The paper introduces an experiment where a small masked language model (DeBERTa) is initialized with visual embeddings for tokens that correspond to image regions, following St. Augustine’s ostensive definition of word learning. The visual initialization leaves a measurable imprint that persists through training, yet it does not improve performance on most BabyLM benchmarks that test abstract grammatical knowledge. However, the seeded models show a consistent advantage in zero‑shot object‑property tasks and in a custom Visual‑Property Swap benchmark that probes color, material, size, and shape knowledge, with the advantage confined to the seeded words and transferable to newly seeded words.
By Lisa Bylinina
The paper introduces a Database Normalization Benchmark (DNBENCH) with 3,275 samples to evaluate how well Large Language Models (LLMs) can perform database normalization from 1NF to BCNF, assessing semantic equivalence, structural accuracy, and logical validity. It identifies common failures in dependency inference, schema decomposition, and inter-table constraint reconstruction across various complexity levels. The authors also propose a Multi-Agent Reasoning for Schemas (MARS) framework that separates evidence extraction, violation diagnosis, and decomposition planning from schema generation, achieving an 82.0% improvement in DNB-SCORE over a single-prompt baseline.
By Dong-Jae Koh, Huisu Kim, SeongHwan Yoon, Lasse M. Jantsch, Chun-Hee Lee, Seonghyeon Lee, Young-Kyoon Suh
AcFlow introduces an inference‑time controller for text‑to‑image diffusion transformers that transports intermediate layer activations through a learned, concept‑conditioned velocity field while keeping the base model frozen. The method allows fine‑grained style intensity control and suppression of unwanted concepts, achieving superior style–content trade‑offs compared to baselines and generalizing to unseen concepts without per‑concept fitting. Experiments demonstrate improved style alignment and qualitative suppression of diverse concepts where direct prompting fails.
By Junran Wang, Zehao Jin, Tianyu Luan, Xinjie Shen
The paper introduces StructMemEval, a benchmark designed to assess how well large language model (LLM) agents can organize their long‑term memory rather than merely recall facts. It compiles tasks that humans typically solve by structuring knowledge—such as transaction ledgers, to‑do lists, and trees—and evaluates agents on these. Experiments show that simple retrieval‑augmented LLMs struggle with such organization tasks, while memory‑augmented agents perform better when explicitly prompted to structure their memory, yet many modern LLMs still fail to recognize memory structures without prompting.
By Alina Shutova, Alexandra Olenina, Ivan Vinogradov, Anton Sinitsin
BridgeMatch is a two‑stage generative solver that preserves the full soft matching matrix for 3D deformable registration. Stage I uses denoising diffusion to estimate a global matching matrix at a coarse resolution, then lifts it to high resolution while maintaining hierarchy and rank constraints. Stage II refines this lifted matrix via a conditional transport bridge, implemented with either a deterministic Flow Matching ODE or a stochastic Brownian‑bridge SDE, and demonstrates improved correspondence accuracy and registration performance on 4DMatch, 4DLoMatch, CAPE, and DeepDeform datasets, especially in low‑overlap scenarios.
By Qianliang Wu, Haobo Jiang, Guangwei Gao, Shuo Chen, Jin Xie, Jian Yang, Yaqing Ding
The paper investigates how data repetition affects Mixture-of-Experts (MoE) language models compared to dense Transformers. Across models from 80 M to 1 B active parameters, MoEs degrade more quickly as data is repeated, with performance dropping significantly beyond 4× repetition and overtaking dense models only when strong regularization is applied. The study also identifies routing stabilization and expert specialization as key factors in MoE overfitting, and explores regularization techniques that can partially mitigate this issue.
By Atindra Jha, Margaret Li, Jure Leskovec, Percy Liang, Luke Zettlemoyer
The paper revisits the continuous diffusion language model Plaid and introduces RePlaid, aligning its architecture with modern discrete diffusion models. RePlaid achieves a compute gap of only 20× compared to autoregressive models, surpasses Duo with fewer parameters, and outperforms MDLM in over‑trained settings. On OpenWebText, RePlaid sets a new state‑of‑the‑art continuous diffusion perplexity of 22.1 and demonstrates superior generation quality, while theoretical analysis links likelihood‑based training to linear cross‑entropy over time and structured embedding geometries.
By Zhihan Yang, Wei Guo, Shuibai Zhang, Subham Sekhar Sahoo, Yongxin Chen, Arash Vahdat, Morteza Mardani, John Thickstun
Datasette 1.0a39 and 0.65.4 are new security patch releases for the current alpha series and the stable 0.65.x family. The updates address security fixes that are important for public-facing Datasette instances, especially those that mix public and private tables. The patches were developed after an extensive audit using advanced AI models and collaborative review, and the process will be incorporated into future development work.
Shopify has decided to move away from React Native and return to separate Swift and Kotlin codebases for its native mobile apps. The company originally adopted React Native in 2020 to avoid duplicating work, enable cross‑stack development, and reduce feature parity overhead, but has since found that modern coding agents can handle much of the implementation, translation, testing, and review work, making native development less of a bottleneck. Shopify also maintains several React Native libraries—react‑native‑skia, flash‑list, and restyle—though the latter will be archived by the end of 2026.
The paper introduces a multi‑signal pipeline for detecting hallucinations in large language model outputs, combining fine‑tuned DeBERTa‑v3 classification, Monte Carlo Dropout uncertainty, and temperature‑scaled calibration. On the HaluEval benchmark it achieves strong performance (F1 = 0.915, AUROC = 0.977) and further improves accuracy to 93.2% with MC Dropout. The authors also demonstrate that applying Direct Preference Optimization to a Qwen2.5‑0.5B generator reduces hallucination rates from 85.5% to 37.7%, and show that domain‑specific fine‑tuning (PubMedBERT on SciFact) yields better results than general‑domain training.
SpecGuard is an inference‑time backdoor detector that leverages speculative decoding—using a draft model to propose tokens and a target model to verify them—without adding extra model computation. By monitoring the draft‑token acceptance rate, SpecGuard identifies when a target model’s behavior shifts toward attacker‑controlled outputs, a signal that appears whenever a backdoor is triggered. The method works across various backdoor types and model families, reliably detecting stealthy attacks that bypass input‑level filters while avoiding the extra generation cost of existing runtime detectors.
RAG-Safety-Bench is a benchmark designed to evaluate how retrieval-augmented generation (RAG) affects the safety of large language models (LLMs). It isolates safety impacts by testing four conditions: non-RAG, RAG with an oracle document, RAG with related but non-answer documents, and RAG with random safe documents. Results on five open-source LLMs reveal an inverse relationship between benign and unsafe capabilities, show that standard safety guardrails do not guarantee safety in RAG, and confirm that even benign documents can trigger unsafe outputs.
César de la Fuente’s laboratory employs Codex and ChatGPT to explore both living and extinct genomes in search of novel antimicrobial molecules. By leveraging these AI tools, the team aims to identify candidates that could help combat drug‑resistant infections.
LOCUS is a post‑training technique that selects a task‑aware low‑rank adaptation subspace to reduce token‑cost while preserving utility. By updating only a tiny fraction of parameters (0.24–0.28 %) on 3 B‑parameter backbones, LOCUS cuts continuation length by up to 39.84 % on Pythia‑2.8B and 14.87–17.58 % on Qwen2.5‑3B, without materially affecting the internal preference diagnostic.
The paper introduces ORCH, a method that applies human organization theory to create task‑specific hierarchical structures for large, heterogeneous embodied AI teams. By combining concurrent and sequential interdependence, ORCH outperforms four existing multi‑agent frameworks across 25 wildfire‑response missions, improving mission outcomes, execution efficiency, exploration, and computational resource use. Both human‑designed and language‑model‑generated ORCH organizations yield significant performance gains, with hierarchical organization enabling sustained concurrent activity and coordinated transitions in long‑horizon missions.
The Eloquence team presents three strategies for the Interspeech 2026 MLC‑SLM Task 2, a multilingual MCQA challenge across 21 languages. They fine‑tune Voxtral‑Mini‑3B with LoRA and cross‑lingual augmentations, achieving 0.72 macro‑accuracy; they use multimodal in‑context learning on the frozen Voxtral‑24B to correct label bias, reaching 0.81; and they deploy a training‑free retrieval system with a voice‑anchored memory, scoring 0.68. All approaches surpass the official baseline.
Negative Self-Distillation (NSD) is a new framework for improving large language models by encouraging them to diverge from their own flawed reasoning rather than imitate privileged solutions. Unlike On-Policy Self-Distillation (OPSD), which can suppress uncertainty and exploratory behavior, NSD generates a question‑specific negative condition (e.g., a careless reasoner) and pushes the student’s distribution away from it. A dynamic gating mechanism isolates reasoning‑critical tokens so that only behavioral flaws are penalized, preserving linguistic capabilities, and empirical results show NSD consistently outperforms OPSD and other label‑free self‑bootstrapping RL baselines.
Ecdysis is a framework for training runtime harnesses for large language model agents more efficiently. It distinguishes between model‑specific issues and systematic harness deficiencies by aggregating failures across multiple task instances and uses Failure‑Driven Collaborative Refinement to diagnose and fix harness problems. The approach reduces training time by up to 1.84× and improves harness reasoning accuracy by 18.56%.
OpenAI’s new Data agent in ChatGPT Work lets users connect company data, uncover insights, and build interactive dashboards using natural language. The feature enables teams to put data to work more easily and interactively with AI assistance.
ZipCodec is a streaming neural speech codec that operates at an ultra‑low frame rate of 6.25 Hz and a bitrate of 0.80 kbps, achieving a theoretical latency of 160 ms. It leverages large‑scale WavLM distillation, a redesigned transformer architecture, a scalar spherical quantizer, and a latency‑aware streaming decoder. Experiments demonstrate that ZipCodec outperforms existing streaming codecs at comparable bitrates in both reconstruction quality and downstream tasks, while remaining real‑time on a consumer‑grade CPU despite its 842 M parameters.