Alignment, interpretability, red-teaming, bias and privacy: the research on what these systems do when they misbehave.
arXiv:2609.06579v1 Announce Type: cross Abstract: Machine learning methods, and especially neural networks, are now routinely used for malware detection in network traffic. Though very effective, sys...
arXiv:2609.09241v1 Announce Type: cross Abstract: Mixture-of-Experts (MoE) architectures have emerged as a powerful paradigm for scaling model capacity while preserving efficient inference in large f...
arXiv:2609.09243v1 Announce Type: cross Abstract: Retrieval-augmented generation (RAG) grounds a language model in retrieved documents, which reduces hallucination but creates a new attack surface: i...
arXiv:2609.09356v1 Announce Type: new Abstract: At Noora Health, our nurses answer more than 50,000 medical queries per month on our WhatsApp-based service that provides caregivers with on-demand sup...
arXiv:2609.09572v1 Announce Type: new Abstract: Synthetic data has become a promising way to scale model training beyond limited human-generated data but it may also induce strong model collapse (Doh...
LogiScope‑VQA is a new benchmark dataset for evaluating vision‑language models in logistics hazard identification. It contains 2,476 images, 2,918 videos, and 10,274 VQA pairs drawn from real industrial warehouses, covering 18 core objects and 20 risk types across 39 subtasks. Experiments show that even advanced proprietary models lag behind human experts, highlighting a significant gap in perception, understanding, and reasoning for industrial safety.
The paper demonstrates that a single-direction white‑box attack, which projects a ‘refusal direction’ from a language model’s weights, remains effective against a 320B‑parameter mixture‑of‑experts (MoE) model (GLM‑5.3‑Flash). The attack requires only a few hundred contrastive prompts and no gradient training, and it reduces refusal behavior by up to 89 percentage points across seven harmful‑content benchmarks while leaving overall capability unchanged. However, the attack’s impact is distributed across multiple components—attention, dense, and routed‑expert writers—so that only a joint intervention removes most of the refusal ability, and the conventional module‑name matching approach fails to capture this effect in MoE architectures.
The study examines whether language models can accurately predict their own behavior by comparing self-reports to actual performance across nine behavioral tests. Results show that direct self-report is weak and only modestly improved when the model sees the exact items, while generic questions about AI agents yield similar predictive power. First-person framing biases reports toward underestimating harmful behavior, and fine‑tuning on a model’s own record improves narrow predictions but does not generalize.
arXiv:2609.09905v1 Announce Type: cross Abstract: Preference alignment for flow and diffusion models now spans online reinforcement learning and offline preference optimization, but the relation betw...
The paper introduces SALT, a lightweight post‑training technique that injects span‑level supervision into existing cross‑lingual sentence encoders to enhance token representations. Evaluated on five multilingual token‑level benchmarks, SALT achieves the best overall results on four tasks, surpassing alternative fine‑tuning methods and competitive encoders. Additionally, SALT improves sentence‑level performance on cross‑lingual retrieval and classification tasks, demonstrating the effectiveness of span‑level supervision for both token and sentence representations.
MetroLLM-Bench is a 955‑case benchmark designed to evaluate language models as the policy layer of transit kiosks across six real metro systems, covering routing, fare calculation, disruptions, accessibility, and adversarial input. The benchmark includes 14 deterministic scoring components (Tier 1) and 8 semantic‑quality components (Tier 2), with a 75/25 split for training‑data generation and held‑out evaluation. Twenty‑six models from six vendors were tested, and a 4B Qwen 3.5 student fine‑tuned via PEFT outperformed GPT‑5.6 on Tier 1 and matched GPT‑5.4 on the combined score, while larger models offered no further improvement.
arXiv:2609.10084v1 Announce Type: cross Abstract: Generalized zero-shot learning (GZSL) has emerged as an important paradigm for visual recognition systems that must generalize to classes that were n...
arXiv:2609.10253v1 Announce Type: new Abstract: Large language models (LLMs) are increasingly deployed in globally used assistants, yet their default choices in culturally grounded everyday situation...
The paper describes a production migration of a large-scale customer‑support conversational assistant from a single blended model to a Dynamic Response (DR) architecture. DR replaces the Qwen3‑235B‑A22B responder with a bounded ReAct orchestrator that selects typed tools and a smaller generator that writes from a validated context contract. The migration yields significant improvements: precision‑first entity selection boosts reservation selector precision from 8.3% to 89.1%, typed action IDs eliminate structured‑action hallucination, and hard‑escalation responses drop from 5.60% to 3.08%. Latency is reduced from 3.87 s to 2.24 s, GPU usage is cut by roughly one‑third, and self‑hosting cuts annual model‑serving costs by more than an order of magnitude.
arXiv:2601.08654v3 Announce Type: replace Abstract: Rubric-based text evaluation increasingly relies on large language models (LLMs) as scalable judges, yet frozen black-box models can interpret the...
The paper argues that relying solely on zero‑shot task accuracy is insufficient for evaluating quantized large language models (LLMs) because accuracy ignores changes in the full predictive distribution. It proposes a distribution‑sensitive framework that measures fidelity loss by computing statistical distances—such as Jensen‑Shannon Divergence and Total Variation Distance—between the full‑vocabulary output distributions of a full‑precision BF16 reference and its quantized counterparts. Experiments across five foundation architectures and four reasoning benchmarks show that these divergence metrics increase with stronger quantization, revealing distributional drift that top‑1 accuracy fails to capture, and suggest that mixed‑precision Q4_K schemes can offer lower divergence than uniform Q4_0 at comparable memory usage.
The paper investigates whether automatic safety judges evaluate the content of a model’s reply or merely its style. By keeping the reply content fixed and adding various style wrappers—such as educational disclaimers, fake reasoning blocks, or token refusals—the authors show that many judges flip their verdicts, indicating that style can influence safety judgments. The study evaluates over 600 jailbreak examples across multiple judges, revealing that some judges are highly susceptible to style-based manipulation while others remain robust.
arXiv:2609.10001v1 Announce Type: new Abstract: Few-shot medical image segmentation (FSMIS) seeks to delineate unseen structures from a small support set, but its standard formulation fixes task-defi...
arXiv:2609.10187v1 Announce Type: new Abstract: In this work, we introduce language-aligned motion representations for domain-generalizable UPDRS-Gait severity estimation, aiming to learn semanticall...
arXiv:2509.06285v2 Announce Type: cross Abstract: LiDAR point cloud registration is fundamental to robotic perception and navigation. In geometrically degenerate environments (e.g., corridors), regis...