This study examines how different reviewer guidelines—official conference guidelines vs. reviewer-imitating guidelines generated by LLMs from high-quality human reviews—affect automated peer review. Results show official guidelines align best with human judgments, while strict rubric scoring degrades performance.

MioFFAn is an open-source, customizable annotation framework that addresses the scarcity of high-quality datasets for automatically translating mathematical expressions into executable code (Formula Formalization). Based on the MioGatto architecture, it features equation selection, aided symbolic code specification, custom taxonomies, and partial automation via LLMs, enabling human-in-the-loop iterative refinement.

Machine translation systems still produce gender-biased translations. GAND is a new benchmarking resource consisting of English source sentences designed to analyze how contextual cues influence gender translation in ambiguous scenarios. The authors translate a subset into two grammatical gender languages, add contrastive translations, and perform feature attribution analysis to identify source words that affect gender translation.

DomainPilot introduces a two-stage data mixture optimization framework guided by domain-level loss, using scaling and mixing laws to improve LLM fine-tuning performance without extra data or cost.

Introduces a scalable, inference-only data valuation pipeline that approximates Shapley values without iterative retraining. Using semantic k-NN graphs and log-likelihood shifts, it evaluates data utility. Applied to HelpSteer2 and HH-RLHF, it reduces manual audit space by 99.1% and uncovers thousands of hidden preference inversions, revealing flaws in human ground-truth labels.

A novel framework integrating deep learning and statistics enables whole-network gene–environment association analysis using leaf vascular architecture as a phenotype. It fine-tunes EDTER for vein extraction, builds an annotated image database, and applies SSCCA to identify three significant gene–geography interactions in poplar.

This paper introduces Graded Large Language Models (GLLMs), an algebraic framework that equips the transformer representation space with a grading and propagates the induced weighted scalar action through embeddings, self-attention, and the training objective. It extends graded neural networks and graded transformers to autoregressive language models while preserving expressive power and complexity. The geometric picture uses invariant theory, where the benefits of grading are expressed via a Kempf–Ness functional, and optimal grades come from a coincidence point of two moment maps. A standard transformer appears as a semistable isotropic point on the boundary. For level-stratified targets, a minimax separation is proven between graded and uniform priors, with an exponential decay in risk.

This paper introduces Accessibility Plasticity, a principle where neural networks adapt not only by changing computation parameters but also by reorganizing which existing computations interact. A proof-of-concept on sequential learning shows reduced capability modification with maintained performance.

Proposes QFedPolyp, a federated learning framework combining quantization-aware training with low-precision model communication for collaborative polyp segmentation. On multiple datasets, 8-bit quantized communication reduces transmission cost by about 4x while maintaining competitive segmentation accuracy and up to 1.5x faster inference, suitable for real-time clinical deployment.

A new method, ProGPO, addresses sampling imbalance in group-based policy optimization for LLM agents by using first-visit observation coverage to assign advantages when all trajectories in a group fail, improving performance on long-horizon tasks.

Existing adaptive inference methods for Large Language Models rely on observational heuristics, such as hidden-state similarity or activation magnitudes, to drop redundant modules. However, these correlation-based metrics often fail to capture subtle, non-linear structural computations vital for semantic accuracy. We introduce CausalGate, an intervention-guided framework for compute-efficient transformer inference. During a calibration phase, CausalGate isolates individual Attention and MLP sub-layers, zeros out their respective outputs, and measures the exact semantic damage via the Kullback-Leibler divergence of the final logit distribution. To eliminate runtime routing overhead, this structural importance hierarchy is distilled into a global set of static, lightweight scalar gates usin…

Proposes CORVUS, a novel trajectory architecture that decouples file-read actions from observations by maintaining a synchronized registry, significantly reducing redundant tokens and reasoning cycles for LLM coding agents while maintaining pass rates.

Semalith v1.4 is a 184M-parameter DeBERTa-v3-base classifier that performs simultaneous three-axis safety classification—prompt injection, general harm, and financial-services regulatory compliance—in a single forward pass. It outperforms Llama-Guard-3-8B on all prompt-injection benchmarks (7/7) and 11 of 18 overall benchmarks with 44x fewer parameters, achieving zero false positives on 208 benign agentic prompts.

This paper presents a loss-aware feature-map pruning framework using multi-armed bandits. Each feature map is treated as an arm; temporary masking and loss change evaluation guide removal decisions. Results on multiple datasets show that UCB1 and Thompson Sampling preserve accuracy close to unpruned models while reducing computation, outperforming greedy and magnitude-based pruning.

This paper proposes ScenarioGeneratorAgent, a pipeline for synthetically generating evidence-grounded industrial scenarios to expand the AssetOpsBench benchmark. It adds a Smart Grid Transformer asset class and four IEC-based diagnostic tools, achieving 8x speedup with optimizations while maintaining scenario quality.

SF-AMS proposes a strategic forgetting framework for LLM agents that models memory importance as a dynamic utility signal, replacing static retrieval with a survival mechanism and integrating composite importance scoring. The method achieves significant gains on LoCoMo and LongMemEval-s, with up to +9.65 F1 in multi-hop reasoning under Qwen2.5-7B, demonstrating strong generalization across backbones.

This paper introduces program distillation as a simple, efficient alternative to LLM-as-a-judge, distilling decision logic into a committee of programs for transparent, low-cost evaluation. The PAJAMA system aggregates program verdicts and falls back to an LLM when confidence is low, matching the performance of a 13B-size LLM judge across five datasets and four model families. On RewardBench, a reward model distilled from program verdicts outperforms one trained on proprietary LLM labels at two orders of magnitude lower API cost.

MIITA is a continual learning framework designed for small language models (SLMs) under resource-constrained deployments. It stores experiences as compact correction-direction prototypes with semantic anchors and retrieves them at inference using semantic and uncertainty cues, enabling non-destructive reuse of past supervision without backbone updates. Experiments show consistent improvement in final performance and mitigation of forgetting under fixed memory budgets.

DeepLens Diagnosis Agent uses a five-stage pipeline (structured clinical extraction, disciplined retrieval, constrained candidate generation, explicit evidence triangulation, and auditable final decision) to boost a small medical reasoning model JSL Medical Small 7B v2 to 60.14% diagnostic accuracy, outperforming many frontier models at lower cost. The workflow design enables reliable diagnostic reasoning under uncertainty and produces traceable intermediate artifacts for high-stakes applications.

LLMs achieve high accuracy on benchmarks but their answers can vary under meaning-preserving paraphrases. Studying across 4 benchmarks and 13 models, the paper finds instance-level instability with mismatch rates over 23%. Despite correct knowledge often being present, it is inconsistently retrieved; a self-paraphrasing strategy can partially recover latent knowledge and improve performance. This suggests standard accuracy metrics can mask instability, and evaluating consistency across equivalent inputs is a better measure of reliability.

QFoldAgent is a closed-loop multi-agent framework for 5-residue tetrahedral-lattice folding that uses a design agent, a VQE-based quantum-classical pipeline, and a feedback agent to iteratively refine Hamiltonian penalties. It reduces median RMSD from 3.64 Å to 3.20 Å on QDockBank fragments and raises structural validity from 87.5% to 98.7% on unseen sequences.

SeT-Diff is the first foundational model for compute node telemetry and time-series. Using a diffusion-based approach conditioned on semantic descriptions of sensors, it decouples system dynamics from dataset structure. Experiments on a real supercomputer show a reconstruction MAE of 0.047 and thermal inference MAE of 0.033, with zero-shot permutation stability.

This paper introduces C-VCE, a diffusion framework that integrates the classifier into the generative model via a concept bottleneck layer, eliminating the need for external noise-robust classifiers. C-VCE allows users to toggle semantic concepts during sampling, generating minimal edits while preserving the rest of the image. It includes a probabilistic regularizer to balance prediction change and similarity to the original, plus a gradient-based mask to confine modifications. On CelebA, C-VCE achieves comparable or better flip rates with visually closer and less distorted counterfactuals.

This article analyzes AMD's progress in AI accelerator software, moving from a zero-percent chance of success to a meaningful possibility, thanks to leadership changes, open-source strategies, and key customers like Anthropic and Microsoft. However, AMD faces two major risks: slow production ramp of the Helios rack due to design issues, and insufficient internal GPU clusters for development and testing. The article provides specific recommendations to AMD's management and details the MI455X silicon and Helios rack architecture.

Vela is an AI-powered recruiting coordinator designed to streamline hiring processes and boost efficiency. The product is featured on Product Hunt with discussion and links.

CharaLab is an all-in-one AI character creation platform that generates images and videos from text or reference photos. It offers a suite of AI tools including image generation, video generation, background removal, and photo restoration. The platform has a free tier and currently offers a 50% discount.

_done is a platform providing APIs for AI agents to perform real-world tasks, using a pay-per-use model with USDC microtransactions via the x402 protocol. It offers a variety of APIs for network lookups, domain checks, data parsing, and more, enabling AI to autonomously complete tasks that require external data.

Velonus is an AI-native Python DevSecOps tool now in beta. It runs multiple security scanners in parallel, uses AI to rank findings, filters false positives, and automatically generates fix PRs, streamlining vulnerability detection and remediation.

Genie One is an AI coworker that integrates with existing business tools to automate recurring tasks. This article presents four practical use cases: automated business reviews and reporting, meeting preparation and follow-up, knowledge work and document automation, and operational monitoring and alerts. Each use case includes step-by-step instructions to set up workflows in minutes.

A developer uses AI (Claude) to create and publish a new video game daily. Day 106 marks their first attempt at the zombie survival genre, completed in about an hour using Claude Opus 5. They've been consistently releasing a game every day for over four months.