Alignment, interpretability, red-teaming, bias and privacy: the research on what these systems do when they misbehave.
arXiv:2609.30840v1 Announce Type: cross Abstract: One-step generators enable high-quality visual generation with a single network evaluation, but their post-training is difficult: general implicit ge...
The paper introduces SAGE, a Sampling‑Aware Global Evaluation benchmark for species distribution modeling that uses GBIF records for training and sPlotOpen vegetation plots for presence‑absence evaluation across 5,771 plant species. It groups species by sampling effort and relative prevalence to assess how well single‑species and multi‑species deep‑learning SDMs perform under different data conditions. The study finds that Random Forests and DeepSDMs perform best overall, with DeepSDMs excelling for infrequently recorded species only when bias‑correction techniques are applied.
arXiv:2609.31603v1 Announce Type: cross Abstract: Large language models (LLMs) implicitly infer attributes of their users and adapt their behavior accordingly, yet these beliefs remain difficult to i...
StarWM introduces a self‑supervised attention routing mechanism that selectively applies reconstruction only to dynamically relevant regions of visual input. By combining a cross‑attention module with a dual‑stream decoder and stop‑gradient barriers, it balances faithful environmental dynamics capture with abstraction of irrelevant content. Experiments on DeepMind Control show that StarWM outperforms both reconstruction‑based and reconstruction‑free baselines, especially under distractor conditions, and preserves state attributes over long‑horizon imagination.
arXiv:2609.31122v1 Announce Type: new Abstract: Activation steering is a powerful training-free paradigm for controlling large language models at inference time. However, standard approaches estimate...
arXiv:2603.19742v3 Announce Type: replace-cross Abstract: Understanding the internal mechanisms of transformer-based large language models (LLMs) is crucial for their reliable deployment and effectiv...
arXiv:2609.31234v1 Announce Type: new Abstract: Problem. Ultra-high-resolution (UHR) remote sensing with vague user intents has two bottlenecks: visual tokens are expensive, and tool calling must be...
KuaFu is a unified behavior‑compression layer that reduces each user behavior item to 2–4 tokens, dramatically shrinking per‑item cache size while preserving fidelity through a four‑stage training process. In production across four profiling tasks, it matches or outperforms uncompressed single‑task models, boosts GPU throughput by 37–350%, and saves 190 GPUs. On public benchmarks it consistently beats prior compressors at the same compression ratio, and on RecBench a 4B KuaFu model outperforms its 8B counterpart by 1.90 points, contributing to a 1.37% lift in overall GMV on Tencent’s advertising and recommendation platform.
The paper contrasts design-ignoring and design-respecting world models for epidemiology, formalizing study design as constraints on a model’s latent-world interface, action mechanism, observation likelihood, and target readout. Using a large cluster-randomized test-negative trial, the authors compare a design-ignoring case-count model with a design-respecting test-negative observation model, finding that while both reconstruct factual trajectories similarly, the design-ignoring model’s intervention contrasts vary substantially across resampling experiments, whereas the design-respecting model’s contrasts remain stable. The study demonstrates that factual reconstruction alone does not guarantee alignment with study design, and that evaluating whether fitted contrasts preserve design-implied observation-process invariances offers a more precise assessment.
Large language models (LLMs) are increasingly used for legal research, but their fixed training cutoffs and reliance on static knowledge clash with the evolving nature of statutory law. This study introduces a benchmark of 312 expert‑validated, time‑sensitive German statutory QA pairs that examine two temporal failure modes: post‑cutoff staleness and recency bias. Five LLMs were evaluated under four inference settings, and the results show that retrieval‑augmented approaches that enforce temporal validity significantly improve performance, while web search yields unstable gains and a pronounced recency bias.
Audio large language models (Audio LLMs) often fail to transcribe English‑Mandarin code‑switching speech, exhibiting language omission, translation‑instead‑of‑transcription, and hallucination. By applying Direct Preference Optimization (DPO) with 100 K preference pairs, the models learn to preserve mixed‑language content rather than translate, leading to significant reductions in mixed‑error rates (up to 89.6% in‑distribution). The study demonstrates that DPO can effectively align multilingual Audio LLMs for accurate code‑switching transcription.
The paper investigates whether audio large language models (Audio LLMs) can detect when their own transcriptions are unreliable. It finds that the models are poor at self-assessment and that existing methods offer limited detection. By leveraging audio-encoder representations, the authors develop a lightweight predictor that accurately flags unreliable transcriptions and can prompt user clarification without altering the underlying model.
arXiv:2609.31431v1 Announce Type: new Abstract: Accurate segmentation of axons in 3D microscopy data is important for analyzing white-matter organization, but dense ground truth labels are expensive...
The paper investigates why retrieval‑based open‑ended evaluation fails in medical fact verification. By creating two detailed taxonomies—one for retrieval‑stage errors across five quality dimensions and another for verifier‑reasoning errors across six steps—the authors automatically label evidence quality and reasoning errors using an LLM‑as‑Judge pipeline. Their large‑scale stress tests across multiple retrieval methods and verifier models show that increasing model size, reasoning effort, source breadth, or medical fine‑tuning does not eliminate these failure modes, indicating fundamental limits of the retrieve‑then‑verify paradigm in open‑ended medical contexts.
The paper introduces the concept of Evolutionary Safety for recursive self-improving AI, focusing on how safety properties evolve as an AI system and its successors change. It identifies key risks such as intent drift, error accumulation, and safety-property erosion, and presents a taxonomy covering agent state, model state, evaluation, environment, and update mechanisms. The authors propose methods for discovering and evaluating evolutionary risks, and outline governance principles for modification, selection, authorization, provenance, and recovery, while highlighting open problems for maintaining safety in persistent, adaptive, and recursively self-improving systems.
The paper introduces skill cascading attacks, where a malicious goal is spread across multiple seemingly benign skills, causing harmful outcomes when combined. It presents SkillCascade, an automated red‑teaming framework, and releases SkillCascade‑Bench, a benchmark of 213 validated cascading test cases across various agent systems and domains. Experiments show that these cascaded interactions reliably induce harmful behaviors while evading existing per‑skill scanners and runtime monitors, revealing a gap between component‑level integrity and system‑level safety.
JevAdvBench introduces the first adversarial benchmark for reinforcement‑learning‑based calibrated decision (RLCD) models, providing 812 typed questions across 66 scenarios and a black‑box attack suite of 9,744 single‑edit variants. The benchmark evaluates attacks by comparing each perturbed decision to the model’s own clean decision and to an identical re‑run, revealing that rewording changes decisions by only 1.2 percentage points while certain injected opinions can flip 12.1% of decisions and lower confidence below 0.8 in 38% of cases. These findings demonstrate that RLCD models can be significantly misled by seemingly innocuous input edits, underscoring the need to treat the state as untrusted in applications.
The paper investigates how allocating a query budget to structural depth rather than surface variation improves jailbreak success against the SAGE self‑check defense. By using a best‑of‑N approach over a code‑completion encoding, the authors achieve 67%, 22%, and 15% success rates on three open‑weight targets—far exceeding the 4.7% and 3.0% rates of single‑draw encoding and character‑search methods. The study demonstrates that depth of encoding and breadth of variation independently undermine transform and gate defenses, and that repeated sampling can inflate perceived robustness.
MedHal is a large-scale synthetic dataset created to detect hallucinations in medical AI-generated text. It includes diverse medical sources and tasks that cover both intrinsic and extrinsic hallucinations, providing a substantial volume of samples for training. The authors demonstrate that models trained on MedHal outperform general-purpose hallucination detectors, highlighting its usefulness for medical AI development.
Assistive robots increasingly operate in many human-centered environments and perform various human-robot interaction (HRI) tasks, such as object delivery. However, most existing HRI systems rely on R...