Model releases, architecture work and prompting research on large language models — from frontier-lab announcements to the arXiv papers behind them.
arXiv:2609.16055v1 Announce Type: cross Abstract: Test-time compute has emerged as a major approach to improving the capabilities of Large Language Models (LLMs). However, existing test-time reasonin...
BiCLIP is a bidirectional multimodal framework that enhances medical image segmentation by allowing visual features to iteratively refine textual representations, improving semantic alignment. It incorporates an augmentation consistency objective to stabilize learning against perturbed inputs. Experiments on QaTa-COV19 and MosMedData+ show that BiCLIP outperforms state‑of‑the‑art image‑only and multimodal baselines, achieving strong performance even with only 1% labeled data and resisting common clinical artifacts such as motion blur and low‑dose CT noise.
arXiv:2609.16937v1 Announce Type: cross Abstract: On-policy distillation (OPD) has emerged as an effective approach for large language model post-training, yet existing objectives face a trade-off be...
The paper reviews ethical and privacy risks of large language model (LLM)–enabled geospatial artificial intelligence (GeoAI), identifying eight recurring issues such as data provenance, spatial privacy, algorithmic bias, and technical risks. It evaluates current responses, noting many remain largely unaddressed or conceptual, and proposes a governance‑aware architecture with enforceable controls illustrated by a flood‑response routing scenario. The authors call for empirical validation, spatially specific interpretability tools, and workforce training to address these emerging risks.
The paper titled "Position: AI Is Not Ready for Strategic Conflicts" argues that language‑model (LM) based open‑ended strategic wargames, while useful for simulating adversaries, institutions, and crisis response, pose significant safety risks. It identifies five failure modes—decision laundering, adjudication opacity, role collapse, escalation‑through‑adjudication, and failure of strategic imagination—and contends that such wargames should not inform real‑world planning or policy without an auditable safety case. Instead, the authors suggest using these simulations primarily as stress tests to expose potential failures in decision‑influencing LM agents.
The paper introduces CLEAR, an agentic framework designed to improve the reliability of large language models (LLMs) in medical contexts by adjudicating evidence from multiple sources. CLEAR generates candidate answers from three distinct pathways—parametric knowledge, locally curated corpora, and dynamically retrieved evidence—and then uses an aggregation verifier to evaluate agreement and conflict among these sources. An adjudication module decides whether to preserve or revise conclusions, employing override-guard and challenge-audit mechanisms, and initiates targeted follow-up searches when conflicts remain unresolved.
The paper investigates whether natural language inference (NLI) can be performed using only interpretable, graph-based representations of evidence. It introduces a pipeline that decomposes sentences into atomic propositions, maps them to ConceptNet triples, and feeds three graphs—premise, hypothesis, and a retrieved ConceptNet subgraph—into a fine‑tuned 0.8‑billion‑parameter language model. On the SNLI dataset the graph‑only model reaches 89.7% accuracy, close to a text‑based baseline, while on ANLI it matches RoBERTa‑large on rounds R2 and R3 but lags on R1, illustrating a trade‑off between interpretability and performance.
The paper introduces MASA, a method for Wild Test-Time Adaptation that uses a frozen multimodal large language model to provide structured semantic anchors, thereby avoiding the self-referential loop common in existing WTTA techniques. MASA selects a small, diverse set of reliability-ranked anchors, encodes their descriptions, propagates them to nearby test samples, and stores this visual‑semantic information in an online prototype memory. The stored descriptors enable lightweight adaptation of normalization parameters, and MASA is evaluated on the WTTA ImageNet‑C benchmark with ResNet and ViT backbones under limited‑batch, mixed‑domain, and imbalanced‑label‑shift scenarios.
The paper introduces a framework that uses Large Language Models (LLMs) to generate synthetic time‑series data for manufacturing processes. By fine‑tuning pre‑trained LLMs on manufacturing instructions and applying Retrieval Augmented Generation (RAG), the method enhances data diversity and realism. Evaluation against traditional models such as ARIMA and LSTMs shows that the LLM‑driven approach produces higher‑quality synthetic data, better capturing temporal dependencies and improving downstream anomaly detection performance.
The paper proposes universal, tool‑based defenses for large language model agents that use external tools, addressing four types of adversarial attacks: direct and indirect prompt injection, memory poisoning, and backdoor attacks. Two main defenses are introduced: Attacker Tool Filtering, which uses anomaly detection to remove suspicious tools, and Normal Tool Recalling, which restores the agent’s original toolset before planning. The authors also add prompt‑based defenses such as Chain‑of‑Thought prompting and self‑reflection, and demonstrate that these methods dramatically lower attack success rates—often to 0%—across multiple open‑source and proprietary LLMs while maintaining or improving task performance.
The paper evaluates five context‑trimming strategies for agentic large language model workflows, comparing them on metrics such as task success, protocol adherence, token savings, and latency. Conventional trimming methods save about 60% of tokens but achieve lower success rates, while protocol‑aware trimming raises success to 92.2% and adaptive guardrails further improve it to 96% success with 56% token savings. The study shows that preserving protocol‑critical state is more important than aggressive token removal, and that adaptive guardrails enhance efficiency, scalability, and reliability for long‑horizon agentic systems.
The paper introduces VT-Transformer, a model that predicts answerability scores for Visual Question Answering by treating the task as a regression problem rather than a binary classification. It leverages visual and textual features within a Transformer architecture and demonstrates improved performance and robustness on the VizWiz 2020 dataset compared to existing baselines.
RoleBreak is an open benchmark designed to evaluate long‑horizon role‑playing robustness in spoken dialogue systems. It includes 310 character‑based and user‑centered roles, 6,688 human‑verified dialogue turns, and 11,743 fine‑grained evaluation criteria, with 1,856 turns specifically targeting expressive vocal emotion. The benchmark stresses role consistency, interaction quality, safety, and affect over extended conversations, and the authors evaluated nine system configurations across full‑duplex, omni‑modal, and cascaded ASR–LLM–TTS paradigms.
The article surveys how large language models (LLMs) can be incorporated into networked control systems, cyber‑physical systems, and multi‑agent networks without violating stability and safety guarantees. It proposes treating the LLM as a slow supervisor that sets high‑level goals, while a fast, certified inner loop preserves physical stability. The survey maps LLM characteristics—such as inference latency, API failures, tokenization, and hallucinations—to classical control challenges and highlights the growing gap between model capability and formal safety assurances, calling for future research on stability proofs.
TAME (Token Attribution and Masking for Emergent misalignment) is a three‑stage framework that identifies which training tokens drive harmful behavior in fine‑tuned language models. It first scores tokens by how much fine‑tuning increases their likelihood, then characterizes patterns among high‑attribution tokens, and finally validates them by masking during training. Experiments on Llama and Qwen show that masking the top‑attribution tokens reduces emergent misalignment by 23‑ to 36‑fold, while random masking has no effect.
The paper investigates how to help users monitor their own and an AI system’s competence when using AI assistance. It identifies 30 interventions from experts and organizes them into a design space based on timing, target competence, and source of cue. A large experiment shows that reliability cards and contrasting replies reduce estimation error and overconfidence, though they do not improve task performance.
The paper presents an AI-driven workflow for optimizing the thermal design of 2D CFET inverters, using an electrothermal model to evaluate structural proposals against electrical costs. At 12 nm, the AI selects a redistributed source‑interconnect geometry and a substrate‑directed heat‑removal path, achieving a 1.67 K reduction in peak temperature rise for a fixed metal volume and 20 µW. Subsequent sensitivity analyses show additional cooling benefits and trade‑offs with nFET on‑current, demonstrating that the AI scientist workflow can propose viable thermal structures and quantify their electrical impact.
The paper investigates how to choose model pools for Multi-Agent Systems (MAS) that combine multiple model outputs to tackle complex reasoning tasks. It evaluates eight selection strategies—such as model size, accuracy, and answer diversity—across both before-generation (routing) and after-generation (majority-voting, LLM-as-a-judge) MAS architectures on scientific benchmarks. The study finds that expanding the candidate pool often harms performance, that selecting candidates within a single model family yields the best relative gains, and that indiscriminate addition of heterogeneous models can destabilize the system.
MiCRo is a two‑stage framework that improves personalized preference learning for large language models. It first uses a context‑aware mixture model to capture diverse human preferences from large binary preference datasets, then applies an online routing strategy to dynamically adjust mixture weights based on context, reducing ambiguity. Experiments on multiple datasets show that MiCRo captures diverse preferences and enhances downstream personalization.
The paper introduces ANCHOR, an external supervisory framework driven by large language models (LLMs) that provides evaluative feedback at multiple stages of self‑evolving agents. By integrating ANCHOR into two open‑source self‑evolving agent frameworks, the authors demonstrate that it significantly improves safety performance while preserving core capabilities across coding, mathematical reasoning, and safety tasks. The study also finds that supervision based on execution results is especially effective and that increasing supervision frequency yields diminishing returns, offering practical guidance for future research.