As language model parameter sizes grow, effective compression is essential to reduce computational and memory overhead. Existing methods suffer from performance degradation at high compression ratios. This paper provides the first mathematical proof that low-rank decomposition and quantization are non-orthogonal—their combination causes significant performance loss. The authors propose Diagonal Adhesive Method (DAM) to effectively combine both techniques and mitigate the loss.

Knowledge editing can degrade math and programming reasoning. Moir estimates preservation covariance by sampling from the model's own decoding distribution, requiring no external data. It significantly improves performance preservation across multiple models and editing methods.

A position paper argues that despite advances in LLMs prompting claims that natural language could replace formal languages like programming languages, natural language is optimized for underspecification in open-ended contexts. The authors introduce a framework of 'task specificity', prove a 'specificity crossover theorem', and demonstrate through case studies that natural language excels at low-specificity tasks while formal languages are advantageous for stricter requirements. They conclude that the two are complementary and advocate hybrid systems.

AlphaAgent is a skill-driven framework that decouples retrieval-based QA from paper-level report generation using explicit skill contracts. It outperforms baselines in blind evaluations, particularly in mechanistic explanation and credibility awareness.

LLM-INSTRUCT, the winning system for the UZH Shared Task at ArgMining 2026, focuses on paragraph-level argument mining in UN and UNESCO resolutions. It uses metadata-aware dense retrieval to narrow candidate tags, constrained decoding with per-dimension caps, and a three-agent debate branch for uncertain cases. It achieved 1st overall, 1st in F1, and 5th in LLM-as-a-Judge. Development improved Task 1b Micro-F1 from 35.83% to 40.08%. The key lesson: reducing decision space before generation improves accuracy and submission robustness.

A new study using a factorial experiment challenges common assumptions about diversity in LLM opinions. Findings show that adding persona details does not monotonically increase diversity, different interaction architectures explore non-overlapping opinion regions, and combining architectures yields broader coverage than optimizing any single one. Low-cost alternatives like higher temperature and diversity instructions have negligible effects.

This study evaluated a retrieval-augmented, multi-agent large language model (LLM)-driven, human-in-the-loop framework for detecting cutaneous immune-related adverse events (cirAEs) from clinical notes. Compared with unassisted manual review, the LLM-assisted workflow improved accuracy (F1 = 0.88 vs 0.77), inter-rater agreement measured by Cohen's kappa (kappa = 0.82 vs 0.50), and reduced average review time by approximately half. This framework pilots how LLMs can be applied to identify immune-related toxicities across organ systems and, more broadly, enable accurate, scalable, and transparent adverse event data extraction.

A new study reveals that Mixture-of-Experts (MoE) routing in large language models follows a principle similar to Huffman coding, where common tokens are processed by sparse experts and rare, complex tasks engage diverse expert committees. The paper introduces the Frequency-Diversity Law, identifies a redundancy trap in certain models, and proposes Subset Difference Pruning to improve efficiency.

Existing LLM hallucination mitigation methods either fail to alter internal knowledge or lack cross-domain generalization. Contrastive decoding uses layer-wise differences but has only been studied in transformers. This research investigates mixture-of-experts (MoE) models, finding that while shared-expert MoEs lack layer-wise differences, higher layers exhibit distinct expert activation patterns between factual and non-factual outputs. The proposed EAACD method splits high-layer experts into reliability groups, contrasts their predictions, and amplifies hallucinations from lower-reliability experts as negative references, outperforming baselines on four QA datasets.

This two-study investigation explores how reasoning-enabled LLMs evaluate literary quality. Study 1 constructs a benchmark of 30 texts across six quality tiers, extracting an implicit theory that values intentionality, craft, depth, and distinctive voice over correctness. Study 2 probes this theory through systematic degradation of canonical prose, finding that vocabulary simplification causes the smallest quality loss while structure and voice loss are far larger. The results suggest LLM judgments are holistic, author-specific, and more sensitive to structural than lexical features, with implications for automated writing feedback and computational aesthetics.

This paper proposes an uncertainty-aware trust estimation method for multi-LLM aggregation, adapting structured expert judgment from decision theory with Cooke-style log weighting to calibrate trust in each LLM. Experiments on MMLU and MMLU-Pro show that while methods perform similarly in homogeneous settings, Cooke weighting is crucial under heterogeneity and contamination, achieving a superior accuracy-reliability balance.

A preregistered experiment tested whether holonomy concentrates on active sparse-autoencoder (SAE) feature planes in Gemma 2 2B. The prediction was falsified in reverse: active-feature planes carried less holonomy than matched mixed-feature controls. The result is a narrow operational reversal, not a causal claim, leaving the cause open.

This paper proposes Generative Bayesian Filtering (GBF), a filtering framework that replaces restrictive observation models with pretrained conditional generative models (CVAE). GBF performs Bayesian prediction-update recursion where the measurement update is a posterior sampling problem combining dynamical prior with CVAE-induced likelihood, transformed into a score-based sampling problem. Experiments on synthetic data and real-world applications like manufacturing monitoring and arrhythmia diagnosis show improved accuracy and robustness.

This study diagnoses adaptive depth in looped transformers through a trajectory-readout lens, finding that fixed-prior depth supervision yields difficulty-aware trajectories and simple post-hoc confidence readouts often match or outperform learned gates. The main failure stems from joint gate training inducing poor trajectories, not limited gate expressivity.

This paper proposes Multimodal CoLRAG-TF, a four-axis fusion architecture that integrates dense text embeddings, BM25 keyword matching, knowledge-graph triple filtering, and image-based similarity for robust retrieval over complex PDFs. The system constructs a multimodal index of 2,403 blocks from 43 Japanese disaster lesson PDFs and extracts 11,414 OpenIE triples for compositional reasoning. Bayesian optimization reveals that the triple axis must dominate (α=0.44) to counter lexical bias. On a 457-pair benchmark, it achieves a Recall of 0.9909 and a 71.6% improvement in multi-hop answer similarity over single-hop queries.

A new study scales LLM-based closed-loop channel configuration search to 250 candidates per cycle, confirming transferable optimization and revealing architectural regularities such as non-power-of-two widths and structured allocation patterns.

A study decomposes Muon's speedup in grokking modular arithmetic, finding orthogonalization (Newton-Schulz iteration) as the key mechanism, not spectral norm constraints. Orthogonalization achieves generalization at 3x lower spectral norm and leads to lower-norm solutions. Reducing Newton-Schulz iterations speeds up grokking but makes solutions fragile.

Research reveals that language models hallucinate significantly more when required to fill mandatory fields like JSON. Even models that answer honestly in free text fabricate responses under format pressure, and most cannot be rescued by explicit options or instructions. The PhantomFill benchmark quantifies this coerced fabrication.

DataPrep-Bench is the first unified benchmark that jointly evaluates LLMs on data construction and data quality evaluation across six domains and multiple base models. It introduces Data-Construction-Skill agent and Distributional Alignment Score (DAS), achieving state-of-the-art results in several domains.

This paper introduces a benchmark for evaluating LLM personalization using Bayesian Persuasion, releasing the SDR-Bench dataset of 6,279 sales stories. Experiments reveal a consistent performance plateau across frontier models, with field validation confirming framework utility.

PlanE proposes a planning framework for building extractive LLMs via data decomposition, instruction tuning, and prompt inference, and introduces a DTI planner to automatically select the optimal base-LLM and combinations, demonstrating effectiveness and generalizability.

Large language models (LLMs) handle many tasks with one set of parameters, but under KV-cached inference it is unclear what task-general structure, if any, is used at decode time rather than during prefill. We propose DecodeShare, a protocol that identifies a low-dimensional subspace consistently shared across tasks in decode-time hidden states, and then tests its causal role by removing that subspace only during decoding. In our experiments, disturbing the discovered shared subspace degrades decision performance far more than disturbing either a prefill-derived or random subspace under the same intervention budget. We further show this decode-shared subspace has practical consequences for activation steering: common steering directions can overlap the task-general decode channel. Project…

InferenceBench evaluates AI agents on open-ended LLM inference optimization. Agents must deploy an inference server and optimize speed under a two-hour budget across four scenarios: prefill latency, decode latency, concurrent throughput, and a balanced mix. Agents improve up to 8x over baselines but fall short of simple hyperparameter search (11.53x). Analysis reveals agents converge on a single framework, lacking diverse exploration.

DC-Leap is a training-free framework that accelerates diffusion large language models (dLLMs) by addressing the Joint Probability Dependence Error (JPDE) that leads to overly conservative confidence thresholds. It introduces Dynamic Contiguous Verification and draft-guided decoding to achieve up to 53.19x speedup on MBPP long-sequence generation and up to 105.02x when combined with KV-Cache, while maintaining generation quality.

JAXBench is a TPU-native benchmark suite of 50 JAX workloads for evaluating AI-generated kernel optimization methods on Google Cloud TPUs. Results show that document conditioning improves correctness significantly, while search structure provides speedups, though high-quality optimization remains challenging.

Research finds that answer variation from repeated runs of a single LLM via temperature sampling (self-consistency) only reveals a single dimension of structure, while an ensemble of multiple different models detects more cross-question signals, better revealing what the model does not know.

A new AI-driven browser extension called ClickGuard uses a hybrid machine learning approach combining transformer embeddings and linguistic features to detect clickbait with 91% F1-score. It provides pre- and post-access warnings, a clickbait likelihood percentage, and a spoiler summary to help users avoid misleading articles.

A new study systematically evaluates the impact of LLM watermarking on medical text performance, finding that watermarks can cause significant degradations like lexical corruption and hallucinated terminology, and that domain-specific evaluation is crucial for safe deployment.

Modern software quality assurance demands intelligent, autonomous systems. This paper presents AINTMA, a multi-agent AI system transforming test management into an autonomous quality intelligence ecosystem. It deploys six specialized AI agents coordinated through secure cloud-native microservices. Evaluation over 18 months on 12 projects shows 88.4% test prioritization accuracy, 43% cycle time reduction, and 340% ROI at 9-month payback.

Local-first desktop agent for everyday work.