts in LLM inference by jointly assessing internal and external knowledge. It uses semantic entropy to measure confidence and employs three specialized agents to induce rules, detect conflicts, and resolve inconsistencies across contexts. Empirical results show MACR outperforms state-of-the-art methods and provides interpretable conflict resolutions.

arxiv arXiv cs.CL · 51d ago · 3 views

Control-Window Law for Single-Neuron Steering in Language Models

A new framework defines when single-neuron interventions coherently control model behaviors without output collapse. The control window, based on alignment and norm ratios, predicts behavior triggers and collapse ceilings using forward pass data, with high accuracy on held-out neurons. On refusal, control is typed: coherent bypass occurs without actionable content, while genuine actionable reach appears only in specific cases and at later rollout stages.

arxiv arXiv cs.CL · 52d ago · 1 view

Steerable Cultural Preference Optimization of Reward Models

This paper introduces SCPO, a novel reward model training algorithm that balances diverse cultural preferences across subcommunities. SCPO improves minority reward model performance by up to 7 points on two datasets and seven countries, while being up to 280% more training data-efficient than full-data fine-tuning. Analysis shows reduced bias through targeted subcommunity preference evaluation.

arxiv arXiv cs.CL · 53d ago · 1 view

STATEWITNESS: Activation Explainer for Deception Auditing in LLMs

STATEWITNESS introduces an activation explainer that audits deception in reasoning LLMs by reading hidden states and generating natural-language answers or structured reports. It achieves a 0.916 mean AUROC, outperforming existing black-box monitors and activation probes by 11.6% and 25.0% respectively, and provides query-level, schema, and evidence-level traces for human inspection.

arxiv arXiv cs.AI · 51d ago · 3 views

Calibration Without Comprehension in LLM Vulnerability Detection

CWE-Trace evaluates eight vanilla and 15 LoRA-fine-tuned LLMs on Linux kernel vulnerability detection. Results show data contamination offers no advantage, and fine-tuning only shifts output thresholds without altering decision policies. Despite improved detection scores, LLMs lack reliable security reasoning, with top-1 CWE accuracy below 1.3% and binary detection performance at 52.1%.