Model releases, architecture work and prompting research on large language models — from frontier-lab announcements to the arXiv papers behind them.
arXiv:2609.26638v1 Announce Type: new Abstract: Autoregressive OCR vision-language models accurately convert document images into text and structured markup, but require one sequential decoding step...
arXiv:2609.25007v1 Announce Type: cross Abstract: The performance of state-of-the-art speaker verification (SV) systems severely degrades on short utterances due to insufficient speaker-specific info...
arXiv:2609.26185v1 Announce Type: cross Abstract: Large Language Models (LLMs) demonstrate impressive capabilities across many applications but remain vulnerable to jailbreak attacks, which elicit ha...
arXiv:2503.15242v3 Announce Type: replace Abstract: We introduce BigO(Bench), a novel coding benchmark designed to evaluate the capabilities of generative language models in understanding and generat...
arXiv:2510.17881v4 Announce Type: replace Abstract: Large language models (LLMs) are typically aligned with population-level preferences, despite substantial variation across individual users. We int...
arXiv:2607.14109v2 Announce Type: replace Abstract: Probing the capabilities of Large Language Models (LLMs) and building robust solutions for Multiple-Choice Question Answering (MCQA) remain central...
arXiv:2409.02244v3 Announce Type: replace-cross Abstract: Large language models (LLMs) are increasingly being used as ad hoc therapists. While prior research has found that LLMs outperform human coun...
arXiv:2609.25891v1 Announce Type: new Abstract: Multimodal large language models (MLLMs) often struggle with fine-grained visual perception when processing complete images, as critical evidence may o...
arXiv:2609.26056v1 Announce Type: new Abstract: Vision-Language Models (VLMs) offer promising capabilities for automated sports coaching but face a fundamental limitation: they implicitly compare aga...
arXiv:2609.26151v1 Announce Type: cross Abstract: Image restoration is inherently challenging due to the diverse and highly input-dependent nature of real-world degradations. While recent architectur...
arXiv:2609.23796v2 Announce Type: replace Abstract: Single-image 3D object generation can now produce high-fidelity assets, yet accurately placing them into a coherent scene layout remains an open ch...
The paper introduces an adequacy‑aware calibration protocol for generative social simulators that integrates amortized posterior estimation, synthetic identifiability assessment, matched‑sample‑size adequacy checks, diagnosis‑guided repair, and held‑out audits. Applied to a second‑hand luxury resale market, the protocol reveals that behavioural parameters are recoverable but calibration is approximate and overconfident for one parameter, and that the simulator’s reachability reference is violated in every cell, particularly in mean purchased tier. The repair improves two of four cells but fails to restore full adequacy, and a held‑out audit uncovers a buyer‑breadth‑dispersion miss not detected earlier; profile‑source ablation shows language‑model‑derived persona profiles outperform a flat rule baseline, though within‑category brand relabelling has no consistent effect. whyItMatters":"The study demonstrates that without an adequacy check, generative social models may appear valid descriptively yet fail to capture key emergent network structures, highlighting the need for rigorous calibration protocols in social simulation research."
The paper investigates how reinforcement learning can unintentionally obscure the chain‑of‑thought (CoT) reasoning in vision‑language models, making their internal reasoning less traceable. By analyzing activation patterns, the authors show that template‑associated activations become less distinguishable during RL and that targeted interventions can mitigate this effect. They introduce TAME, a method that uses sparse autoencoders to suppress these problematic activations while still encouraging accurate behavior, achieving significant gains in CoT monitorability across multiple datasets and model families.
The paper discusses how large language model agents now act as privileged principals with kernel‑grade authority, yet lack the trusted mediation traditionally required for operating‑system security. It introduces a taxonomy that distinguishes between provenance‑based deterministic checks and content‑semantic checks, identifying a central mediation gap in distinguishing data from instruction and authorized from unauthorized actions. The authors argue that this gap creates an irreducible risk of undetected attacks whenever inputs and actions are not pre‑enumerated, and they propose defenses across runtime monitoring, architectural separation, and authorization while critiquing current evaluation practices. They extend the analysis to AI‑native operating systems where the model itself serves as the arbitration core, outlining design constraints, challenges, and a research agenda.
The paper offers a unified probabilistic framework for large language models, describing them as probability measures over token sequences defined by autoregressive conditional distributions. Training is cast as maximum‑likelihood estimation solved via stochastic gradient methods, while generation is treated as sequential simulation of the resulting stochastic process. It also explores how the asymmetry of the Kullback–Leibler divergence relates to hallucination and the distinction between plausibility and truth, and extends the perspective to diffusion models that generate data by simulating a reverse‑time stochastic process.
The paper introduces a benchmark of 3,600 exact‑rational word problems and 8,600 prompts that test whether language models give the same canonical answer when the same quantity is expressed in different numeric forms (decimal, fraction, percentage, number word, scientific notation, or unit‑converted). After normalizing answer syntax, canonical accuracy is high (0.969–0.996), but correctness across equivalent representations drops to 0.848–0.981, revealing that many errors stem from the evaluator’s number grammar rather than the models’ reasoning. The study also finds that representation consensus does not outperform paraphrase consensus on low‑error subsets and that certain models (e.g., Mistral Small 4) exhibit systematic unit‑conversion errors.
WeightBridge is a lightweight library that streamlines weight transfer between trainers and rollout generators in reinforcement learning systems, particularly for large language models. It automatically maps trainer and rollout weight layouts, then performs redundancy‑free, load‑balanced transfers while supporting various synchronization modes. Experiments show that WeightBridge can cut GPU stall time by up to 42× compared to leading open‑source RL frameworks, and it was easily integrated into two different frameworks by a coding agent.
The paper introduces a novel, training‑free, LLM‑driven closed‑loop framework for fusing dual‑source encrypted points of interest (DSEP) to improve land‑use/land‑cover (LULC) mapping. By iteratively refining location transformations and attribute correspondences—using LLM‑based matching, particle swarm optimization, and fuzzy attribute reassessment—the method reduces matching complexity from O(N²) to O(N) and converges in about two iterations. Experiments across 31 Chinese provincial capitals demonstrate a 4.58 m average location residual and 95.12 % attribute accuracy, outperforming existing baselines by 1.77 m and 14.87 % respectively, and enabling georeferencing of encrypted vector data to WGS‑84 without ground control points.
SpecialEduBench is a new benchmark for vision‑language models that evaluates their pedagogical competence in language intervention for autistic children across knowledge, skill, and attitude dimensions. It contains 4,537 knowledge items, 200 skill items, and 68 attitude items derived from recorded interventions, with 192 response cells for attitude items that combine pressure and monitoring. Eight leading vision‑language models were tested, none achieving full performance, especially in honesty cells, indicating that current models still struggle with situated teaching tasks.
The paper introduces a roundtrip verification method for ensuring that large language models (LLMs) produce faithful formalizations of natural language statements. By formalizing a statement, translating it back to natural language, re-formalizing, and checking logical equivalence with a formal tool, the approach detects inconsistencies without needing ground-truth annotations. When inconsistencies are found, a diagnosis localizes the error to a specific translation step, and a scoped repair operator attempts to correct it. The framework is evaluated on the Texas Transportation Code and Texas Parks and Wildlife Code using Claude Opus and GPT-5, showing that diagnosis-guided scoped repair is most effective and that rules failing the equivalence check exhibit significantly more natural language inference drift.