ExtractConf introduces a confidence engine that uses dual LLM readings—field-guided and document-guided—to detect unreliable extractions. It fuses disagreement between calls, LLM uncertainty, and document signals into a classifier, achieving 0.928 ROC AUC on invoices and reducing selective prediction risk by 70%.
Bayesian control improves tool-use decisions in coding agents by modeling uncertainty and dynamically choosing actions. It outperforms fixed-rule orchestrators, especially when verification is costly and critics provide informative but imperfect feedback. The method also produces a more interpretable correctness score than token-probability or raw tool-success metrics.
RaDaR, a compact reasoning large language model, outperformed other open-source models in rare disease diagnosis. In a randomized trial, RaDaR improved physicians' diagnostic accuracy by 21.44 percentage points over internet search alone.
Cross-lingual prompting strategies improve factual knowledge retrieval across 17 diverse languages. The approach outperforms native-language scaling in compute efficiency and enhances cross-lingual consistency beyond accuracy gains.
Qwen-AgentWorld-35B-A3B and Qwen-AgentWorld-397B-A17B are the first language world models that simulate agentic environments across seven domains using long chain-of-thought reasoning. Trained via a three-stage pipeline—CPT, SFT, and RL—these models outperform existing frontier models on AgentWorldBench, a benchmark derived from real-world interactions of five models on nine established tasks.
A study evaluates how large language models preserve cultural meaning when generating narratives from equivalent proverbs across 15 languages. Results show semantic consistency in moral lessons, with systematic shifts in narrative agency and structure, and strong convergence across model families. The research highlights that current evaluations may overestimate cultural preservation by focusing only on semantic similarity.
A multi-agent framework sanitizes retrieved content by removing sensitive identifiers through semantic rewriting, reducing privacy leakage in targeted attacks. It maintains strong contextual fidelity with a BLEU-1 score of 0.122, outperforming SAGE's 0.117, and operates as an asynchronous preprocessing step with no added latency to online inference.
Mem-GF introduces a memory-efficient graph filtering method that approximates polynomial graph filters using Krylov subspaces, eliminating the need to store the full item similarity graph. It achieves up to 5.74× lower memory usage and 4.38× faster runtime while maintaining superior recommendation accuracy compared to state-of-the-art methods, scaling effectively to datasets with tens of millions of interactions.
A new distillation method transfers the observation compression strategy of full-history transformers to recurrent models. By training a teacher model to compress observation histories into fixed-size bottlenecks, the approach aligns the student's memory with the teacher's compression. This enables recurrent transformers to achieve near-full-history performance with linear-time complexity, making them viable for long-horizon robotics applications.
LIG extends Integrated Gradients to set-to-set maps in Transformers, enabling token-level attribution within layers. It analyzes module-wise and layer-wide attribution consistency and tracks information flow via separate attention and MLP contributions, using target token embedding and zero or zero-attention outputs as baselines. LIG operates at module boundaries without retraining or custom interpreters, offering a diagnostic XAI tool for Transformer internals.
A finite-machine inference model uses cost geometry to quantify belief transitions, combining optimal transport with Fisher information. The framework reveals a wall, honesty, and rigidity in belief spaces, with the Gaussian belief achieving maximal hyperbolic curvature. Thermodynamics sets the cost unit, and the geometric floor of precision diverges at certainty, with the value -1/4 representing a key scale.
The paper argues that reference in large language models is not a fixed link but a profile-based, context-sensitive, and numerically structured phenomenon. It proposes that LLMs ground reference through linguistic traces parameterized via optimization, with referential profiles distributed and activated through context-sensitive computation, supported by mechanistic interpretability findings.
A study on neural cellular automata shows that linguistic distance slows consensus and induces mild group divergence without full fragmentation. A collective trained under diverse communication protocols remains robust to mismatch, unlike one trained uniformly, and these results are consistent across ring and 2D grid structures, with parallels to human group dynamics.
Dutch language models exhibit coherence illusions similar to human readers. Surprisal and attention entropy metrics show that models are misled by context matches, with energy from associative memory identifying discourse coherence mechanisms.
ARCO introduces a rubric framework that enables step-level credit assignment for multi-step LLM agents. It jointly updates a shared model with generation and scoring heads, allowing the rubric content and scoring function to co-evolve via on-policy data, improving performance and interpretability across benchmarks.
A study uses FastGAN to generate 10,000 synthetic hyperspectral images of faba bean leaves, preserving real spectral and structural features. Transformer-based models, particularly Vision Transformer, achieve the highest accuracy and F1-scores in classifying healthy versus aphid-infested leaves, outperforming classical CNNs and demonstrating improved disease detection with reduced false negatives.
Topological Neural Dynamics (TND) introduces a neuron-wise framework for sequence modeling, where each neuron evolves independently through a directed graph structure. In a single-player Pong behavior cloning task, TND achieves a mean of 17.47 consecutive catches per round, surpassing all baseline models by more than three times.
NASDAQ addresses low-dimensional observation challenges in reinforcement learning by normalizing observation spaces to balance reconstruction losses across dimensions. The framework combines value learning with short-term value and next observation prediction, achieving competitive or superior performance with less training time compared to existing methods.
A new framework provides instance-level explanations for dysarthria severity assessment by identifying supportive and competing training samples. Using gradient-based influence scores, it links model decisions to perceptible reference cases, enabling auditable and interpretable predictions through controlled deletion experiments.
TASER introduces a framework that dynamically expands and routes atomic skills for continual learning across highly heterogeneous tasks. It reduces catastrophic forgetting and improves plasticity by ensuring semantic distinctness and efficient capacity allocation through skill detection and routing mechanisms. Evaluated on HeteroCLBench, a benchmark of 19 diverse tasks across 9 cognitive dimensions, TASER outperforms existing baselines.