AsyncFlow is an asynchronous streaming reinforcement learning framework designed to improve the post‑training phase of large language models. It introduces a distributed data storage and transfer module that enables panoramic data management and fine‑grained scheduling, allowing automated pipeline overlapping and dynamic load balancing. The framework also employs an asynchronous producer‑consumer workflow to reduce computational idleness by deferring parameter updates within staleness thresholds, and it is architecturally decoupled from training and inference engines, providing modular, customizable user interfaces. Experiments show an average throughput improvement of 1.59× over the state‑of‑the‑art baseline.
By Zhenyu Han, Ansheng You, Haibo Wang, Kui Luo, Guang Yang, Wenqi Shi, Menglong Chen, Sicheng Zhang, Zeshun Lan, Chunshi Deng, Huazhong Ji, Wenjie Liu, Yu Huang, Yixiang Zhang, Chenyi Pan, Jing Wang, Xin Huang, Chunsheng Li, Jianping Wu
The paper investigates whether reinforcement‑learning post‑training of code‑generating large language models can be done entirely offline using existing datasets, avoiding costly online code generation and GPU‑CPU communication. Experiments show that a few hours of offline RL can substantially boost zero‑shot code generation performance across models from 0.5 B to 7 B parameters, though the magnitude of improvement differs by model family.
By Abhinav Anand, Sanjana Reddy Pachika, Shweta Verma, Mira Mezini
The paper investigates on-device language models (ODLMs) for predicting stress in a mobile health context, focusing on privacy-preserving, cloud-independent inference. Using zero‑shot prompting, the authors evaluate ODLMs across multimodal data—objective sensor features and subjective self‑reports—measuring predictive accuracy, latency, and throughput. Results indicate that sensor features slightly outperform self‑reports, and that lightweight sub‑2B models deliver low latency with predictable resource usage, underscoring both the potential and practical limits of ODLMs for mobile mental health.
By Ibukunoluwa Soyebo, Alyssa Donawa, Rodrigo Aguilar Barrios, Brice Patchou, Corey E. Baker
The paper investigates how large language models (LLMs) used as judges in absolute scoring tasks exhibit systematic biases that compromise reliability. It shows that a judge’s task accuracy strongly predicts both its judging accuracy and its directional bias, yet more capable examinee models consistently receive more lenient judgments. To mitigate these biases, the authors propose a calibrated weighted majority voting (WMV) ensemble that estimates judges’ error rates from inter-judge agreement patterns, achieving near-oracle performance without labeled data and improving both accuracy and fairness.
By Gemma Zhang, Prachi Badarayani, Asmi Kumar, Sadid Hasan, Sulaiman Vesal
The paper introduces CLAW, a method that uses a hypernetwork to generate low‑rank adapters for world models during test time, enabling efficient adaptation to new environments with only a few episodes of interaction. By jointly pretraining the hypernetwork and base model on simulated adaptations, CLAW balances computational efficiency and expressivity, outperforming both in‑context learning and gradient‑based adaptation in locomotion and manipulation tasks. The approach also mitigates overfitting in data‑scarce regimes and demonstrates that the benefit stems from expressive adapters rather than context conditioning.
By Fernando Palafox, David Fridovich-Keil
The paper revisits the distortion problem in Reinforcement Learning from Human Feedback (RLHF) for aligning large language models. It shows that exponential degradation in user utility, previously attributed to RLHF, actually stems from a mismatch between the preference data distribution and the KL reference policy. By deriving tight bounds across different KL regularization regimes, the authors demonstrate that when the two distributions match, RLHF achieves near‑optimal distortion, and they recommend using on‑policy data or pre‑fine‑tuning on data close to the true preference distribution.
By Kazusato Oko, Annie Ulichney, Nika Haghtalab, Han Bao
The paper introduces a feedforward graph architecture that uses several frozen large language models as computational nodes connected through a shared continuous latent space via learned linear projections. By jointly optimizing projection matrices through backpropagation, the system combines the representations of three small frozen models with two larger ones, culminating in a lightweight cross‑attention output node. With only 17.6 M trainable parameters, the architecture attains state‑of‑the‑art results on ARC‑Challenge, OpenBookQA, and MMLU, surpassing both individual constituent models and parameter‑matched learned classifiers.
By Marcus Armstrong, Navid Ayoobi, Arjun Mukherjee
The paper introduces a general machine learning framework called the localization method, built on localization kernels and local means that underpin self‑attention. It formally defines the framework through a local‑ized model and a localization trick, and systematically connects it to a wide array of existing models such as kernel methods, MeanShift, Hopfield networks, LLE, fuzzy inference, and denoising autoencoders. The authors also extend the framework with adaptive kernels, hierarchical local models, and non‑local models, showing that the Transformer can be constructed using hierarchical local models, thereby unifying and generalizing state‑of‑the‑art architectures.
By Congwei Song
Dynamic Expert Quantization (DynaExq) is a runtime-aware mixed-precision serving system designed for single‑GPU Mixture‑of‑Experts (MoE) inference under a hard high‑bandwidth memory (HBM) envelope. It treats the problem as an online, budget‑constrained precision allocation task, keeping the most frequently used experts at higher precision while relegating the rest to low‑precision fallbacks. By estimating expert hotness from router traces and asynchronously promoting or demoting experts, DynaExq maintains a fully materialized expert set during the forward pass, improving accuracy and throughput compared to static post‑training quantization and offloading/prefetch baselines.
whyItMatters":"DynaExq enables efficient deployment of large MoE models on memory‑limited GPUs by dynamically allocating precision based on runtime expert usage, thereby reducing memory footprint and latency while boosting accuracy and throughput."
By Kexin Chu, Dawei Xiang, Zixu Shen, Yiwei Yang, Zecheng Liu, Wei Zhang
The study introduces a scalable acoustic‑masking method to quantify how much each consonant contributes to word intelligibility. By silencing individual consonants in isolated words and measuring misrecognition rates with three ASR models, the authors define a mask‑induced misrecognition rate (MMR). Across English, Spanish, German, and Czech, MMR negatively correlates with phoneme frequency and positively with functional load, revealing that consonant importance varies by language.
By Eunjung Yeo, Kwanghee Choi, Krupaben Kothadia, Visar Berisha, Julie M. Liss, David R. Mortensen, David Harwath
The paper introduces an open, modular AI framework that automatically detects and structures evidence of social tipping points in climate literature at the passage level. It integrates a DistilBERT boundary splitter, an iteratively augmented RoBERTa classifier, a Mistral 7B rewrite model, a LLaMA 3.2 3B rating model, and a Milvus vector store, all accessible via a Streamlit interface. Evaluation on a GPT‑4.1‑labelled benchmark and expert‑reviewed set shows the splitter outperforms competitors and the RoBERTa detector achieves high accuracy and agreement.
By Kavindu Perera, Mohammad Abaeiani, Ekaterina Gilman, Lauri Loven, Mourad Oussalah, Tassos Kanellos, Beatrice Gobbo, Dante Adami, Nicol\`o Ferriani, Maximiliano Romero, Pierre Rossel, Marc Bonazountas, Christina Deligianni, Nikos Xyderis, Artur Bogucki, Lampros Argyriou, Prasasthy Balasubramanian
The House with a Million Windows (HWAMW) is an LLM-based interactive fiction system that lets users narrate a story and then view it through a series of AI-generated "windows" that reframe the narrative in various literary styles. The system is grounded in the psychological restorying intervention, aiming to deepen users' exploration of meaning in their personal stories. Empirical results indicate that HWAMW enhances users' sense of narrative identity, and expert reviews suggest it achieves this by facilitating restorying rather than simply generating new content.
By Cody Kommers, Sarah G Immel, Drew Hemment, Mina Lee
The paper introduces Tasks over Application Manuals (TAM), a benchmark designed to test long‑horizon procedural reasoning in large language models. TAM uses real‑world tasks from ICD‑10‑CM clinical coding and U.S. federal sentencing, requiring models to follow extensive, rule‑based manuals and perform interdependent steps to produce exact answers. Experiments with GPT‑5 and various prompting strategies show very low exact‑match accuracy—1% for coding and 15.5% for sentencing—highlighting a gap between current benchmarks and the ability to reliably follow complex procedures.
By Utkarsh Soni, Syed Shariyar Murtaza, Yifan Nie, Sachin Chandrasekhar, Eugene Wen
The study investigates how deictic ambiguity—specifically the shifting reference of expressions like "previous"—affects Draft‑Verify‑Revise pipelines that use multiple large language models (LLMs). Using a synthetic dataset of 10 base examples and 21 reasoning‑effort configurations, six LLMs were evaluated for their ability to correctly resolve the ambiguous expression across the draft, verify, and revise stages. Results show wide variance in balanced accuracy, with GPT‑5.2 improving from 0.156 to 0.942 with increased reasoning effort, while Gemini 3 Pro consistently achieved high accuracy above 0.94 even at low reasoning effort, and meta‑evaluators often relied on surface cues when making errors.
By Obinna I. Ekekezie
The paper introduces UrduFactBench and UrduFactQA, two hand‑annotated benchmarks for claim verification and factual consistency evaluation in Urdu, created through a multi‑stage annotation process with native speakers. It also presents UrduFactCheck, a modular fact‑checking framework that uses both monolingual and translation‑based evidence retrieval to address the scarcity of high‑quality Urdu evidence. Experiments on twelve LLMs show that translation‑augmented pipelines outperform monolingual ones, highlighting ongoing challenges for open‑source models in Urdu.
By Sarfraz Ahmad, Hasan Iqbal, Momina Ahsan, Numaan Naeem, Muhammad Ahsan Riaz Khan, Arham Riaz, Muhammad Arslan Manzoor, Yuxia Wang, Preslav Nakov
The study examines whether Dutch language models exhibit coherence‑illusion effects similar to human readers, using texts that refer back to earlier context with words like ‘again’ and ‘too’. Surprisal at the critical word aligns with human acceptability and eye‑tracking data, showing that models are more surprised by incoherent continuations unless a matching distractor is present. Attention entropy and an energy metric from associative‑memory literature reveal heads that behave differently under coherence versus incoherence, and ablating these heads demonstrates transfer effects across experiments, indicating a shared underlying mechanism.
By Ece Takmaz, Nitin Kumar, Li Kloostra, Jakub Dotlacil
The paper presents an automated pipeline that uses unmanned aircraft systems (UAS) multispectral imagery and a Vision Transformer to detect wheat streak mosaic virus (WSMV) at the plant level. While the model achieved 89% accuracy on over 6,500 test patches using treatment-based labels, ELISA-based ground truth revealed significant label noise, indicating that the high accuracy was largely due to label bias rather than true disease detection. When evaluated against more reliable row‑level symptom severity and plant‑level ELISA labels, both deep learning and classical machine learning models showed limited generalization and weak separability between infected and mock‑inoculated plants, underscoring the importance of biologically grounded labels and realistic data conditions for UAS‑based disease detection.
By Dewi Endah Kharismawati, Sandeep Dhakal, Courtney E. McCusker, Jennifer R. Wilson, Erik W. Ohlson, Sami Khanal
ChitraMiti introduces a synthetic benchmark of 12,874 Bengali planar geometry problems with structured 15‑attribute descriptions, alongside a complementary set of 500 textbook diagrams. Using a three‑phase protocol that tests diagram‑only, diagram‑plus‑description, and description‑only inputs, the study finds that description‑only performance matches diagram‑plus‑description performance across several VLMs, yet models still struggle with cross‑modal verification. Fine‑tuning on ChitraMiti improves results on both datasets, though a gap remains compared to the best zero‑shot model.
By Khan Raiyan Ibne Reza, Sanjana Aktar Maria, Sumaiya Tabassum Nimi, Md Adnan Arefeen
DementiaCare-Bench is a new video benchmark designed to evaluate video-language models (VLMs) on understanding behavioral and psychological symptoms of dementia (BPSD). It contains 56 caregiver training videos, 94 clips across nine BPSD categories, and 2023 questions that are grounded in transcript spans and labeled by their visual demand. The benchmark reveals that most VLMs perform well on questions answerable by language alone but struggle with those requiring ordered frames, and a lightweight LoRA fine‑tune (DemCare‑VLM) can improve video dependence.
By Afrouz Sheikholeslami, Yuankai Qi, Xuyun Zhang, Luping Zhou, Amin Beheshti, Quan Z. Sheng, Ming-Hsuan Yang
AnchorVLN is an open‑vocabulary vision‑language navigation system that separates semantic proposals from geometric metrics. It uses a VLM to generate semantics while a geometry module supplies reliable metric quantities such as range and bearing, all within a Model Context Protocol server. The system achieves 64.4% on instruction following and improves object‑reference accuracy, reducing median center error from 3.37 m to 2.48 m.
By Long Giang Vu, Chengkai Yao, Yuxin Liu, FNU Aryan, Rajath Chandrashekar Aralikatti