The REAP system tackles the AKBC Shared Task 2026, aiming to build knowledge bases from language models without fine‑tuning and within a 32‑B parameter budget. It uses structured chain‑of‑thought reasoning, relation‑specific queries, and a reasoning‑based empty‑set gate to elicit knowledge, then directly extracts it into valid JSON arrays. Evaluated on the test set with the Mistral‑Small‑24B‑Instruct‑2501 model, REAP achieves a macro‑F1 of 0.62, notably high scores on countryLandBordersCountry (0.95), companyTradesAtStockExchange (0.73), and hasArea (0.77).
By Thanh-Dan Bui, Thanh-Trung Do, Tuan-Phong Nguyen
The paper examines how Retrieval-Augmented Generation (RAG) and Named Entity Recognition (NER) affect the quality of lay summaries of radiology reports. Using a framework that extracts clinically relevant findings via NER and grounds them with RAG, the authors evaluate few‑shot and fine‑tuned versions of Qwen and BioBART. Results show that NER consistently improves readability and overall quality, RAG alone offers no benefit and can introduce hallucinations, and the best performance comes from fine‑tuned BioBART with NER.
By Egecan \c{C}elik Evgin, \.Ilknur Karadeniz, Olcay Taner Y{\i}ld{\i}z
TikZilla is a new approach to generating TikZ code from textual descriptions, built on a larger, higher‑quality dataset called DaTikZ‑V4 that includes LLM‑generated figure descriptions. The method uses a two‑stage pipeline: supervised fine‑tuning of small Qwen models (3B and 8B) followed by reinforcement learning with an image encoder that provides semantically faithful reward signals. Human evaluations show that TikZilla outperforms its base models by 1.5–2 points on a 5‑point scale, beats GPT‑4o by 0.5 points, and matches GPT‑5 in image‑based tests while remaining much smaller.
By Christian Greisinger, Steffen Eger
The study investigates cultural biases in large language models (LLMs) by testing their ability to perform author profiling—inferring singers’ gender and ethnicity—from song lyrics in a zero‑shot setting. Evaluating over 10,000 lyrics across several open‑source models, the authors find that most LLMs default toward North American ethnicity, while DeepSeek‑1.5B leans toward Asian ethnicity, and that Ministral‑8B exhibits the strongest ethnicity bias whereas Gemma‑12B is the most balanced. The paper introduces two fairness metrics, Modality Accuracy Divergence (MAD) and Recall Divergence (RD), to quantify these disparities and provides code and results publicly on GitHub and HuggingFace.
By Valentin Lafargue, Ariel Guerra-Adames, Emmanuelle Claeys, Elouan Vuichard, Jean-Michel Loubes
The paper proposes a method to adapt the Depth Anything V2 (DAV2) zero‑shot relative depth model for estimating lunar surface height. By fine‑tuning DAV2 with publicly available stereophotogrammetry‑derived DEM data, the authors achieve a significant performance boost over the unadapted zero‑shot model. This improved estimator can provide more accurate relative height information useful for hazard detection in future ESA lunar landings.
By Patrick Bauer, Marius Schwinning, Melanie Siegel, Andreas Weinmann, Hichem Snoussi
The paper introduces DynaTokens, a transformer-based token generator that produces fine‑tuning tokens on demand for continual VideoQA with multimodal large language models. It uses shared generation weights and meta‑learning‑inspired regularisers to reduce task interference and forgetting, connecting the objective to sharpness‑aware optimisation for flatter cross‑task minima. Experiments on standard continual VideoQA benchmarks show that DynaTokens achieves higher average accuracy, lower forgetting, better zero‑shot generalisation, and robust cross‑modal transfer in a new ImageQA→VideoQA protocol.
By Toan Nguyen, Yang Liu, Celso De Melo, Flora D. Salim
This study experimentally evaluates large language models (LLMs) for legal decision-making in Austrian and EU value‑added tax (VAT) law. It compares fine‑tuning and retrieval‑augmented generation (RAG) on textbook and real‑world tax consulting cases to identify optimal LLM configurations and assess legal‑reasoning abilities. Results show that properly configured LLMs can aid tax professionals with routine tasks and provide legally grounded justifications, though they are not yet ready for full automation due to challenges like hallucination and handling implicit client knowledge.
By Marina Luketina, Andrea Benkel, Christoph G. Schuetz
Freeze, Diffuse, Decode (FDD) is a diffusion-based framework that adapts pre‑trained transformer embeddings to downstream tasks while preserving their geometric structure. By propagating supervised signals along the intrinsic manifold of frozen embeddings, FDD produces low‑dimensional, predictive, and interpretable representations. In antimicrobial peptide design, these representations support property prediction, retrieval, and latent‑space interpolation.
By Pankhil Gawade, Adam Izdebski, Myriam Lizotte, Kevin R. Moon, Jake S. Rhodes, Guy Wolf, Ewa Szczurek
HEAT introduces a fine‑tuning method that treats the number of iterations used to approximate nonlinearities in fully homomorphic encryption (FHE) as learnable parameters, allowing them to co‑adapt with model weights. By optimizing iteration counts per nonlinearity, HEAT reduces the required iterations, bootstraps, and overall latency for encrypted GPT‑2 decoding while improving decode agreement. The approach achieves a 3.1× reduction in iterations, a 1.6× reduction in bootstraps, and a 1.4× speed‑up in end‑to‑end latency without changing the model architecture or requiring retraining from scratch.
By Alessandro Zirilli, Davide Marincione, Evgenios M. Kornaropoulos, Giuseppe Ateniese, Emanuele Rodol\`a
MM++ (Multilayer Mahalanobis++) is a post‑hoc, scale‑invariant framework for out‑of‑distribution detection that builds a joint feature space by selecting discriminative intermediate layers based on entropy density drops and fusing them with the final representation. It uses a Ledoit‑Wolf regularized tied covariance matrix to stabilize the space, allowing reliable distance estimation without requiring auxiliary OOD data, classifier fine‑tuning, or architectural changes. The method achieves robust performance across different architectures for both near‑ and far‑OOD scenarios.
By Rahim Hossain, Md Tawheedul Islam Bhuian, Md Farhan Shadiq, Kyoung-Don Kang
The paper introduces a scalable Kronecker-based approximation that captures cross-layer interactions without storing the full Fisher matrix, making Hessian analysis feasible for billion-parameter language models. It identifies consistent vulnerability patterns, notably that value projection layers are the most sensitive and exhibit strong cross-layer correlations across various model families. Experiments on quantization, sparsification, inter-layer corruption, and fine-tuning show that the approximation correlates strongly with performance degradation and recovery, providing a practical tool for identifying fragile components and guiding compression and optimization strategies.
By Viacheslav Yusupov, Daria Cherniuk, Evgeny Frolov
The paper introduces a method for creating query‑specific rubrics for DeepResearch‑style long‑form report generation by training rubric generators with reinforcement learning. It builds a dataset of queries annotated with human preferences, then uses a hybrid reward that includes preference consistency, format validity, and LLM‑based rubric evaluation. The learned rubrics outperform generic or manually constructed alternatives in distinguishing preferred reports and, when used as rewards, improve performance of both single‑agent and multi‑agent DeepResearch systems.
By Changze Lv, Jie Zhou, Wentao Zhao, Jingwen Xu, Shihan Dou, Zisu Huang, Muzhao Tian, Xiaohua Wang, Zhengkang Guo, Yang Liu, Pluto Zhou, Tao Gui, Le Tian, Xiao Zhou, Xiaoqing Zheng, Xuanjing Huang, Jie Zhou
The paper introduces GPTBIAS, a framework that uses powerful large language models like GPT‑4 to evaluate bias in other LLMs. It employs specially crafted prompts called Bias Attack Instructions to probe for bias and outputs a bias score along with detailed information such as bias types, affected demographics, keywords, reasons, and improvement suggestions. Extensive experiments demonstrate the framework’s effectiveness and usability.
By Jiaxu Zhao, Meng Fang, Shirui Pan, Wenpeng Yin, Mykola Pechenizkiy
DeepAffinity is a model designed to predict eCommerce users’ future preferences for product aspects such as brand, size, and color, treating this as a temporal prediction problem. It uses small language models with structured prompts and specialized prediction heads, outperforming standard generative fine‑tuning and general‑purpose open‑source LLMs that lack task‑specific tuning. The approach improves recommendation quality on a large multinational eCommerce platform.
By Yotam Eshel, Guy Hadad, Guy Feigenblat, Yuri M. Brovman, Matt Gearhart, Bracha Shapira
SpeakPay is a voice‑first digital wallet designed to make mobile payment apps in Nepal accessible to visually impaired users. The paper introduces NepFinSpeech‑403, a 403‑utterance Nepali financial voice command dataset, and demonstrates that fine‑tuning Whisper large‑v2 with LoRA reduces the Word Error Rate from 129.95% to 42.58% and improves Devanagari numeral recognition from 0.0% to 73.9%. Domain adaptation also boosts the Transaction Success Rate from 1.67% to 33.33%, with as few as 100 domain‑specific utterances halving the zero‑shot WER.
By Biraj Subedi
The paper reports a specialized training pipeline for large language models to excel in competitive programming, combining problem curation, synthetic reasoning traces, supervised fine‑tuning, and reinforcement learning. Using 22,000 curated problems, the authors train two models—Nemotron‑3‑Nano‑CC (30B) and Nemotron‑3‑Ultra‑CC (550B)—and introduce GenCorrect, a test‑time refinement strategy. On the IOI 2025 benchmark, Nano‑CC scores 468 points with GenCorrect, surpassing the gold‑medal threshold, while Ultra‑CC reaches 502; in IOI 2026, a competition‑specific Ultra‑CC system scores 535.4, exceeding both the gold threshold and the top human score of 498.27, marking the first AI system to outscore the highest‑scoring human contestant on an IOI problem set.
By Aleksander Ficek, Sean Narenthiran, Mehrzad Samadi, Somshubra Majumdar, Boris Ginsburg
IDEA is a training‑free, input‑dependent steering method for large language models that matches activations to cluster‑specific directions aligned with a target concept. It clusters positive and negative activation supports per attention head, solves an optimal‑matching problem to create a pool of cluster‑conditional directions, and selects the best match for each input at inference time. This approach preserves the input’s original representation while improving the truth × info rate on TruthfulQA by an average of 9.9% (up to 23.5%) over input‑independent baselines.
By Zheng Wang, Muchen Li, Renjie Liao, Yan Leng
RouteGraph-Mona introduces a confusion‑aware routing fine‑tuning approach for mineral image classification, building on the Mona visual adapter. It replaces Mona’s static multi‑scale aggregation with sample‑adaptive routing, creating a compact routing space that reflects each image’s scale preferences. The method regularizes routing signatures using class‑wise route anchors and confusion‑weighted margins, encouraging consistent routing patterns and greater separation between visually similar mineral categories. Experiments on three public mineral image datasets with two visual backbones show that RouteGraph‑Mona consistently outperforms Mona and remains competitive with other fine‑tuning methods and classification baselines.
By Jierui Li, Zhiyuan Qi, Hao Zhu, Yufan Liu, Jixian Liu, Shaojie Jiang, Jianda Wang, Yaqi Liu, Xiaotong Li, Wei Wang
The paper introduces a black-box membership inference attack framework tailored for fine-tuned text-to-speech models, addressing challenges in query generation and representation engineering. It evaluates five query types, finding recitation queries most effective, and uses multi-level speech embeddings with temporal alignment for fine-grained comparison. Experiments on CosyVoice2, F5-TTS, and XTTS-v2 trained on VCTK and British Dialect datasets show high privacy leakage, with speaker-level AUC above 0.80 and record-level AUC between 0.80 and 0.90.
By Kunlin Cai, Kaiyuan Zhang, Zihang Xiang, Jinghuai Zhang, Abeer Alwan, Fnu Suya, Yuan Tian
The paper introduces Self-Geometry, a plug‑and‑play test‑time adaptation framework that enforces explicit multi‑view geometric constraints on Vision Foundation Models (VFMs) using 2D pixel correspondences as pseudo ground truth. It combines Geometric Disentanglement Optimization—mixing Multi‑View and Epipolar Consistency losses with Gradient Disentanglement—to avoid gradient conflicts, a Frame Angular‑Neighbor sampler based on SO(3) geodesic distances to select informative views, and a Lightweight TTA module that adapts VFMs via LoRA. Experiments on six VFMs and four benchmarks (7Scenes, ETH3D, ScanNet++, HiRoom) show consistent improvements in pose and geometry estimation.
By Seokhyun Youn, Dahyeon Kye, Sung-Ho Bae, Jihyong Oh