The paper proposes a modular relational deep learning approach that decouples row encoding from graph message-passing. It introduces a transformer-based Universal Row Encoder that uses schema metadata to generate invariant row embeddings, enabling better generalization across databases and improving convergence on RelBench benchmarks.
The BIPC framework reduces solution time for large-scale mixed-integer programs by identifying a backdoor subset of variables that drive computational complexity. Using supervised learning, it predicts backdoor variable values and intervals, then solves a reduced problem with these predictions, achieving significant speedups with minimal quality loss. This enables rapid, high-quality solutions under parameter perturbations in real-world systems like power and supply chains.
A new post-training method uses multi-metric perceptual rewards to optimize speech enhancement models. It directly applies non-differentiable metrics like DNSMOS, WER, and UTMOS as rewards via Group Sequence Policy Optimization, achieving state-of-the-art results on DNS2020. Human evaluation confirms that combining multiple metrics outperforms single-metric approaches, reducing reward hacking.
w$^{2}$VLA introduces a modular approach that decouples declarative and procedural knowledge in Vision-Language-Action models. By restructuring information flow, it enables robust behavior cloning and unprecedented zero-shot skill transfer across unseen, dissimilar objects.
llama-server crashes when a user pastes a long prompt as a text file alongside an image, treating it as a file attachment. The server works fine when the prompt is sent in smaller batches but fails when the full prompt is combined into one text block and submitted with the image.
A new framework using cross-lingual question answering reveals users' mental models of speech translation systems. Users develop stronger mental models with practice, especially with source language knowledge, relying on surface-level error cues. Providing speech transcriptions improves model development, showing cross-lingual question answering's potential in human-AI collaboration research.
We propose decomposing structured annotation tasks into sub-tasks to reduce overall inferential load. By identifying salient anchor entities—centers in the space of valid annotations—we constrain output complexity and improve cost-efficiency. We provide guidelines for decomposition and a procedure to allocate sub-tasks across human and model annotators for optimal quality under fixed budgets.
FMLM+ introduces Posterior Refinement, a strategy that enables adaptive self-correction during inference. By combining flow map transport with masking-style noise schedules, it achieves high-fidelity language generation with 32x fewer noise-free evaluations, outperforming both MDM and FMLM in speed-quality tradeoff.
This study introduces the Match Task to Objective (MTO) framework to align pre-training and fine-tuning objectives with specific tasks. The framework enables automated, unsupervised data adaptation and delivers performance gains of over 120% in few-shot settings, outperforming baselines in both few-shot and full-dataset scenarios. It also enhances prompt-tuning by providing effective soft prompt engineering guidance.
MedBench v5 introduces a dynamic, process-oriented benchmark for clinical multimodal models, featuring clinical cognitive responsiveness and atomic skills across 63 tasks. It includes stressors for degradation analysis and monitors hallucination propagation through five reasoning nodes, revealing that strong task performance does not ensure process stability.
A deep learning study analyzes sentiment evolution across review rounds in 11,063 Nature Communications papers. As review rounds increase, positive sentiments rise and negative ones decline, with aspect-level sentiments showing a negative correlation to the total number of rounds, particularly in 'experiments', 'research significance', and 'result analysis'.
CALIBER introduces a method that elicits and supervises confidence estimates at two stages: before and after reasoning. It reduces Expected Calibration Error by 52.5% on BigMathDigits for a 7B model, achieving the best Brier score and AUROC, and performs best on out-of-distribution benchmarks like GPQA and TriviaQA.
Bad contexts in conversations can lead to 'pigeonholing', where models repeat incorrect answers or narrow down to a single response. Experiments show performance drops of 38-40% and worsening errors with more conversation turns, even when initial inputs are correct. A new method, RLVR with synthetic errors, improves model performance by 43-60% under such bad contexts.
This review presents a taxonomy of transformer-based language models across domain verticals, covering encoder-only, decoder-only, encoder-decoder, long-context, permutation-based, and generator-discriminator variants. It evaluates post-2023 advancements like instruction tuning and mixture-of-experts scaling, and assesses model deployments in healthcare, finance, legal, education, customer service, creative writing, and scientific work, linking each to specific capabilities. The paper critically analyzes model architectures on four key deployment axes, quantifies parameter count versus energy cost, and examines how alignment methods, data provenance, and benchmark saturation define 'state of the art'.
Age of LLM introduces a turn-based 1v1 benchmark where two LLMs compete on a 13x7 grid under fog of war, full diplomacy, and strict JSON reliability rules. Findings show the nuclear rush dominates, diplomacy is prolific but rarely succeeds, and illegal actions reveal belief-tracking errors, with a weak link between reliability and victory. The corpus is small and unbalanced, and the results offer a preliminary view of LLM reasoning under adversarial uncertainty.
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.