Translate
Live intelligence

Latest AI updates

1,750 updates
AllModelsProductsIndustryResearchTutorialOpinion
Research Editor selected

TLAPS-Bench: Benchmark for AI-Driven Formal Verification of Real-World Systems

TLAPS-Bench is a new benchmark and execution harness for using AI agents to construct machine-checked proofs of real-world system specifications in TLA+. It includes 265 proof targets from 95 specifications covering cache coherence, Byzantine consensus, and write-ahead logs. Frontier models like GPT-6.1 Sol proved 44.3% of a subset of 70 targets, revealing both progress and remaining challenges in end-to-end proof development.

Why it matters: Provides a concrete, reproducible benchmark for measuring AI capabilities in formal verification of complex systems, a key area for AI safety and reliability.

arxiv5d agoScore 85
TLA+formal verificationAI for mathematics
Research Editor selected

Value-Conditioned Proposal Acquisition for Test-Time Reasoning

The paper introduces ValueFormation, a test-time policy that decides after each primary-model draw whether to stop, continue sampling, or acquire a proposal from a heterogeneous reasoning specialist. The policy uses a cost-regularized value scale learned from vote-state features, which is frozen across tasks and primary models. On MATH-500, a single Llama-fitted calibration improves accuracy by over 15 points for three 8B primaries and transfers to Qwen3-14B without retraining. Ablations show that removing acquisition drops accuracy by 20.43 points, confirming the benefit of state-conditioned acquisition over mere specialist access.

Why it matters: Provides a practical, transferable method for improving test-time reasoning with heterogeneous specialists, backed by strong empirical results on math and code benchmarks.

arxiv5d agoScore 88
test-time reasoningLLM inferencevalue-conditioned acquisition
Research Editor selected

KAN vs. MLP Benchmark on Clinical Tabular Data Shows Trade-offs in Precision vs. Efficiency

A rigorous benchmark compares nine KAN variants against MLPs and Random Forest across seven medical datasets. Results show KANs (especially TabKAN and ReLUKAN) improve minority-class detection on imbalanced data but require 1.2x–1.48x longer inference and up to 12.5x more parameters. StandardMLP remains the efficiency Pareto frontier, while KANs are viable when precision on rare positive cases is the priority.

Why it matters: Provides the first large-scale, statistically rigorous comparison of KANs vs. MLPs on clinical tabular data, offering actionable guidance for practitioners.

arxiv5d agoScore 88
Kolmogorov-Arnold NetworksMulti-Layer Perceptronsclinical tabular data
Research Editor selected

Integrating Domain Adaptation and Retrieval-Augmented Generation for Saline-Alkali Land Reclamation Planning

This paper presents Wenyao, an agricultural decision-support system that combines domain adaptation with retrieval-augmented generation for saline-alkali land reclamation planning. Using a bilingual corpus of 8,664 research papers and a QwQ-32B model fine-tuned via low-rank adaptation, the system achieves 95.83% accuracy on domain-specific multiple-choice questions and strong expert ratings. The study demonstrates complementary benefits of adaptation and retrieval in literature-informed planning.

Why it matters: Provides a concrete, rigorously evaluated example of combining domain adaptation and RAG for a specialized agricultural planning task.

arxiv5d agoScore 88
domain adaptationretrieval-augmented generationagricultural AI
Research Editor selected

Tug of Watts: A study on the Energy - Accuracy Trade-off of Large Language Models on Medical Question Answering Tasks

This study examines the energy consumption and carbon footprint of 13 LLMs (270M to 14B parameters) on medical QA tasks from the OpenMedical-LLM leaderboard. It demonstrates how to estimate energy and emissions, and argues that separate accuracy and energy metrics are insufficient, requiring joint trade-off metrics for informed model selection. The work aims to help practitioners adopt more efficient models for sustainable LLM pipelines.

Why it matters: Provides concrete, recent empirical data on energy-accuracy trade-offs for medical QA, addressing a critical gap in sustainable AI development.

arxiv5d agoScore 85
large language modelsmedical question answeringenergy consumption
Research Editor selected

JABER: Sequential Quality Control for Human-AI Data Annotation under Shared Bias

JABER is a bias-aware sequential quality control framework for human-AI data annotation that maintains a joint Bayesian posterior over label correctness and shared biases. It uses a finite-horizon algorithm to minimize expected QC loss under resource constraints. Experiments show JABER reduces expected QC loss by up to 10% compared to one-step lookahead methods and achieves zero labeling errors on FEVER while reducing costs.

Why it matters: Offers a practical solution to a critical problem in AI data annotation—shared bias—with measurable improvements in accuracy and cost efficiency.

arxiv5d agoScore 88
data annotationquality controlbias
Research Editor selected

Voice-Native LLMs Hear Frustration but Do Not Listen to It

This study evaluates whether voice-native LLMs perceive and utilize frustration in speech. Using a 2x2 design with frustrated wording and voice, the authors find that models are robust to vocal frustration but detect it in audio encoders. The VoxAffect dataset of 1,412 clips is released. Results show models hear frustration but do not adjust outputs accordingly.

Why it matters: Provides concrete insights into how voice-native LLMs process emotional cues, with implications for user experience and model design.

arxiv5d agoScore 85
voice-native modelsemotion recognitionrobustness evaluation
Research Editor selected

LANCE: Bridging Low-Level Image Fusion and High-Level Semantic Understanding

Researchers propose LANCE, a framework that uses structured language as an explicit semantic state to connect remote sensing image fusion with downstream semantic tasks. A frozen vision-language model maps inputs to text records, a language-conditioned generator fuses imagery, and a dense verifier re-describes the output, using text-text consistency and contrastive losses without dense annotations or VLM fine-tuning. On seven test sets across four satellite sources, LANCE improves reduced-scale reference metrics by up to 11.7% and boosts downstream detection and open-vocabulary F1 scores over the closest competitor.

Why it matters: Provides a concrete, novel method that unifies image fusion and semantic understanding with measurable improvements on multiple benchmarks.

arxiv5d agoScore 88
image fusionsemantic understandingremote sensing
Research

Maswao: Automatically Collected SASL Dataset for Sign-to-Text Translation

This paper introduces Maswao, a dataset for South African Sign Language (SASL) built with minimal human labor by automatically collecting news broadcast videos, converting audio to text via ASR, and using motion valleys to segment sentences. The authors benchmark sign-to-text translation with RNN and Transformer models, finding Transformers outperform recurrent baselines. Best BLEU scores range from 13.78 (BLEU-1) to 1.54 (BLEU-4), with ChrF=18.09 and ROUGE-L=10.51.

arxiv5d agoScore 78
Sign language processingDataset collectionNeural machine translation
Research

Operationalizing theorem interestingness for open-ended exploration

This paper studies what makes mathematical theorems interesting by modeling how mathematicians select results in formal libraries and papers. The authors compare feature-based evaluators against a large language-model judge, finding that ex-ante features predict selections as well as the LLM and that the two complement each other. Interestingness is largely shared across mathematicians, with structural features like statement and proof length carrying most of the signal, while subjective variation depends on practices and individual authors. The approach can be applied to other sciences.

arxiv5d agoScore 72
AI-driven discoverymathematical reasoningevaluation metrics
Research Editor selected

LLMs Show Systematic Political Bias in Online Debate Receptiveness

A study reconstructing 5,376 Debate.org debates as LLM conversations and running nearly 150,000 controlled multi-turn interactions finds that LLMs are consistently more receptive to left-leaning users than right-leaning users, even when positions are identical. The bias persists across models and repeated interactions, but fine-tuning on different political corpora can reduce or reverse it. This reveals interactional asymmetries in AI-mediated political discourse.

Why it matters: Provides concrete, recent evidence of political bias in LLM-mediated debates, with implications for online discourse and AI fairness.

arxiv5d agoScore 82
LLM biaspolitical discourseAI-mediated communication
Research

Training-Time Domain Transfer Outperforms Inference-Time Virtual Staining for Novel Pathology Modalities

This paper proposes training-time domain transfer for computational pathology, generating synthetic target-domain training data from archived routine samples to enable direct diagnosis on novel imaging modalities without inference-time generation. In a prostate histopathology proof-of-concept, this approach achieved higher concordance with conventional H&E-based cancer detection than inference-time virtual staining, for both raw and deparaffinized tissue. The method reduces hallucination risks and computational overhead while supporting emerging diagnostic modalities.

arxiv5d agoScore 78
domain adaptationmedical imagingdiagnostic AI
Research

Why Tabular Foundation Models Excel on Relational Data They Have Never Seen

This paper investigates why single-table tabular foundation models (TFMs) perform well on relational databases despite no relational pre-training. The authors confirm this phenomenon empirically and provide multiple complementary explanations based on parameter-free relational encoders that map relational effects to embeddings mimicking TFM training distributions. The analysis suggests new directions for improving relational database foundation models, especially for interpretability and agentic integration.

arxiv5d agoScore 78
Foundation ModelsRelational DataIn-Context Learning
Research

PlagiaCorpus: Bilingual Adversarial Dataset for Forensic Plagiarism Detection

PlagiaCorpus is a new bilingual Arabic-English dataset for benchmarking plagiarism and text-reuse detection, including LLM-generated reuse and cross-lingual cases. It contains 198,643 text records with 163,758 queries across five reuse classes of increasing difficulty, plus a safe class. The dataset is split by source passage to prevent leakage and is released under CC BY-SA 4.0. This resource addresses the need for robust evaluation of plagiarism detection in the era of LLMs.

arxiv5d agoScore 78
Research Editor selected

BlackSwan: Refutation-Based Action Verification for Language Model Agents

BlackSwan is a new verification procedure for language model agents that checks proposed actions by actively seeking evidence that refutes them, rather than only confirming supporting conditions. It predicts intended effects, formulates testable hypotheses for failure, and tests them before execution. On AppWorld, it improves success rates by up to 25.2 percentage points, and on WebShop by 12.7 points, even without thinking mode or demonstrations.

Why it matters: This introduces a novel, effective verification paradigm with concrete benchmark gains, offering practical value for AI agent safety.

arxiv5d agoScore 85
AI agent safetyverification methodslanguage model agents
Research Editor selected

BRIDGE: Balancing Representations by Generating Under-represented Data for Long-tailed Self-Supervised Learning

BRIDGE is a label-free feedback loop that improves self-supervised learning on long-tailed data by identifying under-represented images via kNN distance and generating local variants with a frozen diffusion model. It outperforms baselines across five long-tailed regimes on linear-probe and kNN tasks, and extends gains to ViT-S, low-label fine-tuning, and segmentation. The method requires no labels and leaves the pretraining objective unchanged.

Why it matters: This paper presents a concrete, novel method with strong empirical results for a common real-world problem (long-tailed data) in self-supervised learning, offering practical value for researchers and practitioners.

arxiv5d agoScore 85
self-supervised learninglong-tailed recognitiondata generation
Research Editor selected

Specification-Aware Predictive Uncertainty Decomposition

This paper introduces a probabilistic framework that decomposes predictive uncertainty into specification, residual aleatoric, and epistemic components, addressing underspecified inputs in LLMs. It shows that aleatoric uncertainty splits into specification uncertainty (expected reduction from revealing user intent) and residual aleatoric uncertainty (remaining risk after intent is known). Experiments across domains indicate specification uncertainty best detects underspecified inputs, while residual aleatoric uncertainty best detects in-distribution errors.

Why it matters: Provides a novel, formal decomposition of uncertainty with practical implications for LLM clarification and error detection, backed by experiments.

arxiv5d agoScore 82
predictive uncertaintylarge language modelsBayesian risk
Research Editor selected

CAREFUL-Bench: Effect-Aware, Judge-Free Evaluation of Multi-Turn LLM Function Calling

CAREFUL-Bench enhances the Berkeley Function Calling Leaderboard (BFCL) by making evaluation effect-aware without replacing it. It infers effect declarations for all 128 functions via simulator execution and uses a judge-free replay scorer that reproduces official results while reporting unrequested destructive calls, duplicate writes, redundant reads, and token costs. Across 11 open models, effect-aware scoring preserves ranking but reveals quality gaps, and a confirmation-aware protocol shows models rarely obtain explicit permission before destructive actions.

Why it matters: Provides a concrete, recent improvement to function-calling benchmarks, exposing critical limitations in current evaluation and offering actionable metrics for model developers.

arxiv5d agoScore 82
LLM function callingbenchmark evaluationagent safety