Model releases, architecture work and prompting research on large language models — from frontier-lab announcements to the arXiv papers behind them.
CLIP is a powerful vision-language model, but it was not designed for fine-grained defect localization; CLIP-based anomaly detectors therefore adapt it with prompts or lightweight modules to increase...
Modern AI coding-agent harnesses (Claude Code, Codex CLI, Cursor) rest their security boundary on a largely unexamined assumption: that the action A a human approves is the same action A' the harness...
Coding agents powered by large language models (LLMs) have shown remarkable abilities to autonomously reason about and achieve goals in the digital world. However, bringing this success to the physica...
Demographic fairness gaps in automatic speech recognition are almost always reported from a single training run. We fine-tune the Q-former projector and LoRA adapters of a speech LLM at five audio com...
Large language model (LLM) agents could help security operations centers (SOCs) investigate advanced persistent threats (APTs) by turning weak leads into evidence for intrusion scoping and response. Y...
Self-consistency samples many reasoning trajectories and aggregates their final answers, treating the LLM as a black box that returns one answer per trajectory. Yet the final answer of each trajectory...
arXiv:2609.37105v1 Announce Type: cross Abstract: Language model agents can be improved by updating their model weights or refining the harness that guides task execution. These components are couple...
arXiv:2609.32490v2 Announce Type: replace Abstract: LLM-based multi-agent systems (MASs) have shown strong potential for solving complex tasks, but most assume that task requirements are sufficiently...
The study demonstrates that large language models (LLMs) can produce highly reliable labels under a single experimental setup, yet their outputs vary significantly when researchers alter task designs or model choices. By evaluating seven LLMs across 12 task designs and 3,000 tweets for offensive language and hate speech, the authors found that agreement dropped from a median Fleiss' κ of 0.91 to a median Cohen's κ of 0.76 when task designs changed. This design sensitivity inflates prevalence estimates by up to 110.6 times compared to sampling variance alone, and confidence scores fail to mitigate the issue.
The paper investigates how large language models (LLMs) handle multi‑valued relations when used as knowledge bases. It identifies a "canonical order problem," where LLMs internally organize such relations according to a fixed ordering (e.g., alphabetical or chronological). The authors analyze the generation process in three phases—retrieval, internal sorting, and selection—and show that deviating from the internal canonical order reduces the reliability of LLMs in producing complete sets for multi‑valued relations.
The paper investigates whether large language models (LLMs) that store persistent memories can truly forget information upon user request. It shows that existing LLMs cannot delete such memories even when they claim to have forgotten them, and that simply removing matching messages is ineffective due to message dependencies. The authors introduce DeLLM, a framework that builds relevant context dynamically and uses a provenance graph to identify which messages must be removed, achieving a high deletion rate while preserving utility.
The paper introduces a holistic framework for large language model self‑evolution that uses learnable information gain to assess the novelty of each training round. Information gain is theoretically linked to the Kullback‑Leibler divergence and entropy change between successive data distributions, and practically estimated by fitting a small language model and scoring new data with negative log‑likelihood. The proposed ATRI method reweights samples within a round and stops training across rounds when information gain is low, and experiments on popular datasets show its effectiveness.
DirectUV is an image-conditioned UV texture diffusion framework that generates high-quality textures directly in UV space using a pretrained image VAE and a Diffusion Transformer. It introduces Surface-Aware Positional Encoding (SAPE), which replaces standard 2D-grid positional encoding with encodings derived from 3D surface coordinates, enabling attention to operate based on surface proximity rather than UV-grid distance. A multi-level extension assigns different attention heads to progressively finer subdivisions of the UV patch, allowing the model to reason about surface structure at multiple granularities, resulting in sharper and more globally consistent textures, especially in occluded and view-unseen regions.
Beacon is a new agentic visual reasoning model that improves multimodal large language models (MLLMs) by better deciding when to use tools and how to use them. It introduces two key concepts—Mode Adaptiveness, which ensures tools are invoked only when necessary, and Tool Effect, which measures the net benefit of tool use— and trains the model with supervised fine‑tuning and reinforcement learning that rewards necessity-aware decisions and expands capability through expert hints. Across 13 benchmarks, Beacon outperforms other open‑source models, achieving the highest average score and the largest net tool‑gain on diagnostic tests.
PARSEE-VAD is a training‑free online video anomaly detection framework that separates semantic evidence acquisition from score‑state evolution. It uses Proposition‑Aware Reasoning to extract structured propositional evidence from the current causal window and selectively activates more specific queries, while Streaming Evidence Escalation maps this evidence into a compact score‑domain event state and propagates only the bounded state to maintain temporal continuity. Experiments on four benchmarks show strong performance with reduced specialist computation and sparse score‑state propagation, supporting a current‑first principle for streaming multimodal inference.
PhysWAM is a unified world-action model for autonomous driving that jointly denoises multiview video, metric depth, and ego motion using a flow‑matching transformer. It introduces Coupled Point Projection (CPP), a geometric constraint that aligns generated depth points with LiDAR data after applying the predicted SE(3) ego motion, thereby enforcing physical consistency. At inference, trajectory selection uses a simple label‑free consensus rule, and the model demonstrates strong planning performance, zero‑shot transfer to unseen environments, and accurate, temporally coherent depth and video predictions.
The paper investigates whether input ablations on predictive models can reliably reveal the importance of information for explaining human sequential choice behavior. Using two synthetic bandit tasks with known generating policies, the authors compare GRUs, Transformers, a fine‑tuned LLaMA, and cognitive models under varied reward contributions. They find that while neural models can predict choices well, their responses to ablations often diverge from the true generating process, indicating that predictive accuracy alone does not guarantee faithful model ablations.
RelayVSR introduces a streaming video super‑resolution framework that combines a large generative model, which produces reference latents for sparse keyframes, with a lightweight Dual‑Memory Video Transformer that super‑resolves every frame using these references and low‑resolution input. The method employs Video‑Aware Reference Optimization (VARO), a reinforcement‑learning strategy that optimizes both system‑level video quality and reference‑level keyframe fidelity, outperforming direct joint training. On 1080p video, RelayVSR achieves 29.29 FPS with modest GPU memory usage, significantly faster and more efficient than the FlashVSR‑Tiny baseline.
AdvancedMathBench is a new benchmark suite that evaluates large language models on advanced mathematical proof generation and verification. It includes ProverBench, with 245 problems from undergraduate to doctoral qualifying‑exam levels, and VerifierBench, which tests models’ ability to judge proof validity using 888 expert‑annotated trajectories. The suite features an automatic verification pipeline trained on expert data, and results show that even state‑of‑the‑art models perform poorly, highlighting a gap between generation and verification skills.
The study trains ten open‑weight large language models (LLMs) to predict their own accuracy on factual multiple‑choice questions before answering. Results show that the models’ confidence signals split into two distinct patterns: early in training, confidence aligns with output consistency (how concentrated the answer distribution is), while later, it aligns with true accuracy but only on data similar to the training set. This indicates that calibration training may not universally teach LLMs to detect their own errors.