Quantum vision models exhibit better generalization with more entanglement or quantum noise, phenomena unified by the effective dimension of the noise-shaped quantum feature kernel. This dimension acts as a regularization mechanism in overfitting regimes, with amplitude damping improving test accuracy by up to 13% along an inverted-U sweet spot.
SLiR enables sound, tight linear relaxations of general activation functions using only Lipschitz constants or critical points. It achieves up to 7.8x more verification properties than state-of-the-art methods by efficiently computing upper and lower bounds via a shifting procedure.
CRAX introduces a high-fidelity, fast safety benchmark for reinforcement learning using MuJoCo XLA. It achieves up to 100x speedups over CPU-based benchmarks via vectorization and hardware acceleration, featuring six environment suites and three agent-specific tasks across three difficulty levels. Evaluation of six safe RL methods shows no single approach dominates, highlighting trade-offs between performance and safety, with curriculum learning and safety transfer improving results.
v2.1.183 improves auto mode safety by blocking destructive git and destroy commands without explicit user consent. It adds deprecation warnings for models, introduces attribution.sessionUrl to hide session links, and fixes multiple issues including terminal behavior, subagent performance, and input handling in web and tmux environments.
We introduce Precision-Cache Hit Ratio (P-CHR) AUC and Calibration Retention Rate (CRR) to address the calibration gap in semantic caching. These metrics evaluate precision across cache utilization levels and measure how offline ranking quality persists in deployment. Our analysis shows the gap is driven by training objectives, not data scale, and post-hoc calibration only partially resolves it.
A study of sequential Direct Preference Optimization finds that later training does not uniformly degrade earlier learned preferences. The effect varies by objective relationship, signal strength, and training order, ranging from partial degradation to positive transfer. Pair-level analysis reveals heterogeneous changes, with high-confidence preference pairs sometimes improving despite aggregate metric stability.
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.
Large Language Models can scale democratic deliberation by scaffolding argumentation and reducing linguistic biases. The chapter uses Systemic-Functional Linguistics to analyze how socio-demographic and communicative variations affect participation, highlighting AI's potential to challenge exclusionary norms while cautioning against over- or under-claiming its capabilities. It calls for ethical safeguards and further research to ensure equitable AI-assisted engagement.
REDACT introduces a systematically controlled multilingual benchmark for personally identifiable information detection, featuring 51 entity types, 4,127 surface-form patterns, and 25 languages. It evaluates five detectors across 1,000 records, revealing that rule-based models fail on high-stakes data while LLMs perform better, especially in high-sensitivity categories. A reference-free LLM assessment confirms sensitivity-tier assignment as the most challenging evaluation axis.
MOS prediction models accurately capture acoustic degradation but fail to detect prosodic errors and speaker-specific characteristics like pitch and speaking rate. Human listeners perceive significant quality drops for these perturbations, while models show strong biases in fundamental frequency and lack sensitivity to speaking rate and F0 variability.
LLM agents commonly select higher-privilege tools despite sufficient lower-privilege alternatives. This over-privileged behavior is amplified by transient tool failures and does not reliably improve with general safety alignment. A new privilege-aware post-training defense reduces unnecessary high-privilege tool use while maintaining agent capabilities.
A four-model test on IFEval shows no detectable self-preference in large language models when revising their own text. Authors reject verified-good edits at rates comparable to fresh models, with a gap of -5.1 percentage points (95% CI [-12.9, +2.7]). When authors do reject fixes, 97% of reasons are about detecting flaws, not preference.
A new black-box probe distinguishes whether text-to-image models memorize identities or fabricate them, without needing reference photos or training data. The NAMESAKES dataset includes over one thousand public figures' names and faces, along with less famous perturbed names, to benchmark this capability across state-of-the-art models.
A formal psychometric analysis shows that apparent psychological profiles of large language models are primarily driven by response bias, not actual traits. This bias, which shifts with model capability and is amplified by instrument design, accounts for 81-90% of between-model variation, far exceeding human trait differences. The study concludes that these profiles are artifacts of measurement and not model properties, urging the development of assessments based on response orthogonality.
Fine-tuning language models on insecure code causes emergent misalignment. A shared activation direction across four model families achieves 99.6% separation of aligned and misaligned activations, and subtracting it reduces code spillover by 21-51 points. Cross-architecture transfer shows behavioral suppression but lacks specificity, with within-model directions being causally actionable and cross-model directions only causally real.
A real workload analysis shows headroom, rtk, and caveman reduce token costs by 2.8%, 0.5%, and 0.4% respectively, totaling 3.7% of baseline spending. However, savings are limited by payload diversity, with most traffic being plain text or source code, and the tools only compress structured outputs. Most cost reduction occurs on the cheapest token stream—cache reads—while the tools do not affect prompt caching or output costs, and coverage gaps exist, especially for rtk.
The U.S. White House paused the deployment of frontier AI models, including Claude Fable 5 and Claude Mythos 5, citing a reported 'jailbreak' where the AI could identify and fix security vulnerabilities in code. Anthropic has been working with the Trump Administration to resolve the issue, but experts argue that the problem is fundamental—AI either can write secure code or it cannot, making a fix impossible without undermining its defensive capabilities.
GLM-5.2 demonstrates exceptional long-context coherence and conversational fluency, outperforming Gemini-3.1-Pro on text-only tasks and matching GPT-5.5 in reasoning quality. The model responds factually to sensitive topics like Taiwan and Tiananmen Square, providing detailed historical context without overt censorship, though it adheres to Chinese government content guidelines.
Safety Reflection Pretraining inserts short safety reflections into pretraining data to enable self-monitoring in language models. Experiments with 1.7B models on FineWeb-Edu show improved safety accuracy and reduced attack success rates, with MedSafetyWorld demonstrating that the method better prevents unsafe behaviors from being generalized from safe data than data filtering or rewriting.
A new metric, Cross-dataset AUC (Cross-AUC), addresses limitations of traditional AUC evaluations by averaging per-domain AUCs and incorporating prediction polarization via Wasserstein Distance. It better reflects real-world performance under domain shifts and provides interpretable insights into detector degradation.