Model releases, architecture work and prompting research on large language models — from frontier-lab announcements to the arXiv papers behind them.
arXiv:2609.13288v1 Announce Type: new Abstract: Video-language models can answer multiple-choice questions with high confidence yet be wrong. We study whether answer-level reliability scores can be i...
arXiv:2609.13617v1 Announce Type: new Abstract: We evaluate whether the improvements advertised for ChatGPT Images 2.5 translate into better performance on forgery tasks with predetermined answers. W...
arXiv:2609.13815v1 Announce Type: new Abstract: Text-centric Visual Question Answering (VQA) requires reading and reasoning over text embedded in images, a task made substantially harder when images...
arXiv:2609.14080v1 Announce Type: new Abstract: Promptable models such as microSAM segment electron microscopy (EM) images from point prompts, but automation requires generating prompts without user...
arXiv:2609.14521v1 Announce Type: new Abstract: Recovering editable 3D parametric curves from 2D images is a fundamental challenge in computer graphics, bridging pixel-based perception and vector-bas...
arXiv:2609.15457v1 Announce Type: new Abstract: Vision-Language Models (VLMs) enable autonomous GUI navigation, but agents still struggle to process and learn from dense, continuous visual histories....
arXiv:2609.13606v1 Announce Type: cross Abstract: Multi-arm robotic harvesting offers a promising path to improve harvesting efficiency and reduce reliance on manual labor. However, practical deploym...
The paper investigates how large language models perform propositional logical reasoning by conducting a causal mechanistic analysis on the PropLogic-MI benchmark. It identifies four interlocking mechanisms—Staged Computation, Information Transmission, Fact Retrospection, and Specialized Attention Heads—that organize the reasoning process across layers. The study demonstrates that these mechanisms recur across different model families, rule categories, and reasoning hops, indicating a structured, layer‑organized internal process for propositional reasoning.
arXiv:2605.18177v2 Announce Type: replace Abstract: Query-based Vision Transformer segmentation models typically reconstruct dense spatial feature maps to predict masks, inheriting design patterns fr...
The paper investigates how to evaluate generative audio large language models (Audio‑LLMs) on known closed‑set tasks by separating the decision to call a generative model from the use of acoustic evidence. It introduces a controlled call‑decision framework where a policy can choose between a transcript label, encoder evidence from CLAP, AST, or WavLM, or a generative call to Qwen2‑Audio, Qwen2.5‑Omni, or MOSS‑Audio, and measures the impact of generative calls on accuracy. Results on the VocalSound dataset show that while transcript‑only accuracy is low (0.296), encoder‑based controls achieve high accuracy (≈0.85) without any generative calls, and adding generative calls yields only a marginal improvement (0.925 vs. 0.921).
arXiv:2606.08918v2 Announce Type: replace Abstract: Worldwide image geo-localization aims to determine where on Earth a single image was captured. However, visually similar scenes may lie thousands o...
arXiv:2606.23843v2 Announce Type: replace Abstract: Vision-language models (VLMs) achieve strong cross-modal alignment but remain brittle to negation, often relying on shallow word associations rathe...
The paper investigates how vision‑language models (VLMs) read exact values from vertical bar charts. Using controlled counterfactual activation patching on Qwen2.5VL‑7B‑Instruct and InternVL3.5‑8B, the authors find that the bar‑top region contributes more to answer accuracy than the bar body, and that legend and series states lose recoverability earlier than geometry and axis‑scale states. They also show that prompt‑series positions mediate legend influence and that the two models differ in how they combine geometry and scale information from separate image donors.
The paper introduces the Romantic Relationship Advice-Seeking Prompts (RRASP) dataset, comprising 2,400 prompts across five relationship themes, to study how query formulation affects sycophancy in large language models. Using the ELEPHANT framework, the authors evaluated GPT‑5 Mini and Gemini 3 Flash, finding that grammatical mood alone does not drive sycophantic behavior, whereas perspective‑driven framing does, with models increasingly accepting user premises over successive turns. Gemini 3 Flash showed smaller increases in moral sycophancy than GPT‑5 Mini, indicating greater resistance to reinforcing ethically problematic positions.
The paper presents an empirical study of factual errors in human-written text, focusing on corrections in newspaper articles to build a taxonomy of common mistakes such as kanji misconversions and unit errors. It evaluates large language models’ ability to detect these errors, finding that even advanced models like GPT‑5.4 achieve only a 52% word‑level F1 score on synthetic data, underscoring the difficulty of the task. The work highlights the gap in research on factual error detection in human writing compared to LLM hallucinations.
The paper introduces LLM-Microscope, a toolkit for measuring how Large Language Models encode contextual information at the token level. It shows that seemingly minor tokens—such as determiners, stopwords, and punctuation—carry surprisingly high contextual weight, and removing them degrades performance on benchmarks like MMLU and BABILong-4k. The study also finds a strong link between contextualization and linearity, indicating that the transformation between layers can be approximated by a single linear mapping when tokens are well contextualized.
The paper proposes Bypass Observation, a non‑intrusive layer‑wise readout architecture that attaches read‑only observation heads to selected Transformer layers without feeding their outputs back into the backbone. Three variants are explored: a shared language‑model head across layers, layer‑specific heads, and a layer‑ or step‑adaptive head. The authors provide a closed‑form overhead estimate (≈ V/(12d)) and discuss ways to reduce cost, while distinguishing bypass chain‑of‑thought from conventional chain‑of‑thought and outlining potential applications to looped and recurrent‑depth Transformers.
The paper reports experiments with six frontier AI models from OpenAI, Anthropic, xAI, and Google DeepMind, using ten independent sessions per model type and a three‑stage prompt sequence that shifts from architectural preference to a full ASCII backbone. Under a school‑audience framing, responses consistently converged on a shared architectural pattern featuring persistent latent state, adaptive computation, memory, specialist routing, verification, stopping control, and delayed decoding, while removing the framing led to more heterogeneous results. A notable observation is the close overlap between GPT‑5.6 Sol’s elaborate successor architecture and GPT‑6 Astra’s independently sketched design, raising questions about shared design priors or independent convergence.
CoMem introduces a memory architecture for multi-agent systems that blends private experience with shared knowledge. It includes Private Experience Sedimentation to retain useful individual memories, Collective Wisdom Curation to filter widely proven ideas for sharing, and Parallel Dual-Stream Retrieval to draw from both personal and group memories while maintaining diversity. Experiments on ALFWorld and PDDL benchmarks demonstrate that CoMem improves overall performance and reduces memory pollution.
The paper "Towards Optimizing SQL Generation via LLM Routing" proposes a routing approach for Text-to-SQL tasks that dynamically selects the most cost‑effective large language model (LLM) for each query. Two routing strategies—score‑based and classification‑based—are introduced, achieving accuracy comparable to the best LLM while reducing latency and monetary cost. The authors design the routers for easy training and efficient inference, and demonstrate a practical accuracy‑cost trade‑off on the BIRD dataset.