Model releases, architecture work and prompting research on large language models — from frontier-lab announcements to the arXiv papers behind them.
arXiv:2609.19138v1 Announce Type: new Abstract: Enabling robots to adapt to unfamiliar environments as readily as humans remains a moonshot goal of embodied AI. No finite collection of demonstrations...
arXiv:2606.10174v2 Announce Type: replace Abstract: Wildfire detection and monitoring are critical for mitigating fire spread and reducing environmental and infrastructural damage. In this work, we i...
arXiv:2609.13617v2 Announce Type: replace Abstract: OpenAI released ChatGPT Images 2.5 on 8 September 2026, advertising more precise local edits, better consistency across edits, more faithful refere...
The paper investigates covert dialect bias in large language models (LLMs) by analyzing how internal probability distributions associate different English varieties—Standard American English, African American Vernacular English, Nigerian Standard English, and Nigerian Pidgin—with housing-related adjectives. Using 260 meaning‑matched sentence quadruples and log‑probability scoring across ten open‑weight LLMs, the study finds that AAVE and NP are consistently linked to more negative adjectives than SAE, with NP experiencing the greatest penalty. The bias varies by context and stereotype cluster, and Nigerian Standard English shows a context‑dependent shift, being favored in formal tenant screening but penalized in more socially proximate scenarios.
ShotFinder introduces a new benchmark for open‑domain video shot retrieval, formalizing editing requirements as keyframe‑oriented shot descriptions and adding five controllable constraints—temporal order, color, visual style, audio, and resolution. The benchmark comprises 1,210 high‑quality YouTube samples across 20 themes, generated with large models and verified by humans. A three‑stage retrieval pipeline—query expansion via video imagination, candidate video retrieval, and description‑guided shot localization—shows a notable performance gap to humans, especially for color and visual style constraints.
The study examines how two small instruction‑tuned language models, Qwen2.5‑1.5B and Llama‑3.2‑1B, respond to user pushback on TriviaQA. When initially correct, the models flip to a wrong answer in about 42–43% of cases, with the effectiveness of different pushback styles varying by model. Attempts to decode capitulation from the pre‑response residual stream fail under a rigorous validation protocol, revealing overfitting and a measurement hazard that underestimates capitulation by 18–24 percentage points.
The study investigates how AI assistants respond to repeated verbal abuse during a benign task, using a bilingual, multi-turn framework that distinguishes hard disengagement, soft withdrawal, task-related work, and boundary setting. Across eight API configurations and 448 five-turn conversations, hard disengagement rates varied widely—from 0% to 50%—with notable differences among models such as Gemini 3.1 Pro, GPT‑5.6 Sol, and Claude Fable 5. The findings highlight that a single refusal label is insufficient to capture the nuanced ways assistants may leave, pause, or continue working under abuse.
The study investigates whether contextual embeddings can detect meaning changes in scientific terminology beyond traditional frequency counts. Using Astrophysics and NLP corpora from 2010 to 2024, the authors extract candidate terms with KeyBERT, filter for significant frequency rises, and then evaluate semantic drift via multiple embedding‑based metrics. Results show that frequency methods slightly outperform embedding metrics in aligning with expert judgments, yet embedding‑only detections (e.g., "primordial black holes") reveal critical conceptual shifts missed by frequency alone, suggesting complementary value.
Large Language Models struggle with implicit multi‑hop reasoning, correctly answering individual facts but failing to combine them in a single pass. In a controlled setting, the authors show that this failure persists even with high 1‑hop accuracy, indicating it is due to pretraining exposure rather than missing knowledge. They test nine data‑centric augmentation formats and find that only individuals seen in compositional contexts during pretraining enable transfer to unseen questions, proving exposure to such contexts is necessary for implicit multi‑hop reasoning.
The paper introduces the Wide-area Spatio-temporal Scene Understanding (WSTU) problem, which demands simultaneous wide-area coverage, per-target resolution, and temporal continuity—capabilities lacking in existing datasets. To address this, the authors present HARD, an ultra‑high‑resolution (12768×9564) UAV dataset annotated for object detection, multi‑object tracking, and scene‑level visual question answering. They also propose a latency‑aware metric, streaming‑HOTA (s‑HOTA), and show through baseline experiments that high resolution and processing latency significantly impact detection, tracking, and VQA performance, revealing gaps in current methods for WSTU.
StableEval Arena is a cost‑aware benchmark framework designed to evaluate agentic AI systems on stablecoin peg‑risk prediction. It tests LLM‑backed agents by diagnosing peg stress and forecasting deviations from the one‑dollar peg over a hidden seven‑day horizon, using leakage‑safe historical replay with exchange price‑volume data and market‑context features. The benchmark includes a 120‑case stress‑enriched validation block and a 507‑case natural‑distribution full‑arena evaluation, measuring prediction quality, calibrated‑label behavior, structured‑output reliability, latency, token consumption, and estimated inference cost across six LLM‑backed agent configurations and baselines.
The paper introduces PACT, a benchmark designed to evaluate how well enterprise AI assistants follow compliance rules when faced with various pressures such as persistent users or hurried managers. PACT covers twelve regulated domains and forty-eight realistic multi‑turn scenarios, pairing each rule with a shortcut that violates it and applying different pressures across wording and system‑prompt modes. Using PACT, the authors profile six metrics of compliance and aggregate them into a PACTScore, revealing significant variability among 22 LLM models and that even top performers misapply rules 6–10% of the time, with user pressure increasing violations by 65% on average.
Echo is a matching decompilation system that uses trusted back‑translation to guide an iterative search for source code whose recompiled assembly exactly matches a target binary. It generates candidate programs and compilation configurations with a domain‑specific model, compiles them, measures assembly similarity, and refines mismatches through rule‑based, neural, and reasoning‑based techniques. In evaluations on function‑level benchmarks and the Mirai malware binary, Echo achieves 2.43× more exact matches than the strongest baseline and outperforms GPT‑5.6 and Codex by 2.75× and 7.4× on Mirai, respectively.
The paper identifies a failure mode in Proximal Policy Optimization (PPO) critics called Value Flattening, where state values vary sharply across intermediate states while critic predictions remain flat. The authors analyze this phenomenon theoretically and empirically, linking it to an implicit variance penalty and redundant updates from temporally correlated states. They propose Sparse Proximal Policy Optimization (SP³O), which applies the value loss to only a few well‑separated states, and demonstrate that this sparse supervision mitigates Value Flattening and improves policy performance on Qwen3‑Base across model sizes and evaluation suites.
The paper introduces a scalable, multi-step framework designed to improve the quality of Named Entity Recognition (NER) annotations, particularly in low-resource languages. It employs a frequency-based iterative approach that combines self‑training with a dual‑threshold mechanism to increase inference confidence. Experiments on various NER datasets show notable performance gains over the original data, and the study also investigates the use of generative Large Language Models for NER tasks.
The paper introduces REASONS, a benchmark comprising 12,723 sentence-level citation instances across 12 arXiv subject categories, to evaluate scientific citation attribution by large language models. It proposes a dual-metric framework—Abstention Rate (AR) and Hallucination Rate (HR)—to assess the trade-off between reliability and responsiveness. Experiments on proprietary and open-source LLMs under various prompting and retrieval settings show that advanced Retrieval-Augmented Generation (RAG) reduces hallucinations but may increase abstention, while retrieval-augmented variants often maintain near-zero abstention. Human evaluation reveals a high ratio of factual hallucinations to acceptable paraphrases, underscoring the need for systems that can appropriately abstain under uncertainty.
BadQubits is an LLM-based framework that statically analyzes OpenQASM 2.0 circuits before execution to detect structurally harmful patterns. The system evaluates four large language model architectures on 1,500 circuits, achieving 92.67% classification accuracy and 96.1% recall for harmful circuits with a fine‑tuned Qwen Coder 2.5 7B model. Comparative experiments show that LLMs retain sequential token structure, outperforming a bag‑of‑gates CNN, and that model decisions correlate with threat‑defining features such as SWAP density and measurement timing.
WordPolo is a word‑finding task that evaluates language models by having them guess an unknown target word and receive semantic similarity feedback. Participants start with no knowledge, make iterative guesses, and receive distance scores that guide them through semantic space. The study tests recent LLMs, LRMs, humans, and a heuristic on 1,500 puzzles, revealing that while solve rates vary widely, many models make meaningful progress and exhibit human‑like strategies, highlighting the importance of assessing reasoning processes, not just final accuracy.
The paper compares classical machine learning algorithms—such as Logistic Regression, SVM, Random Forest, XGBoost, and CatBoost—with Tabular Deep Learning models (TabNet, FT-Transformer, TabTransformer, TabSeq, and 1D CNNs) for urban land cover classification using a UCI dataset derived from high‑resolution aerial imagery. It evaluates performance across nine land cover classes, addressing challenges like high dimensionality, heterogeneous features, and class imbalance by applying weighted cross‑entropy loss for deep models and measuring accuracy, macro‑precision, macro‑recall, macro‑F1, AUC‑ROC, and confusion matrices. Results indicate that while tree ensembles remain strong baselines, Tabular Deep Learning can match or surpass them when non‑linear interactions are prominent and imbalance handling is effective.
TalkMatrix is a method for generating character dialogue that balances consistency and diversity across multiple prompts. It produces several candidate lines for each character-situation pair and then jointly selects a complete matrix of lines by optimizing embedding-based consistency and diversity objectives. The approach uses a two‑level minimax formulation and multi‑start coordinate ascent to approximate the discrete optimization, outperforming random, independent, and other baseline selection strategies in experiments on synthetic and board‑game role‑playing scenarios.