GAP3D introduces a modular diffusion-based approach that aligns VLM-generated latents directly to the patch-level feature space of a pre-trained image encoder, enabling frozen generative models to use VLMs as prompt encoders while preserving spatial structure. It trains primarily on image-text pairs, avoids large-scale 3D data, and demonstrates zero-shot multimodal capabilities, though it currently prioritizes high-level semantics over fine-grained detail.

A new approach called Noise-Aligned Diffusion Bridge (NADB) addresses underfitting near the target endpoint in diffusion bridge models, improving image restoration and translation tasks.

This paper proposes S3MEM, a structured scene-event memory framework for long-horizon interactive question answering. By writing trajectories into structured memory units, using anchor-sensitive retrieval, and exposing a compact token-budget-aware evidence interface, S3MEM significantly improves accuracy and efficiency in answering questions about early events. Experiments on multiple environments show that S3MEM achieves a better accuracy-efficiency frontier than existing methods.

A comprehensive evaluation of 14 open-source safety guard models on a benchmark of 79,331 samples reveals that Qwen Guard (4B parameters) achieves the highest recall (83.97%), while larger models like Llama Guard (12B) miss up to 75% of unsafe content. Model size does not correlate with safety performance, and general-purpose guard models outperform specialized ones.

Aryabhata 2 is a reasoning-focused language model for competitive STEM exams like JEE and NEET, fine-tuned via reinforcement learning on GPT-OSS-20B using PhysicsWallah's question banks. It achieves up to 64% fewer output tokens while outperforming the base model on multiple benchmarks.

Large language models suffer from hallucination in long-form generation. Existing retrieval-augmented models cannot ensure key information stays close to outputs. This paper proposes Micro-Macro Retrieval (M2R), a retrieve-while-generate framework that retrieves coarse-grained evidence externally and extracts key information from a reasoning-built repository, significantly reducing hallucination. It uses curriculum learning-based reinforcement learning for stable training.

This paper presents RightNow-Arabic-0.5B-Turbo, a 518M-parameter Arabic-specialized LLM built on Qwen2.5-0.5B using vocabulary injection and edge-first deployment. It achieves 35.9% mean accuracy on Arabic benchmarks, outperforming all same-class open models, and ties Falcon-H1-1.5B on COPA-ar at one-third the size. The quantized model is 398 MB and delivers 635 tokens/s on a single H100, enabling efficient edge deployment.

A new paper analyzes 17 LLMs (410M-100B+ parameters) and documents that instruction-tuned systems systematically collapse language entropy along discourse and structural dimensions (mean amplification: 1,949-16,853%, peaks: 5,181-209,675%), while suppressing complex punctuation to 3.2-23.2% of baseline. These effects do not worsen under RLHF. Weak intervention (lambda=1.0) exacerbates collapse by 240%, while strong control (lambda=5.0) achieves 40.5% improvement and outperforms frontier models by 96.7-98.2% despite 200-1000x scale disadvantage. Strong control also delivers 15% higher distinct-4, 27% higher vocabulary diversity, and 78% lower repetition than moderate regularization. The findings underscore that alignment requires sufficient control strength, not merely distributional smoothing.

A new framework, MechELK, uses mechanistic interpretability to extract hidden knowledge from LLMs. It combines sparse autoencoders, causal probing, and representation engineering to achieve 84.7% accuracy, outperforming existing methods. The framework is particularly effective at surfacing knowledge when models give incorrect or evasive answers.

A modular framework for generating pronounceable, typologically plausible, and semantically structured artificial lexicons. It samples phoneme inventories from PHOIBLE, generates word forms under interchangeable phonological grammars (deterministic, OT, MaxEnt), and assigns meanings via Swadesh-Leipzig-Jakarta ontology. Evaluation shows probabilistic grammars outperform deterministic and random baselines in phonotactic coherence and typological realism.

As large language models (LLMs) grow in influence, understanding their decision-making becomes crucial. This paper introduces a method to detect concepts within LLM embeddings using low-cost linear probes, enabling monitoring of what models "think" during normal operation. The authors demonstrate concept delineation, probe training, and cross-context tracking across four concepts and three LLMs, paving the way for scalable model transparency.

Multimodal learning often suffers from modality imbalance, where faster-converging modalities dominate optimization. Existing methods typically strengthen weak modalities or adjust gradients, but may compromise strong modalities. This paper proposes Balanced Multimodal Label Reshaping (BMLR), the first label-side approach to promote balance. BMLR reshapes the cross-modal label space to equalize mapping difficulty across modalities, enhancing interaction and injecting rich inter-class information. Extensive experiments show consistent improvement and compatibility.

Metagenomic taxonomic annotation identifies microbial origins of DNA fragments. Traditional similarity-based methods struggle with high diversity and incomplete databases. TaxDistill uses a knowledge distillation framework with a 500M-parameter genomic foundation model (GenomeOcean) as teacher to generate soft labels, reducing label noise. Experiments on seven CAMI2 datasets show TaxDistill outperforms baselines, e.g., improving F1 score on Gastrointestinal dataset from 0.763 to 0.941.

PrismFlow introduces a novel flow matching method that uses Koopman-inspired dynamical experts to correct estimator smoothing in standard flow matching, recovering high-frequency and fine-grained temporal structures. It achieves state-of-the-art performance with 15.6% gain in Context-FID and 38.6% improvement in Discriminative Score.

This paper proposes COM, a strategy that integrates geometric constraints into token initialization and training to preserve the inherent continuity and ordinality of time series tokens, consistently improving the performance of token-based time series LLMs on multiple benchmarks.

This paper studies self-play reinforcement learning in the four-player imperfect-information card game Big 2. PPO outperforms Monte Carlo Q approximation, SARSA, and Q-learning against various opponents. Moderate entropy regularization improves PPO by preventing overdeterministic policies, and current-policy self-play provides a stronger finite-budget curriculum than alternatives.

TRACE is a trajectory-aware LLM-reasoning agent for molecular lead optimization that treats tool selection as a sequential decision-making problem, enabling forward-looking structural refinement under constraints, achieving higher success rates and property improvements on ADMET tasks.

Recent work shows RL retains prior capabilities more effectively than SFT. This paper extends to the mechanistic level, introducing differential circuit vulnerability to measure circuit degradation. On Qwen2.5-3B-Instruct for scientific QA, SFT adapts faster but causes greater circuit disruption and forgetting, while RL preserves circuits at the cost of slower adaptation. Results suggest circuit preservation explains RL's robustness against catastrophic forgetting.

This paper studies behavioral alignment and representation dynamics of LLM agents in financial environments using TradeArena. It identifies measurable pre-failure signatures like planning embedding drift and effective-rank contraction. Structured risk feedback can serve as an external alignment signal but is not a universal performance enhancer. A 51-stock experiment reveals a correlation blind spot where LLM rationales justify concentrated exposure to coupled assets.

A study reveals that knowledge editing methods like ROME and MEMIT share a common mechanism despite fact-specific weight changes. A binary mask trained over edited weights reverses over 70% of edits, indicating a shared functional subspace. The mask works by eliminating overattention in later layers, and injecting it during editing drops success from 98% to 38%. Edits suppress rather than overwrite knowledge, explaining failures to propagate changes. This work aids detection and defense against unwanted edits.

VFEAgent is an end-to-end multi-agent system that automates finite element analysis (FEA) modeling and simulation directly from input images and problem descriptions. It combines a multimodal vision-language multi-agent pipeline with a verification-first code synthesis framework, using ReAct-driven reasoning to extract structured FEA specifications and incorporating self-debugging and fallback mechanisms for executability and physical validity. Experiments show high success rates in generating complete, physically valid simulations, outperforming LLM-based baselines in reliability and correctness, and promising to free engineers from tedious manual analysis.

A new study uses five frontier LLMs from Anthropic and OpenAI as 'agentic curators' in a self-contained workspace to automate phenotype annotation. The agents achieved consistency within the range of human curators and substantially outperformed traditional NLP tools, addressing the scalability bottleneck in ontology curation.

This paper introduces Orthogonal Concept Erasure (OCE), which uses multiplicative parameter updates for precise concept removal while preserving generative capacity, supporting multi-concept erasure with high speed.

This paper empirically evaluates LLM-generated reviews for scientific papers, finding limited alignment with human reviews that varies significantly across prompts and models. It also shows that authors can game the system by iteratively revising papers based on LLM feedback, achieving statistically significant score increases for up to 35% of papers.

The Cognitive Categorical Transformer (CCT) is a 306M-parameter architecture that augments GPT-2 Small with cognitive and category-theoretic components, achieving 21.27 perplexity on WikiText-103, a 2.92 (12%) reduction over a fine-tuned baseline. Ablations attribute 84% of the improvement to GT-Full simplicial message passing. The study also identifies a structure/consistency distinction among categorical priors.

This paper proposes behavior-aware auxiliary corrections to stabilize off-policy temporal-difference learning. By replacing the auxiliary covariance matrix with the behavior Bellman matrix, the authors introduce BA-TDC and BA-TDRC algorithms. Theoretical analysis proves fixed-point preservation and almost-sure convergence. Experiments on standard benchmarks show that the behavior-aware replacement improves performance, but regularization is needed for robust results.

This paper proposes STHTD-MP, a behavior-induced Mirror-Prox temporal-difference method that replaces the covariance metric with the symmetric part of the behavior-policy Bellman matrix to improve off-policy prediction speed. Theoretical convergence analysis and numerical experiments on several benchmarks show improved performance over GTD2-MP.

Yukihiro Matsumoto, creator of Ruby, is building Spinel, an experimental ahead-of-time compiler for Ruby with AI assistance from Anthropic's Claude. Spinel compiles Ruby to C code, achieving significant performance gains but with many limitations including unsupported features like eval and threads.

The article explores the concept of 'disposable software' in the AI era, arguing that AI-generated code should be treated as disposable to accelerate development, much like mass-produced furniture replaced artisan craftsmanship. A case study demonstrates successful AI refactoring, and a 'Disposable Code Manifesto' is proposed with three pillars: intent, requirements, and safety.

OpenAI launches Rosalind Biodefense, expanding trusted access to GPT-Rosalind for vetted developers and U.S. government partners advancing biodefense, public health, and pandemic preparedness through frontier AI.