Model releases, architecture work and prompting research on large language models — from frontier-lab announcements to the arXiv papers behind them.
arXiv:2609. 31176v1 Announce Type: new Abstract: Repository-level issue localization aims to identify and rank the files and functions relevant to resolving a reported issue.
The paper introduces Stackelberg Program Optimization (SPO), a framework that uses large language models to discover adaptive destroy‑repair operators for large neighborhood search. SPO conditions operator decisions on a compact state representation, enabling state‑dependent behavior, and frames the discovery process as a Stackelberg game where destroy operators act as leaders and repair operators as conditional followers. Experiments on the traveling salesperson and capacitated vehicle routing problems show that SPO outperforms strong baselines, generalizes to larger instances, and exhibits coupled improvement in operator behavior during discovery.
The paper argues that the current LLM-as-a-judge evaluation method, which compensates for systematic measurement bias by increasing the number of comparisons, is statistically unsound and computationally wasteful. It identifies that treating LLM judges as neutral ignores documented biases such as position bias, verbosity bias, judge severity, and self‑enhancement. The authors propose a unified latent variable framework that jointly models pairwise and ordinal data while explicitly correcting for these confounders, enabling reliable rankings with far fewer comparisons and negligible additional compute.
The paper investigates how the definition of influence—specifically the behavior being attributed, the intervention on training data, and the counterfactual training process—affects rankings produced by influence estimators. It formalizes influence as a counterfactual estimand, distinguishes specification mismatch from approximation error, and categorizes existing estimators by their implied specifications. Experiments demonstrate that different specifications can lead to markedly different rankings, and that careful specification choice improves attribution quality in tasks such as noisy label detection and large‑language‑model attribution.
The paper introduces DIAL, a framework that uses large language models (LLMs) as judges while mitigating position bias and aligning their judgments with human preferences. DIAL separates judge‑specific position effects, learns shared structure in debiased LLM preferences, and adaptively calibrates this structure toward human targets using limited human comparisons. Experiments on simulations and three human‑preference benchmarks show that DIAL remains robust to unbalanced response order, achieves strong human‑aligned rankings with few labels, and adapts when LLM information is imperfect, supported by a real‑data study of over 410K judgments from 21 LLM judges.
The study investigates how the choice between processing page images or parsed text in an information extraction pipeline depends on the document’s layout, focusing on privacy‑sensitive, on‑premise scenarios with small models (≤8 B parameters). It evaluates accuracy and energy consumption across input representations, model families, and inference settings, finding that batching dramatically reduces energy use, FP8 quantization offers modest savings, and neural OCR is far more energy‑intensive than classical OCR. The optimal representation varies: vision‑language models excel on layout‑rich documents, while small text‑only models with a cheap parser perform best on near‑plain‑text contracts, achieving higher accuracy and lower energy than any vision‑language setup.
The paper examines a method called Program-of-Layers (PoLar) that allows transformer layers to be dynamically routed rather than processed in a fixed sequence, mirroring the brain’s thalamic routing. Reproductions across five models confirm that skipping, repeating, and combining layer blocks improve performance, with shorter programs for easier inputs and more repeats for harder ones. However, the study could not replicate the claimed advantage of a learned single‑shot router, noting that its top prediction defaults to the standard pass while the top‑k predictions still yield accuracy gains. The authors also analyze the robustness of correction programs, finding them brittle to single edits, and release their code publicly.
The paper introduces SeLATM, a framework that improves topic modeling by generating topics at the segment level and refining them through agentic feedback loops. This approach addresses limitations of LLM-based topic assignment methods, such as the inability to produce topic distributions, overly broad or narrow topics, and high resource consumption. Experiments on multiple datasets show that SeLATM reduces LLM resource usage while maintaining superior performance.
Game Arena is an open, continuously expanding platform that evaluates large language models through competitive games, allowing head‑to‑head matchups in structured environments. Unlike static benchmarks, it prevents performance saturation by increasing gameplay difficulty as models improve. The report outlines the infrastructure and presents three pilot games—Chess, Poker, and Werewolf—covering perfect information, imperfect information, and multiplayer settings, and details evaluation metrics and competition results.
The paper investigates prompt minimization, aiming to reduce prompts to their smallest, most information-dense form without losing output fidelity. It argues that shorter prompts lower computational overhead and inference latency, especially when large contexts are unnecessarily included, and that longer prompts can harm LLM reasoning and accuracy. The authors propose three frameworks to identify minimal prompts and show that these often produce outputs comparable to longer versions, highlighting redundancy in the input space and opening new avenues for efficient prompt engineering.
The paper introduces Steiner’s taxonomy of group tasks to study how multi‑agent large language model (LLM) teams scale on disjunctive versus compensatory tasks. By modeling agents as conditionally independent given the item, it shows that plurality voting converges to the modal answer while averaging converges to the item‑level bias. Experiments with 13 open‑weight models and up to 30 agents reveal that disjunctive tasks benefit from larger teams, whereas compensatory tasks like Fermi estimation see little improvement, highlighting that task structure and aggregation method fundamentally determine team scaling.
DeepEdu‑v1 is an AI‑tutoring system tailored for Vietnamese education that addresses data‑sovereignty and local curriculum alignment issues. It uses a long‑context inference engine to reduce retrieval calls and prefill latency by about 35%, and a self‑improving agentic layer that curates verified local knowledge without fine‑tuning. In deployment, DeepEdu achieves nearly twice the speed of standard vLLM serving and raises agentic accuracy from 70.0% to 79.5% on complex tasks, especially in financial reasoning and interactive‑agent benchmarks.
The paper demonstrates that fine‑tuning reasoning models to predict their own confidence at intermediate steps—using only 600 self‑supervised examples—substantially improves inference efficiency. Without adding any explicit stopping or length penalties, the models generate up to 25 % fewer tokens while maintaining accuracy on mathematical, scientific, and coding benchmarks across several architectures. The study finds that confidence supervision preserves the models’ high‑level reasoning structure rather than merely suppressing specific behaviors.
The article surveys fake review detection research, focusing on how pre‑trained language models (PLMs) and large language models (LLMs) influence both the generation of deceptive reviews and their detection. It reviews 211 studies from 2018 to early 2026, categorizing methods by evidence source—such as review text, sentiment, rating behavior, temporal metadata, user‑product graphs, multimodal content, external knowledge, and LLM‑generated signals—and by fusion level. The survey traces the evolution from traditional machine learning to PLM‑based and LLM‑based approaches, evaluates performance on Amazon, Yelp, and OpSpam benchmarks, and highlights open challenges including adversarial generation, cross‑domain transfer, uncertainty‑aware fusion, robustness to missing sources, interpretability, and trustworthy evaluation of AI‑generated deceptive content.
SignTrace is a system that tackles the reverse‑lookup problem in Chinese sign language by allowing users to describe a sign in natural language and retrieve the corresponding entry from a dictionary of 6,699 signs. It combines large‑language‑model–based dictionary enrichment, action extraction, seven‑channel retrieval, and candidate reranking, achieving a 94.0% Hit@1 and a mean reciprocal rank of 0.9540 on a benchmark of 500 queries. The system has been deployed for user trials, receives positive informal feedback, and processes median queries in 13.37 seconds.
The article evaluates the Sola Security Brain, a purpose-built security intelligence layer, against a general-purpose coding agent (Claude Code) on 28 cloud‑security investigation tasks. The Sola Security Brain achieved 0.693 coverage versus 0.387 for the coding agent, a 79.2% relative gain, and outperformed the agent on 25 of 28 tasks while incurring far lower reasoning and cost per unit of coverage. The study also identifies a ‘sample‑and‑generalise’ pattern where the live agent reports universal negatives based on limited sampling, illustrating a potential efficiency trade‑off in cloud investigations.
The paper "Strategic Self-Consistency" investigates how large language model providers might exploit the self‑consistency technique—generating multiple reasoning paths and selecting the majority answer—to overcharge users. The authors present a simple, efficient algorithm that strategically generates and reorders extra reasoning paths so that each appears necessary for the majority vote, thereby evading detection by auditors. Experiments on Llama, Qwen, and DeepSeek-R1 models across math, science, and QA benchmarks show that the added paths follow a heavy‑tailed distribution and that significant overcharging can persist even under stringent audits with a false‑positive rate below 0.1.
The paper introduces a new variant of the multi‑armed bandit problem that incorporates dueling feedback—pairwise comparisons of model responses—and heterogeneous sampling costs to identify the best large language model (LLM) from a set with varying query costs. Assuming a Condorcet winner, the authors propose a Track‑and‑Stop style algorithm that guarantees asymptotically optimal cost as the error probability approaches zero. Extensive experiments on synthetic and real‑world data show that this cost‑aware approach consistently outperforms both classical cost‑unaware algorithms and other cost‑aware extensions.
The paper introduces CLUE, a framework that lets robots actively resolve contextual uncertainty for underspecified natural language tasks. CLUE employs an LLM-derived policy to generate task-relevant hypotheses and plans, then uses an online language-embedded map to ground these into actions, refining its plan through closed-loop interaction. Experiments on a Boston Dynamics Spot across diverse indoor and outdoor settings show CLUE achieving near-oracle performance and outperforming LLM planners without closed-loop feedback by a significant margin.
The paper introduces ContextXR, a benchmarking framework designed to evaluate context‑aware interfaces in Extended Reality (XR) systems. It models an XR application as a graph of functional facets—semantically coherent groups of capabilities that support a shared user intent—and uses this representation to build MineXR++, a dataset enriched with facet‑level annotations. The framework defines three core tasks—context factor analysis, initial facet suggestion, and next facet suggestion—and evaluates suggestion methods via a simulated interaction metric that measures navigation and search cost to reach desired functionality.