This paper proposes casting multi-label legal annotation as a retrieval task, using a frozen retrieval model to embed documents and label descriptions, then predicting labels via k-nearest neighbors. Experiments on three legal datasets show competitive accuracy, high data efficiency, and zero hallucination compared to generative models.

CHI-Bench evaluates AI agents on realistic healthcare workflows emphasizing policy density, multi-role composition, and multilateral interaction. Best agents achieve only 28.0% task resolution, with strict pass rates under 20% and single-session performance dropping to 3.8%.

This paper presents a computational approach using LLMs and a RoBERTa-based classifier to automatically identify manner and result verbs in sentence context, achieving up to 89.6% accuracy. It provides a scalable measurement tool for verb semantics in developmental research.

This paper introduces a novel generative framework for emotion intensity evaluation, moving beyond discrete classification. By constructing a dataset of continuous intensity scores and fine-tuning open-weight LLMs to output 0-100 values, the framework outperforms classification baselines and shows generalization to related constructs, particularly beneficial for finance.

arXiv reports progress on its HTML Papers offering since 2023: community-driven improvements resolved roughly half of 6,000 user reports; corpus-scale conversion aims for 90% error-free HTML (currently 75%); initial MathML 4 Intent annotations for accessible speech; and an in-progress Rust port of LaTeXML to reduce compute costs and enable faster previews. The project remains experimental but is maturing.

PQR is a framework that generates diverse and realistic user queries to uncover failures in QA agents, using iterative interaction between query refinement and prompt refinement modules. It detects 23-78% more unhelpful responses in e-commerce QA agents, with greater diversity and realism than prior methods.

A study reveals scaling laws for skill libraries in LLM agent systems: routing accuracy decays logarithmically with library size, while correct execution can improve downstream decisions by ~4×. Optimization raises routing accuracy from 71.3% to 91.7% and reduces hijack rate from 22.4% to 4.1%.

AdaGraph is a graph-native clustering algorithm from the Structure-Centric Machine Learning (SC-ML) paradigm. It operates entirely within kNN graph topology, replacing Euclidean distances to dissolve the curse of dimensionality. It requires no pre-specified number of clusters k, natively handles noise, and scales via the SLCD framework. Paired with Graph-SCOPE for unsupervised tuning, it achieves mean ARI=0.900 on synthetic benchmarks from d=10 to 5000, outperforming traditional validity indices. It demonstrates superior performance in gene co-expression, text clustering, and materials science.

A new study introduces a Residual Gap-Aware Transformer that combines mixed-effects statistical reference with transformer residual learning to predict 24-month CDR-SB changes using ADNI data. The model outperforms baselines, reducing MSE by 13.1% and increasing prediction-observation correlation by 26.4%.

This paper systematically investigates how action information can be incorporated into the state update function of recurrent neural networks (RNNs) in reinforcement learning (RL). The authors discuss multiple encoding choices, provide empirical evaluations across illustrative domains, and outline future work and challenges specific to RL.

A study shows that in self-play reinforcement learning, removing all positive-reach contingent decision points causes agents to rapidly collapse to a near-maximal loss fixed point, while preserving even one such point prevents collapse. The mechanism is co-adaptation under constraint, not the perturbation itself, and is timing-invariant, reversible, and intensified by function approximation.

We study adversarial action masking in self-play reinforcement learning: an attacker selectively removes legal actions from a victim's action set. Experiments across poker and non-poker domains show learned masking causes more damage than random masking and perturbation baselines, persists across multiple RL algorithms, transfers across agents, is amplified by self-play, and shows no recovery under extended masked training.

SignMuon is a 1-bit, matrix-aware optimizer that combines majority-vote sign aggregation from signSGD with the polar-step framework of Muon. It achieves non-convex convergence rate of O(1/√T) under spectral-norm smoothness and bounded-variance stochastic gradients. Evaluated on 330 CIFAR-10/ResNet-50 configurations, SignMuon attains the best validation accuracy of 92.15%; its 4-GPU majority-vote variant reduces training time by 37% at matched effective batch. On nanoGPT, SignMuon achieves lower perplexity and better anytime performance, with favorable weak-scaling up to 16 GPUs.

This paper introduces a counterfactual comparison-based credit assignment framework for reinforcement learning in multi-step reasoning with LLMs. By sampling multiple reasoning trajectories under the same input and treating their differences as implicit approximations of alternative decisions, it constructs an implicit process-level advantage estimator that transforms sparse terminal rewards into step-sensitive learning signals. Based on this, Implicit Behavior Policy Optimization (IBPO) is proposed, which significantly improves training stability and performance upper bounds on mathematical and code reasoning benchmarks.

This paper proposes mirror descent-type algorithms for variational inequality problems with functional constraints. The algorithms switch between productive and non-productive steps based on constraint satisfaction, achieving optimal convergence rates for bounded monotone operators and Lipschitz convex constraints. A modification that considers only the first violating constraint saves runtime for many constraints. The analysis extends to δ-monotone operators, allowing application to constrained minimization without exact subgradient information.

This study conducts comprehensive 10-phase optimization experiments on Apple M3 Ultra (60-core GPU, 512 GB unified memory) to achieve real-time camera img2img transformation. By combining CoreML conversion of the distillation-specialized model SDXS-512 with a 3-thread camera pipeline, it reaches 22.7 FPS at 512x512 resolution. The work demonstrates that CUDA optimization insights do not transfer to Apple Silicon's unified memory architecture, with quantization showing no speedup, parallel inference being ineffective, and the Neural Engine unsuitable for large models, providing practical guidelines for Apple Silicon diffusion model inference.

LinAlg-Bench is a diagnostic benchmark that evaluates 10 frontier LLMs on structured linear algebra problems across 3x3, 4x4, and 5x5 matrices. With 9 task types and 660 SymPy-verified problems, it assesses 6,600 model outputs. Beyond binary accuracy, it uses a three-stage forensic pipeline to classify 1,156 failures into 10 error tags. The central finding is a sharp behavioral threshold at 4x4: below, models fail due to execution errors; above, they shift to computational abandonment, fabricating responses. This transition is near-universal, suggesting a working memory limit rather than a knowledge gap. The study also shows solution strategy rigidity is a strong predictor of 5x5 determinant accuracy and documents a novel structured hallucination mode called constraint-aware confabulation.

This research proposes a knowledge-adaptive architecture that separates visual perception from reasoning, combining a visual encoder with a dynamic knowledge base to tackle environmental variability and resource constraints in edge AI deployments for ecological monitoring. By replacing implicit parameter encoding with an explicit knowledge base, it supports knowledge sustainability and involves cross-disciplinary collaboration with biologists and Indigenous communities for ethical AI co-development.

Recent research shows Universal Multimodal Embedding (UME) benefits from Chain-of-Thought (CoT) reasoning, but explicit CoT traces are computationally expensive. This paper proposes replacing explicit CoT with latent think tokens, which serve as latent variables that can produce explicit CoT traces as observed variables. By optimizing think tokens using CoT generation loss and embedding tokens using contrastive loss, the model achieves high-performance, reasoning-aware representations at constant inference cost. The introduced TTE-Flash-2B model outperforms its explicit-CoT counterpart on the MMEB-v2 benchmark, with interpretable think tokens. Zero-shot evaluation across 15 video datasets shows scaling behavior with more think tokens, motivating adaptive think budget allocation.

PRISMat is a cost-effective, permutation-invariant model for generating crystal slabs. It outperforms LLMs in predicting surface properties like cleavage energy and work function, reducing error by 4× with faster inference. The model addresses the inefficiency of LLMs in high-throughput materials discovery.

A new study reveals that while LLM agents can accurately model a counterparty's preferences in multi-attribute bargaining, they fail to leverage this information for strategic advantage, leading to suboptimal agreements driven by initial anchors rather than utility weights.

Skim is a speculative execution framework that leverages the predictable structure of purpose-built websites to reduce cost and latency without accuracy loss.

Researchers present an AI agent architecture that integrates large language models with laboratory orchestration, enabling scientists to create and monitor automated lab protocols using natural language. Integrated into the Experiment Orchestration System (EOS), it achieves a 97% first-attempt success rate and an order-of-magnitude reduction in interface actions across three simulated labs.

ANNEAL is a neuro-symbolic agent that converts recurring failures into governed symbolic edits of a process knowledge graph without modifying foundation model weights. Its FDKA mechanism localizes, synthesizes, and validates patches, achieving 0% recurring failure rates in tests, outperforming baselines like ReAct and Reflexion.

AgentWall is a runtime safety and observability layer for local AI agents that intercepts every proposed action, enforces declarative policies, requires human approval for sensitive operations, and logs complete execution trails. It achieves 92.9% policy enforcement accuracy with sub-millisecond overhead across 14 benchmarks.

Odyssey launches Agora-1, a research preview of a multi-agent world model that allows up to four humans or AI agents to share and interact inside the same real-time generated GoldenEye-style simulation, maintaining consistency across different viewpoints.

Multi-Claude is a native macOS app that allows users to run multiple Claude accounts side by side with isolated profiles, sessions, and settings. It solves the problem of constant logging in and out between personal and work accounts. The app is a one-time purchase with a launch discount.

Deutsche Börse Group's StatistiX team faced a notebook migration challenge affecting 2,000+ users with a 2027 deadline. They built a Databricks App that handles structural conversion automatically and generates context-aware Genie prompts for AI-assisted logic reconstruction. Notebook redevelopment time dropped from hours to 15–20 minutes per notebook.

Tomorrow, May 20, the China AIGC Industry Summit kicks off in Beijing. Twenty heavyweight guests will discuss Agent, multimodal AI, applications, and computing power. Check traffic and set your alarm – we'll see you there!

TuriX AI has released an updated version of its open-source computer-use agent, achieving 80% success on macOS benchmarks and ranking 3rd on OSWorld. The agent supports desktop automation, hot-swappable models, and skills via markdown playbooks. Latest releases include TuriX SuperAgent and TuriX SuperPower 3.0.0-alpha with expanded capabilities.