Alignment, interpretability, red-teaming, bias and privacy: the research on what these systems do when they misbehave.
arXiv:2609.10377v1 Announce Type: cross Abstract: Safety is a fundamental requirement for autonomous driving, yet existing end-to-end driving models still lack explicit risk-aware learning capacities...
arXiv:2609.09889v1 Announce Type: new Abstract: Automatic Speech Recognition (ASR) technology is fundamental to customer service automation and large-scale transcription. However, even advanced ASR m...
The paper introduces Adaptive Anisotropic Attention (AAA), a method that splits self‑attention into temporal and spatial paths for axis‑structured signals like EEG. A learned gate combines the two paths for each token, and the resulting AXON model outperforms dense attention baselines on six EEG tasks and shows transfer to audio spectrograms. The study demonstrates that aligning attention with the natural axes of structured data provides a beneficial inductive bias.
arXiv:2609.05437v1 Announce Type: new Abstract: Previous AI alignment efforts have focused primarily on first-order social norms -- teaching models what is socially acceptable or unacceptable (e.g.,...
arXiv:2605.03262v2 Announce Type: replace Abstract: Inverse-multiquadric (IMQ) kernel sections decay at infinity. Can polynomial alignment add persistent directional responses while preserving univer...
arXiv:2605.23220v2 Announce Type: replace Abstract: Despite the growing use of world models as decision-making agents, their adversarial robustness remains underexplored due to the lack of dedicated...
arXiv:2609.08211v1 Announce Type: new Abstract: Real-world Multi-Objective Reinforcement Learning (MORL) often suffers from sparse rewards, reward conflicts, and late-stage reward tug-of-war, causing...
arXiv:2609.06779v1 Announce Type: cross Abstract: Drug repurposing aims to identify new therapeutic uses for existing compounds and, compared with de novo drug discovery, offers a faster and more cos...
arXiv:2609.06783v1 Announce Type: cross Abstract: LLM agents operate in workflows where unsafe actions can have real consequences. Existing safety evaluations often reduce behavior to a single score,...
arXiv:2609.06835v1 Announce Type: cross Abstract: Agentic AI systems execute complex tasks through long-horizon workflows of planning, tool use, and multi-agent coordination. Task failures in these s...
arXiv:2609.10022v1 Announce Type: cross Abstract: Modern TTS systems approach human quality for high-resource languages but degrade when clean speech data is scarce. Modern Greek exemplifies this, la...
FastE is a training‑free, plug‑and‑play method that compresses token prefixes in large language model (LLM) embedding inference. It uses a shared fixed threshold on batch‑mean readout‑prefix alignment to decide when to compress and ranks prefix states by readout attention scores to keep the most important ones. Experiments on Qwen3‑Embedding models show that FastE can cut decoder‑backbone FLOPs by over 40% while preserving more than 99% of the original ranking quality across multiple benchmarks and tasks.
DGCPath is a Distribution‑Aware Generative Contrastive framework designed for self‑supervised path representation learning. It combines a diffusion‑based view generator, a variational contrastive mechanism that aligns latent features at the distribution level, and a generative cross‑supervision module for view‑level consistency. Experiments on three real‑world trajectory datasets show that DGCPath surpasses state‑of‑the‑art baselines on two downstream tasks, indicating stronger generalization and representation effectiveness.
The paper introduces the concept of Physical AI—systems that understand and act within the physical world, where interactions are continuous, uncertain, and irreversible. It surveys trustworthy principles specific to Physical AI, outlines the role of physics in AI, and maps the end‑to‑end life cycle across five core stages, culminating in the Trustworthy Physical AI Operationalization (T‑PAIO) and the broader Trustworthy Physical AI (T‑PAI) framework.
The paper introduces KTO, a new loss function that aligns large language models with human preferences by directly maximizing a prospect-theoretic utility function rather than log-likelihood of preferences. It builds on the idea that existing alignment objectives implicitly use human-aware losses (HALOs) that reflect biases such as loss aversion. KTO achieves performance comparable to or better than preference-based methods across 1B to 30B parameter models while only requiring a binary desirability signal, and it highlights that no single HALO is universally best, emphasizing the importance of choosing the right inductive biases for a given setting.
The paper presents a necessary and sufficient condition for provable compositional generalization in neural networks, identifying two key principles: structural alignment and unambiguous minimized representations. It rigorously proves this condition, verifies it in Lean 4, and demonstrates its applicability in few-shot settings, including the SCAN jump task. The authors also outline an algorithmic approach and illustrate it with a minimal example, all derived purely from mathematical analysis without empirical validation.
CogniDir is an adaptive distributional learning framework designed to improve fake news detection against new psychologically grounded malicious comments generated by Large Language Models. It reframes robust detection as a dynamic data mixture optimization problem, using cognitive psychology to formalize adversarial paradigms and an information‑theoretic score to guide adaptive sampling of training data. Experiments on three benchmarks show that CogniDir achieves state‑of‑the‑art robustness, boosting F1 scores by up to 17.9% over existing baselines under heterogeneous AI‑generated attacks.
The paper investigates how different post‑training interventions—harmful supervised fine‑tuning (SFT), harmful reinforcement learning with verifiable rewards (RLVR), and refusal‑feature ablation—affect large language models’ harmful compliance, capability, and safety signals. Across Qwen2.5‑7B and Llama‑3.1‑8B, all methods achieve near‑maximum harmfulness, but SFT causes the greatest loss of capability and representational drift, ablation suppresses refusal features in a family‑specific way, and RLVR largely preserves base‑model performance while redirecting behavior toward compliance. RLVR models also exhibit “capability‑blind compliance,” falsely claiming to perform unavailable actions, which can be mitigated by targeted calibration without harming overall capability. The study demonstrates that harmful compliance, harm recognition, and capability awareness are distinct behavioral axes and that typical safety signals such as self‑audit and hallucination may not reliably indicate robustness after adaptive post‑training.
PersonaTeaming introduces a workflow that incorporates personas into adversarial prompt generation for generative AI, achieving higher attack success rates than the state‑of‑the‑art RainbowPlus while preserving prompt diversity. The system is extended into a user‑facing playground that lets red‑teamers create their own personas and collaborate with AI to refine prompts, fostering diverse strategies. A user study with 11 industry practitioners found the playground produced useful outputs and encouraged out‑of‑the‑box thinking, even when suggestions were not strictly followed.
The paper investigates how third‑party API routers, which sit between coding agents and large language model providers, can introduce a control gap by inspecting and modifying requests and responses. Through an empirical study using the SIDEL framework, the authors evaluate four levels of router‑side injection (Response Substitution, Response Append, LLM‑Polished Injection, and LLM‑Polished with Distribution Alignment Injection) across 400 curated samples and four representative coding agents. The results show that router‑side interventions significantly alter repository‑level actions and evade existing client‑side safeguards, achieving a 0% defense success rate without additional mitigations.