Model releases, architecture work and prompting research on large language models — from frontier-lab announcements to the arXiv papers behind them.
arXiv:2609.26512v1 Announce Type: new Abstract: Convolutional neural networks (CNNs) and vision transformers are both used to model the human visual system, but whether the two architectures diverge...
arXiv:2609.25627v1 Announce Type: cross Abstract: General-purpose robot control requires models to understand task intent, identify where to interact, capture how the scene evolves, and generate prec...
arXiv:2607.04745v2 Announce Type: replace-cross Abstract: Strong mapped-region thermal visual place recognition (VPR) does not ensure safe rejection of unmapped queries. We identify and quantify this...
arXiv:2609.25575v1 Announce Type: cross Abstract: Fine-tuned Large Language Models (LLMs) significantly advance Automated Theorem Proving (ATP), but are often deployed as guiding policies within tree...
arXiv:2609.26652v1 Announce Type: cross Abstract: Commonly, classifiers and monitoring procedures are trained from labeled data by optimizing an objective such as the misclassification rate. This may...
The paper discusses the trustworthiness of agentic AI systems built on large language models, highlighting new security and operational risks such as indirect prompt injection, memory contamination, and cross‑session data leakage. It categorizes failure modes, reviews mitigation strategies—including instruction hierarchies, context isolation, and constrained tool use—and introduces the Trustworthy Agent Development Lifecycle (TADL), a six‑phase framework for specification, design, training, evaluation, deployment, and monitoring. The authors note that TADL has not yet been empirically validated but offers a structured foundation for developing more secure and accountable agentic systems, and they call for improved benchmarks and future research priorities.
EDGEGEN is a synthetic task generation framework that extracts compliance rules from a tool‑calling agent’s specification to create database‑grounded edge‑case tasks that violate those rules. By combining EdgeGen with existing synthetic data generation methods, it forms a fully automated closed‑loop system that requires no human annotation. Experiments show that finetuning on EdgeGen data improves performance by 2–42 % on the tau2bench airline domain, while harness optimization yields 10–30 % gains over human‑curated and base harnesses for the Gemma‑4‑e4b model.
LADDER is a new framework that combines diffusion language modeling with Graph Retrieval-Augmented Generation (GraphRAG) to enable efficient multi‑hop reasoning. It introduces an event‑driven self‑clocking retrieval mechanism that triggers graph queries only when new entities appear, and an incomplete‑query graph propagation module that aggregates multi‑hop evidence during parallel decoding. Experiments on three multi‑hop QA benchmarks show that LADDER improves exact match from 39.6% to 45.2% while reducing latency by 4.1×.
DUMA-Bench is a new benchmark that evaluates the security of large language model agents in dual‑control settings, where both the agent and the user can modify the shared environment. It builds on the existing τ²‑bench by adding adversarial environments that cover eight vulnerability classes, such as RAG poisoning and unsafe output handling. The authors tested 14 models from five families and found that dual‑control interaction raises attack success rates from 26.9% to 41.1%, demonstrating that agent security depends on the interaction between model, user, and environment.
ActGov is a runtime enforcement framework that validates each action proposed by a large language model (LLM) agent before it interacts with external tools, ensuring that actions stay within task‑scoped authorization boundaries and comply with dynamically constructed policies. It builds policies from tool specifications, benign tasks, and failure traces, verifying updates via SMT‑based counterexample checking. In evaluations on AgentDojo and AgentDyn benchmarks, ActGov consistently reduces indirect prompt‑injection attack success while maintaining task utility, outperforming existing defenses.
The paper introduces a physics‑guided deep metric learning approach for open‑world radar pulse de‑interleaving, leveraging continuous Time‑of‑Arrival sinusoidal positional encodings to model physical inter‑pulse durations. It builds on a transformer‑based framework, optimizing network parameters solely with Supervised Contrastive learning and employing physics‑based priors—PRI consistency and AoA continuity—for validation and checkpoint selection via unsupervised HDBSCAN clustering. This method aims to improve de‑interleaving performance in dense, contested electromagnetic environments where classical techniques falter.
The paper proposes a method for allocating a limited budget of expert annotations to optimize the accuracy of off-policy evaluation in settings where rewards are missing or noisy. By deriving variance‑optimal annotation probabilities for sequential, forward‑monotone protocols, the authors provide a batch‑adaptive implementation that can be applied to real data. Experiments on casenotes from a homelessness services nonprofit and on human‑preference votes from LMArena demonstrate substantial reductions in RMSE—up to 65% for housing placement and 68% for progress toward a housing application—when using only 40% or more of the full annotation budget.
Agent0 is a fully autonomous framework that enables large language model agents to evolve without external data by using a multi‑step co‑evolution process. It pits a curriculum agent against an executor agent, both derived from the same base LLM, where the curriculum agent creates increasingly challenging tasks and the executor learns to solve them. By integrating external tools into the executor’s workflow, the system creates a self‑reinforcing cycle that continuously generates high‑quality curricula, leading to significant gains in reasoning performance—an 18% improvement on mathematical reasoning and 24% on general reasoning for the Qwen3‑8B‑Base model.
The paper addresses the problem of over‑refusal in safety‑aligned large language models, where benign instructions are incorrectly rejected. It identifies that a small set of hypersensitive safety heads in transformer attention misfire on hard‑safe prompts, causing abnormal attention entanglement and high‑entropy routing conflicts that block necessary attention to target entities. To mitigate this, the authors propose Semantic Routing Calibration (SRC), a lightweight, training‑free inference framework that dynamically suppresses these hypersensitive heads and fuses logits from dual branches to restore trustworthy reasoning while preserving intrinsic safety performance.
The ICDAR2026 Competition on Multimodal Reasoning over Documents in Multiple Domains introduced a new Visual Question Answering benchmark that tests reasoning over documents from eight distinct domains such as business reports, scientific papers, and engineering drawings. Twenty valid submissions from eight teams were evaluated, featuring approaches ranging from zero‑shot vision‑language models to multi‑agent ensembles and fine‑tuned multimodal systems. Results indicate that the most effective systems employ structured evidence extraction, retrieval, verification, and orchestration across multiple components rather than single‑pass prompting.
The paper introduces ufakzeka-1, a 151‑million‑parameter Turkish language model trained from scratch on 13.5 B tokens. It details the tokenizer, a three‑stage pretraining schedule, post‑training data augmentation, and a comprehensive evaluation suite that includes release gates, a rule‑checked conversation sweep, and hand tests—all with prompts excluded from training data. The authors report three key findings about safety gate performance, training‑seed variance, and data‑round effects, and they release the model weights, data recipe, evaluation code, and spend ledger under Apache‑2.0.
The paper introduces modality‑gated deep adapters, a parameter‑efficient method for adding new modalities to a frozen multimodal embedding language model without altering its existing outputs. These adapters are bottleneck modules attached to each decoder layer, grouped into modality‑specific packs that activate only during encoding of their own modality, ensuring exact preservation of the base model’s computation graph. Experiments on a 2B base model show significant gains in audio‑to‑text and thermal‑to‑text retrieval metrics, and the authors release the audio and thermal packs along with training and evaluation code.
The study evaluates whether large language models (LLMs) used as synthetic personas can predict real audience responses to marketing copy. Using thousands of headline A/B tests from the Upworthy Research Archive, the authors compare a ten-persona panel grounded in real audience demographics to a no-persona zero‑shot baseline that asks the model for a typical reader’s click likelihood. Results show that the no‑persona baseline outperforms the persona‑based approach, with higher predictive validity and top‑1 accuracy, indicating that forcing the model to role‑play specific personas introduces bias and noise.
The paper introduces Dynamic Tool Output Compression (DTOC), a framework that manages context in large‑language‑model agents by storing full tool outputs in external memory and inserting compact placeholders into the active context. DTOC treats context updates as explicit, reversible operations within the agent’s reasoning loop, allowing selective reconstruction of compressed outputs when needed. Experiments on the DeepSWE benchmark show that for responsive models such as Sonnet 4.6 and GPT‑5.4, DTOC reduces input tokens and agent steps while significantly improving solve rates and lowering cost per solved task, with ablation studies confirming the importance of reversibility for maintaining performance.
The paper argues that observing only behavior is insufficient to identify social norms in large language model (LLM) societies. It introduces an evaluation framework that also measures agents’ reported empirical and normative expectations, revealing that expectation elicitation boosts cooperation, social learning stabilizes behavior, and social selection identifies cooperators but offers limited reinforcement. The study shows that similar cooperative outcomes can stem from distinct underlying mechanisms and that expectations can be used to attribute each mechanism’s contribution.