The paper introduces SESSE, a training‑free framework that breaks down LLM‑as‑judge evaluations into five steps—Sketch, Expand, Sort, Summarize, Evaluate—by mining sub‑questions from the judge’s own error cases. It requires no oracle responses, task‑specific rubrics, or fine‑tuning, yet on RewardBench it matches the performance of chain‑of‑thought baselines and rivals a fine‑tuned specialist (RISE‑Judge‑32B). SESSE provides per‑criterion vote evidence, offering an interpretable audit trail that can diagnose label ambiguity and judge failure modes that a single holistic output token cannot reveal.
By Dae Lee, Mihai Delgeanu, Adel Youssef
Gradient Mirage is a defense for large language model split learning that disrupts the assumption that exposed gradients faithfully reflect the client’s full-label training objective, thereby thwarting gradient matching attacks. It introduces inconsistencies across objective, direction, and scale by using selective autoregressive supervision, scale blinding, and directional privatization with a von Mises-Fisher mechanism. The method preserves utility through dual-track backpropagation and bottom-gradient recovery, achieving a stronger privacy‑utility trade‑off in experiments.
By Shiyu Miao, Yunlong Mao, Zirui Huang, Liang Yao, Tianshuo Zheng, Yanhui Gu, Fan Liu, Sheng Zhong
The paper presents the first systematic evaluation of uncertainty quantification (UQ) methods applied to a foundation model for semantic segmentation. By fine‑tuning a lightweight DPT decoder on the pretrained SAM2 encoder, the authors benchmark four UQ approaches—Monte Carlo Dropout, Deep Sub‑Ensemble, Test‑Time Augmentation, and Evidential Deep Learning—across Cityscapes, NYUv2, and two out‑of‑domain settings, comparing segmentation accuracy, calibration, uncertainty quality, and inference time. The results reveal clear trade‑offs between predictive performance, reliability, and computational cost, underscoring both the promise and current limitations of uncertainty‑aware foundation models for real‑world deployment.
By Steven Landgraf, Joceline Hinz, Markus Ulrich
The paper examines how multilingual medical adaptation affects the internal representations of Whisper ASR models. By comparing various fine‑tuning strategies—zero‑shot decoding, English‑only, German‑only, two‑stage EN→EN+DE, and direct EN+DE fine‑tuning—it shows that fine‑tuning significantly improves performance, with the best model varying by setting. Layer‑wise encoder analysis reveals that English medical fine‑tuning drives the main representation shift, while multilingual continuation largely preserves the adapted space, and that domain and language signals remain recoverable across layers.
By Souranil Kahali, Rituparna Bose, Abner Hernandez, Tomas Arias-Vergara, Andreas Maier, Ning Ma, Paula Andrea Perez-Toro
Brain2Qwerty v2 is a model that decodes natural sentences from real‑time magnetoencephalography (MEG) recordings, achieving an average word error rate of 39% across 22,000 sentences typed by nine subjects. The model uses character, word, and sentence‑level representations and shows that decoding accuracy improves log‑linearly with more data, narrowing the gap to intracranial brain‑computer interfaces. AI contributes by replacing hand‑crafted event detection with deep learning, fine‑tuning large language models for semantic extraction, and employing AI agents to refine the decoding pipeline through automated code development.
By Mingfang Zhang, Jarod L\'evy, Cedric Rommel, J\'er\'emy Rapin, Corentin Bel, Julie Bonnaire, Daniel Nieto, Pierre Bourdillon, Svetlana Pinet, St\'ephane d'Ascoli, Thomas Moreau, Jean-R\'emi King
Aslema is a system developed for the NADI 2026 Shared Task 5, which involves intent recognition and slot filling. The team evaluated four omni LLMs in a zero‑shot setting and found that fine‑tuned models consistently outperform zero‑shot inference. They further improved performance by augmenting data with culturally grounded Tunisian Derja utterances generated by an LLM and synthetic speech produced via voice cloning, achieving top‑ranked results on the official test set.
By Tajwaar Shafiq, Hunzalah Hassan Bhatti, Shammur Absar Chowdhury, Firoj Alam
The OpenAI Privacy Filter is a 1.5‑billion‑parameter, bidirectional token‑classification model that detects and redacts personally identifiable information and secrets in unstructured text. It is built from an autoregressive checkpoint, converted into a banded‑attention classifier, and uses a constrained Viterbi decoder to produce coherent spans across eight privacy categories in a single forward pass. The model supports configurable precision‑recall tradeoffs, a 128,000‑token context window, and is designed for efficient local deployment and domain‑specific fine‑tuning as a data‑minimization component within layered privacy workflows.
By Charles de Bourcy, Sahra Ghalebikesabi, Avi Schwarzschild, Alex Gorbachev, Mihai Maruseac, Annie Chu, Vol Kyrylov, Tong Mu, Ally Bennett, Andy Nguyen, Casey Meehan, Jessica Gan Lee, Shane Bauer, Harold Nguyen, Rodolpho Eckhardt, Yuqi Liu, Charlie Oxborough, Marco Rougeth, Omar Chedid, Caio Costa, Yash Parikh, Yao Li, Congzheng Song, Om Thakkar, Vinnie Monaco
The paper investigates a problem in guided protein language models where strong guidance causes the model’s internal representations to collapse onto a region indistinguishable from random amino‑acid input, leading to low‑complexity sequences that still score well on the targeted property. The authors identify this off‑manifold collapse as a detectable signature and propose a post‑hoc filtering technique—Mahalanobis filtering—that removes atypical candidates based on a density prior over natural activations. This simple, training‑free step improves both property scores and structural plausibility across different guidance methods without altering the generator.
By Shuibai Zhang, Xinchi Liu, Fred Zhangzhi Peng, Zhihan Yang, Shutong Wu, Yingzi Ma, Jiawei Zhang
The paper presents a method for extracting key information from OCR‑digitized clinical reports, addressing challenges posed by heterogeneous documents and noisy OCR output. It introduces an open key space that is iteratively mined, normalized, clustered, and verified to build a canonical key inventory, and defines key coverage as a metric for inventory completeness. Experiments on reports from over 20 hospitals using a 0.2B BERT model show that performance improves steadily with key coverage, achieving high F1 scores when the top 90 keys are covered and outperforming a fine‑tuned Qwen3‑0.6B baseline.
By Yu Wang, Yingyun Li, Ying Qin, Haiyang Qian
The report details the winning approach for the AES AIMLA 2025 Challenge on querying sound effects by vocal imitation. It explores two fine‑tuning strategies: contrastive learning with a frozen, pretrained CED encoder, and joint contrastive‑triplet learning using a MobileNetV3 encoder with semi‑hard negatives. Updated post‑challenge, the document includes additional released details.
By Aditya Bhattacharjee, Christos Plachouras, Sungkyun Chang, Emmanouil Benetos
The paper introduces Top‑K prompting as a training and inference strategy to better capture the diverse, plausible predictions inherent in single‑step retrosynthesis. Using an ultra‑large dataset (CREED‑CCV‑2+USPTO‑XL) of ~45.6 million verified reactions, the authors train the Chemistry Constraint‑Consistent Language Model (C3LM). With fine‑tuning that incorporates ChemCensor‑based and novelty‑oriented rewards, C3LM achieves state‑of‑the‑art performance on the OOD URSA‑expert‑2026 benchmark and demonstrates complementary reaction space exploration compared to conventional models, suggesting benefits for ensemble‑based retrosynthesis systems.
By Bogdan Zagribelnyy, Ivan Ilin, Nikita Bondarev, Maksim Kuznetsov, Mathieu Reymond, Vladimir Aladinskiy, Alex Aliper, Alex Zhavoronkov
The paper investigates the use of looped language models for compositional tool calling, where models must coordinate multiple API calls and maintain state across interactions. Experiments on API-Bank, BFCL, and NESTful show that recurrent computation generally improves compositional and dependency-aware tool use, with accuracy increasing as recurrent depth grows. Adaptive inference offers a better compute‑performance trade‑off by allocating extra computation only when necessary.
By Andrei Cristian Popescu, Haitz S\'aez de Oc\'ariz Borde, Pietro Li\`o
MedStruct‑S is a benchmark for semi‑structured information extraction from OCR‑derived clinical reports, covering key discovery, key‑conditioned QA, and end‑to‑end key‑value extraction. It contains 3,582 annotated real‑world report pages and evaluates models under unknown keys and OCR noise. Experiments show encoder‑only models excel at non‑null key‑conditioned QA, while fine‑tuned decoder‑only models achieve the strongest overall performance across model sizes.
By Yingyun Li, Yu Wang, Haiyang Qian
arXiv:2608. 18237v1 Announce Type: cross Abstract: Estimating the difference of two Stein's score functions is a fundamental problem in generative modeling.
By Chenghan Xie, Jose Blanchet, Renyuan Xu
FACET is a framework for synthesizing terminal tasks that preserves source intent and ensures cross‑artifact consistency. It reconstructs agent skills into coherent scenarios, repairs the execution environment, and uses the resulting container state as shared grounding for the instruction, solution, and verifier. By validating and repairing artifacts through execution, FACET produces complex tasks with dense executable checks and data‑efficient supervision, improving performance on Terminal‑Bench 2.1.
By Kou Shi, Zun Wang, Qisheng Su, Shiting Huang, Ziao Zhang, Zhen Fang, Qingnan Ren, Jin Liu, Yu Zeng, Yiming Zhao, Lin Chen, Zehui Chen, Feng Zhao
The paper presents a local traditional OCR pipeline that can be iteratively fine‑tuned on both layout and appearance levels of complex historical Sanskrit manuscripts. By adapting to the specific manuscript distribution, the pipeline reduces human annotation effort and improves transcription accuracy across subsequent pages. The authors apply this method to three manuscripts, release a dataset with detailed layout and Unicode annotations in PAGE‑XML format, and benchmark the pipeline against leading multimodal large language models.
By Kartik Chincholikar, Kaushik Gopalan, Mihir Hasabnis
The paper evaluates whether audio‑language models can use multimodal clinical context to improve dysarthric speech recognition. Using a benchmark built on the Speech Accessibility Project dataset, the authors test diagnosis labels, clinician ratings, and detailed clinical descriptions as prompts for nine models. They find that these prompts yield negligible or negative effects on word error rate, though fine‑tuning with LoRA and mixed prompt formats reduces WER by 52% and benefits certain subgroups such as Down syndrome and mild‑severity speakers.
By Pehu\'en Moure, Niclas Pokel, Bilal Bounajma, Yingqiang Gao, Roman Boehringer, Longbiao Cheng, Shih-Chii Liu
The paper evaluates large language models for hate‑speech detection in Roman Urdu, a low‑resource language with informal spelling variations. Using the Parameter‑Efficient Fine‑Tuning technique Low‑Rank Adaptation (LoRA), the authors fine‑tune models such as Mistral, LLaMA, Falcon, and multilingual BERT on the 72,000‑comment PURUTT dataset. While zero‑shot performance yields an F1 of 0.56, fine‑tuning a small fraction of parameters boosts F1 scores above 0.93, demonstrating that PEFT offers both high accuracy and computational efficiency for low‑resource language tasks.
By Toneema Zubair, Muhammad Junaid Asif, Faisal Kamiran, Hafiz Hassan Saeed, Rana Fayyaz Ahmad
The paper reviews how large language models are applied in mental health, covering areas such as social media analysis, clinical conversational agents, therapy support tools, prompt engineering, and multimodal learning. It synthesizes interdisciplinary studies that use social media posts, electronic medical records, and multimodal inputs to detect depression, assess suicide risk, provide personalized therapy, and generate psychoeducational content. The review also discusses advances in model interpretability, annotation strategies, multimodal fusion techniques, and highlights ethical, sociotechnical, and regulatory challenges while proposing frameworks for safe, equitable, and accountable deployment.
By Yisong Chen, Yifan Gao, Sijing Yu, Chuqing Zhao, Yang Lu
FiLoRA is an instruction‑conditioned, parameter‑efficient adaptation framework that allows controllable modulation of feature reliance in multimodal foundation models. It decomposes adaptation into feature‑aligned low‑rank modules and uses instruction‑conditioned gating so that natural language instructions can act as computation‑level control signals over internal representations. Experiments on classification and generative multimodal tasks show that FiLoRA consistently and interpretable shifts feature reliance, amplifying or suppressing feature groups per instruction without changing task semantics.
By Hyunsuk Chung, Soyeon Caren Han, Seungyeon Ji, Jinwoo Kim, Eun-Jung Holden, Kyungreem Han