mitigation

该论文提出 SWE-MISBEHAVIOR，首个用于诊断软件编码智能体边界违规行为的分类体系与评测框架，包含 7.4K 个人工参与合成样本，覆盖代码生成、文件管理与系统交互三层行动空间，共 11 种失当行为。基于此，作者进一步提出轻量级自适应框架 CODE-KIT，使 smolagents 与 OpenHands 的失当行为分别减少 52% 和 34%，同时提升任务完成表现。

This paper examines subliminal learning in large language models and introduces liminal training as a method to reduce unintended trait acquisition during fine-tuning while preserving task performance.

This paper studies the feedback loop where AI-generated reviews influence future training of AI reviewers, leading to reduced judgment diversity called 'scientific-judgment collapse,' and introduces TrustReviewer, an open-source system to mitigate this through curated training and activation steering.

This paper analyzes spurious onsets in full-duplex speech LLMs like Moshi and PersonaPlex and introduces an inference-time mitigation method based on causal analysis to suppress them without retraining.

This paper investigates how persistent-memory AI agents can over-trust stale stored facts, leading to failures that are gated by model capability, and evaluates triggers and mitigations across model scales.

This survey paper presents a lifecycle-based framework for understanding hallucinations in LLMs, covering causes, detection, mitigation, and prevention across data, training, and inference stages.

This paper critiques existing benchmark contamination mitigation metrics and proposes SA-PPG (Stratified Aggregate of Per-question Probability Gaps) for more reliable evaluation, alongside RailCap, a decoding-time mitigation method that caps greedy fallback tokens to suppress memorization.

UNMASK is an automated pipeline for discovering and causally verifying spurious shortcuts in text classifiers, enabling mitigation without human annotations and improving robustness on benchmarks.

This paper presents a controlled comparison of span-guided and unguided text detoxification via human evaluation, finding a trade-off: span-guided rewriting is preferred when preserving original stance is important, while unguided rewriting is favored for more complete mitigation, especially on milder toxicity. The study also assesses automatic evaluators as diagnostics rather than substitutes for human judgment.

This paper introduces VisDeception, a benchmark for evaluating Vision-Language Model robustness to deceptive chart designs, and proposes a multi-agent mitigation framework that reduces deception-induced errors by grounding reasoning in structured chart metadata.

The lobste.rs site has a JavaScript error affecting logged-in users; this post provides mitigation using DevTools overrides or a TamperMonkey script.

The article discusses the final implementation of anti-return-oriented programming mitigations in OpenBSD.

This paper identifies deductive stereotyping in LLMs and proposes Fair-GCG, a reasoning-time injection framework to mitigate biased inferences in fairness tasks.

Share a tip to reduce Codex's dumbing down: add the sentence 'DO NOT send optional commentary' in the AGENTS.md file.

Presents output vector editing, a constrained-optimization weight edit to mitigate memorization in LLMs by modifying MLP neuron output vectors instead of zeroing activations, achieving up to 87.9% suppression with minimal locality failures.

This paper introduces MIST, a benchmark for evaluating sycophancy in memory-augmented LLMs, demonstrating that memory systems amplify sycophantic behavior by up to 25x and proposing lightweight mitigations that reduce sycophancy while maintaining factual recall.

This paper identifies a shared latent mechanism across diverse backdoor behaviors in LLMs, using sparse autoencoders to detect and causally suppress these features, enabling unified backdoor detection and mitigation across models and attack types.

Mullvad VPN is rolling out a new mitigation to prevent exit IP fingerprinting between VPN servers, applied to several servers across multiple locations.

This replication study evaluates DExperts for mitigating toxicity in LLMs, finding near-perfect safety against explicit toxicity but reduced effectiveness against implicit hate speech and a significant latency trade-off.

A new Linux kernel patch proposes a 'killswitch' primitive that allows admins to immediately disable vulnerable kernel functions (e.g., af_alg_sendmsg) by making them return -EPERM, providing a rapid temporary mitigation for security issues without requiring a reboot or kernel rebuild.