The paper introduces AssayBench-Loop, a large benchmark of 1,389 CRISPR screens across five phenotype categories, and builds on it to develop AssayLoop, a sequential experimental design framework that combines a transformer-based acquisition policy (AssayFormer) trained on historical data with LLM-derived biological priors. AssayLoop achieves a 5.67‑fold enrichment over random selection, recovering 27.7% of hits after testing only about 5% of the library, and outperforms existing adaptive-design methods and standalone LLMs. The authors also present AssayLLM, extending the approach directly to an LLM via task‑specific post‑training, and show that performance improves with more historical training data and transfers to unseen phenotype categories.
By Carl Edwards, Edward De Brouwer, Xiner Li, Namkyeong Lee, Ehsan Hajiramezanali, Anne Biton, Sara Mostafavi, Gabriele Scalia
ExpTest is an autonomous learning‑rate controller that uses the training loss curve as an online signal to perform sequential statistical tests on theoretically motivated windows, detecting convergent behavior and triggering learning‑rate reductions. It combines a covariance‑based initial learning‑rate estimate, curvature‑motivated window sizing, and a two‑phase test‑driven decay, relying on the approximately exponential decay predicted under linearized network dynamics. Experiments on regression, classification, forecasting, and natural‑language tasks across various architectures show that ExpTest achieves competitive performance compared to hand‑tuned SGD baselines and recent learning‑rate‑free methods, without requiring manual initial learning‑rate selection or predefined scheduling.
By Zan Chaudhry, Naoko Mizuno
The paper addresses training instabilities in large language model pretraining, specifically output logit divergence that occurs near the end of training. By analyzing the geometry of output embeddings, the authors identify anisotropic embeddings as the root cause and propose Output Embedding Centering (OEC) as a mitigation strategy. OEC can be applied deterministically as μ‑centering or as a regularization loss μ‑loss, and experiments show both variants outperform the existing z‑loss method while matching logit soft‑capping in stability, even without weight tying. Additionally, μ‑loss is less sensitive to hyperparameter tuning than z‑loss.
By Felix Stollenwerk, Anna Lokrantz, Niclas Hertzberg
The paper investigates how large language models (LLMs) can exhibit stigma toward people with psychological conditions by examining their intermediate reasoning steps rather than just final answers. Using clinical expertise, the authors develop a framework to identify and rate stigmatizing language in LLM reasoning, distinguishing between overt prejudice and subtler biases. They also expand an existing mental health stigma benchmark to include more psychological conditions, finding that reasoning analysis reveals far more stigma than traditional multiple-choice evaluations and exposes flaws in the models’ logic and understanding of mental health.
By Sreehari Sankar, Aliakbar Nafar, Mona Barman, Hannah K. Heitz, Ashwin Kumar, Pouria Tohidi, Dailun Li, Danish Hussain, Russell DuBois, Hamed Hasheminia, Farshad Majzoubi
The paper introduces an AI economist agent that integrates large language models, retrieval‑augmented generation, knowledge graphs, and quantitative models to conduct evidence‑based economic and financial scenario analysis. The framework orchestrates LLM agents to plan analyses, retrieve relevant evidence, and structure economic mechanisms, while registered quantitative models produce numerical outcomes and predefined tests validate intermediate results for inclusion in the final report. Applied to European macro‑financial stress scenarios and bank capital analysis, the empirical study demonstrates the agent’s ability to combine flexible evidence retrieval and scenario construction while maintaining traceability to sources and explicit model calculations.
By Masahiro Kato
SpecGuard is an inference‑time backdoor detector that leverages speculative decoding—using a small 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 detects when a target model shifts toward attacker‑controlled behavior while the draft model does not, signaling a backdoor trigger. The method reliably identifies a range of backdoor types, including stealthy attacks that bypass input‑level filters, across multiple model families.
By Rui Wen, Ahmed Salem, Andrew Paverd, Mark Russinovich, Zheng Li
PathoHR is a new pipeline for predicting breast cancer survival from high‑resolution pathological images. It uses a plug‑and‑play Vision Transformer to enhance patch‑wise whole slide image representations, evaluates multiple similarity metrics to optimize feature learning, and shows that smaller, enhanced patches can match or surpass the accuracy of larger raw patches while cutting computational cost. The authors provide experimental evidence that this approach improves both accuracy and efficiency in computational pathology.
By Yang Luo, Shiru Wang, Jun Liu, Jiaxuan Xiao, Rundong Xue, Zeyu Zhang, Hao Zhang, Yu Lu, Yang Zhao, Yutong Xie
BruNet is a new cross‑domain transfer framework for automatic bruise segmentation that combines a ViT‑based visual encoder (either self‑supervised DINOv3 or pretrained LingBot‑Vision) with a SAM‑based mask decoder. The model is trained on the HAM10000 skin lesion dataset and evaluated on a separate bruise dataset without any fine‑tuning, achieving superior performance over CNN‑based models, state‑of‑the‑art segmentation models, ChatGPT‑4o/5‑assisted SAM2 zero‑shot baselines, and the medical‑oriented MedSAM. This work represents the first study to address pixel‑level localisation of bruises, demonstrating strong cross‑domain generalisation.
By Qiming Wang, Richard J. Motley, Ebube E. Obi, Xianfang Sun, Paul L. Rosin
FluxMoE introduces an expert paging system that decouples Mixture-of-Experts (MoE) model experts from permanent GPU residency, allowing dynamic adaptation to available memory. By combining PagedTensor, a bandwidth‑balanced memory hierarchy, and a budget‑aware residency planner, FluxMoE streams expert weights on demand while keeping computations on GPUs. Experiments on GLM‑4.5 and Mixtral‑8×7B‑Instruct show significant throughput gains and reduced time‑per‑output‑token compared to existing inference engines, without compromising model quality.
By Qingxiu Liu, Yongchao He, Runhan Jiang, Zion Wang, Bohan Zhao, Mi Zhang, Patrick P. C. Lee
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, which can suppress uncertainty and exploratory behavior, NSD generates a question‑specific negative condition (e.g., a careless reasoner) and uses a dynamic gating mechanism to target only reasoning‑critical tokens for penalization. This approach preserves foundational language capabilities while consistently outperforming OPSD and other label‑free self‑bootstrapping reinforcement learning baselines.
By Rongcan Pei, Zhepei Wei, Shuyao Xu, Xinyu Zhu, Wei-Lin Chen, Yu Meng
SparseDitto is an agentic sparse compilation framework that jointly synthesizes representation, execution schedule, and hardware mapping for sparse matrix computations on GPUs. It uses structural analysis, a learned template-ranking prior, and LLM-guided lowering to generate CUDA code, with target-GPU profiling refining the plan. The framework supports multiple operators such as SpMV, SpMM, and SpGEMM, adapts to different hardware, and achieves significant speedups over cuSPARSE, including up to 146.61× on certain matrices and 3.39× acceleration for full-batch GCN training.
By Shiyang Li, Guangyan Sun, Jinwei Tang, Yanzhi Wang, Mingyi Hong, Caiwen Ding
SearchAtlas is a framework that transforms raw search trajectories of large language model (LLM) agents into structured evidential query graphs, where edges capture how evidence is propagated from queries to the final answer. The automated parsing pipeline achieves a mean edge F1 of 86.0% against human-annotated graphs and remains consistent across repeated runs. Using SearchAtlas, the authors analyze five search agents on three benchmarks, uncovering systematic differences in search scale and evidence aggregation, and revealing process failures such as fragmented answer support, unmet question constraints, and unverified parametric knowledge that correlate strongly with incorrect answers.
By Jiacheng Sang, Mengyuan Li, Sanxing Chen, Yukun Huang, Yu Feng, Bhuwan Dhingra
The paper investigates K/V-cache interventions—transplanting a target-conditioned key/value trajectory into a source-persona generation—as a method for controlling persona in decoder-only language models. Experiments on Llama‑3.1‑8B across 13 configurations reveal that strong representation alignment (measured by V‑gap) does not guarantee behavioral persona expression, with only mid‑layer replacements achieving both alignment and lexical diversity. Position perturbations uniformly suppress persona expression, highlighting that representation similarity alone is insufficient to predict downstream behavior.
By Yu Sun, Mengyin Lu, Cong Feng, Guangming Lu, Huimin Han
The study examines how large language models (LLMs) predict depression scores from language responses. In a "Mirror" setup, participants answered structured diagnostic interviews that the LLMs used to predict scores, yielding near-perfect predictions. In a "Non-Mirror" setup, participants gave life history interviews; the LLMs still achieved outstanding prediction accuracy, and both conditions correlated similarly with PHQ-9 scores, indicating that the Mirror advantage disappears when predicting an independent measure. Topic modeling showed different depression themes across interview types, suggesting Mirror evaluations are more about reliability than validity and that Non-Mirror approaches may enhance clinical relevance.
By Tong Li, Rasiq Hussain, Mehak Gupta, Joshua R. Oltmanns
The paper introduces the first benchmark for evaluating confidence estimation in large language models during multi‑turn medical consultations, combining three types of medical data and an information sufficiency gradient to capture how confidence and correctness evolve as evidence accumulates. Experiments with 27 methods reveal that token‑level and consistency‑level confidence approaches are limited by medical data, and that medical reasoning must be judged on both diagnostic accuracy and information completeness. Building on these findings, the authors propose MedConf, a retrieval‑augmented, linguistically grounded self‑assessment framework that aligns patient information with supporting, missing, and contradictory relations, producing interpretable confidence estimates that outperform existing methods across multiple datasets and LLMs.
By Zhiyao Ren, Yibing Zhan, Siyuan Liang, Guozheng Ma, Baosheng Yu, Dacheng Tao
The paper introduces Inverse Turing Bench, a benchmark designed to assess how well language models can distinguish between human-only and human-AI dialogues in multi-turn text. It provides paired dialogue transcripts and evaluates models on correctly identifying the type of conversation. Preliminary results show GPTZero, Claude Opus-4.6, and GPT-5.5 achieving the highest accuracies of 89.41%, 77.92%, and 75.94% respectively, highlighting both the strengths and limitations of statistical versus semantic detection approaches.
By William Hager, Ishika Rathi, Masum Hasan, Cameron Jones
The paper investigates whether language models can identify sentences from their training data by using exact duplication counts from publicly released corpora for two model families, OLMo‑2 and Pythia. It finds that for typical duplication levels, models show only a weak trace of exposure, with a rank correlation near –0.08, and that strong signals only appear when a sentence appears roughly a thousand times, at which point fame rather than memory dominates. The study also demonstrates that common membership tests can be misleading, as changing a single word does not alter the model’s preference, and that controlling for register can significantly improve detector performance.
By Arman Nik Khah
The paper systematically studies how anonymizing input data affects large language models (LLMs). Five prominent LLMs were evaluated on eleven benchmarks, comparing performance on original versus pseudonymized inputs. Results show that anonymization generally degrades performance, with larger drops for more capable models and task-dependent effects; reversible anonymization preserves entity uniqueness better than irreversible redaction, and prompting about anonymization offers no benefit.
By Tobias Deu{\ss}er, Max Hahnb\"uck, Lorenz Sparrenberg, Tobias Uelwer, Christian Bauckhage, Rafet Sifa
The paper evaluates six automated jailbreak evaluators—HarmBench, JailbreakBench, JailbreakRadar, StrongReject, JADES, and JailMeter—using human-labeled data from JailbreakQR and JailMeter-Eva. It measures each evaluator’s agreement with human judgments, error types, and consistency across attack families, controlling for model-specific variation by using a shared LLM judge where needed. The study finds that JADES performs best overall, with HarmBench and StrongReject also showing strong performance.
By Yujie Mu
MHE-Former introduces a Transformer-based multi-hypothesis framework for monocular 3D hand and body mesh recovery, addressing occlusion and ambiguity by maximizing entropy during exploration. The method generates diverse, plausible mesh hypotheses efficiently and includes a context-aware hypothesis selection process that leverages visual language models and natural language intent. Experiments show state‑of‑the‑art accuracy and diversity, and a user preference study confirms the practicality of the selection mechanism.
By Boshu Jia, Rongyu Chen, Linlin Yang, Zihao Liu, Yingjie Chen, Zhongqun Zhang, Zhulin Tao, Shaohui Lin, Xiaoyu Wu, Libiao Jin, Baochang Zhang, Angela Yao