The paper proposes a two‑stage framework, SL+LHF, that first learns low‑dimensional representations from noisy labeled data and then refines model alignment using human comparison feedback via a probabilistic bisection approach. It introduces the label‑noise‑to‑comparison‑accuracy (LNCA) ratio to theoretically identify when this framework outperforms pure supervised learning, showing that trading labels for comparisons reduces sample complexity when labels are scarce. Experiments on a high‑dimensional crowdfunding prediction task and an Amazon Mechanical Turk study confirm that incorporating human or large language model evaluators improves accuracy under a fixed query budget.
By Junyu Cao, Mohsen Bayati
arXiv:2602. 06699v2 Announce Type: replace-cross Abstract: We propose a variational quantum implementation of self-attention (QSA)-the core operation in transformers and large language models-which predicts future elements of a sequence by forming overlap-weighted combinations of past data.
By Alessio Pecilli, Matteo Rosati
The paper introduces MultiwayPAM, a tensor clustering technique designed for analyzing LLM-as-a-Judge score tensors that capture evaluations across questions, answerers, and evaluators. By simultaneously estimating cluster memberships and medoids for each mode, the method reveals the structure of bias inherent in large language model evaluators. Experiments on two practical datasets demonstrate the effectiveness of MultiwayPAM in uncovering these bias patterns.
By Chihiro Watanabe, Jingyu Sun
BanglaTurn is a new corpus of 35,374 Bangla podcast speech samples, each 3 to 15 seconds long, labeled for end‑of‑turn detection through speaker diarization, an LLM pass, and human verification. A Whisper‑based model with task‑specific classification heads achieves 84.33 % accuracy on a balanced test set, outperforming the Smart‑Turn v3 baseline (69.28 %) and reducing the false‑negative rate from 51.57 % to 7.55 %, though with a higher false‑positive rate. The study also details the contributions of encoder‑layer fine‑tuning, multi‑scale pooling, INT8 quantization, and reports inference latency of 165–191 ms on CPU.
By Mizbaul Haque Maruf
The paper investigates how large language models (LLMs) handle bug fixing versus problem solving in competitive programming. Using a dataset of ~3,000 Codeforces submissions and their human fixes, the authors compare LLM-generated patches to human patches and assess whether LLMs prefer to modify buggy code or generate new solutions. Results show that LLMs often alter more lines than necessary and sometimes produce entirely new solutions, performing better when allowed to generate solutions from scratch rather than patching existing code.
By Alexandru Stefan Stoica, Traian Rebedea, Marian Cristian Mihaescu
The paper investigates whether the Voxtral audio‑language model can detect speech spoofing. It shows that without task‑specific adaptation, the model’s language‑model layers prioritize semantic content, making spoof‑discriminative acoustic cues less separable. By applying lightweight weight‑decomposed low‑rank adaptation (DoRA), the authors create Spooftral, which achieves an equal error rate of 4.25% on the ASVspoof5 evaluation set.
By Avishai Weizman, Yehuda Ben-Shimol, Itshak Lapidot
The paper introduces a method that allows automatic speech recognition systems to learn new words during test time using unlabeled data. It combines a frozen CTC acoustic model for spellings, a frozen language model for detecting out‑of‑vocabulary words, and an adaptation module that expands the vocabulary by learning lexical token representations from CTC-generated candidates. Experiments on LibriSpeech and dysarthric speech data show relative character‑error‑rate reductions of up to 14.97% and 6.67% for recurring OOV words, respectively.
By Mengqi Wang, Mark A. Hasegawa-Johnson, Haolong Zheng, Chang D. Yoo
STRAND is a new benchmark that tests multimodal large language models’ ability to track objects, their states, and relationships over time in videos. It evaluates intermediate reasoning by breaking queries into sub‑questions and uses Faithful Accuracy to ensure all parts of an answer are correct. The authors also propose an object‑centric framework that builds structured trajectories and shows reduced hallucinations and better temporal consistency compared to existing models.
By Thong Nguyen, Tri Cao, Khoi Le, Cong-Duy Nguyen, Quynh Vo, See-Kiong Ng, Bryan Hooi Kuen-Yew
The paper introduces an LLM-assisted system for correcting speaker attribution errors during meetings. It combines streaming ASR, diarization, and concise LLM-generated summaries to guide users in providing brief corrective feedback, which updates the transcript and adds online speaker enrollments. The approach includes mechanisms to accurately interpret user corrections and a simulation for large-scale evaluation, achieving significant reductions in DER and speaker substitution error on the AMI headset test set, with a pilot usability study highlighting further improvements.
By Xinlu He, Yiwen Guan, Badrivishal Paurana, Pitipat Kongsomjit, Zilin Dai, Jacob Whitehill
The paper introduces a new evaluation framework for confidence estimation in large language models, focusing on three properties: robustness to prompt changes, stability across semantically equivalent answers, and sensitivity to semantically different answers. It demonstrates that existing confidence estimation methods perform well on robustness and stability but often fail to detect differences in answer meaning, revealing gaps in current evaluation practices. The framework aims to guide the selection of confidence estimators for practical applications.
By Yuxi Xia, Dennis Ulmer, Terra Blevins, Yihong Liu, Hinrich Sch\"utze, Benjamin Roth
The paper investigates continued pretraining (CPT) of the Finnish BERT model (FinBERT) on a Finnish histopathological dataset, noting that CPT train‑time loss curves vary significantly across domains. Since the histopathology data lacks labels, the authors use public Finnish datasets as proxy data to examine whether CPT‑derived signals correlate with downstream classification gains. Their exploratory analysis finds that certain CPT features are associated with improved proxy classification performance, adding to the sparse literature on Finnish healthcare NLP.
By Rami Luisto, Liisa Pet\"ainen, Tommi Gr\"onholm, Jan B\"ohm, Maarit Ahtiainen, Tomi Lilja, Ilkka P\"ol\"onen, Sami \"Ayr\"am\"o
The paper introduces a cause-aware error recovery framework for cascaded Automatic Speech Recognition – Large Language Model (ASR‑LLM) pipelines in Spoken Dialogue Systems. It replaces simple ASR confidence filtering with precision‑focused detectors that use deep ASR latent representations to classify token‑level errors into perception, comprehension, and deletion failures. This fine‑grained diagnosis enables the LLM to execute targeted, multi‑turn clarification strategies, leading to a more than two‑fold increase in recall on domain‑shift errors and significant reductions in word error rate and downstream task errors across varied accents, distortions, and domains.
By Yizhou Peng, Ziyang Ma, Changsong Liu, Yi-Wen Chao, Xie Chen, Eng Siong Chng
The paper investigates how to close the quality gap in low‑resource text‑to‑speech for Khmer and Korean using the VoxCPM2 model. By training a single low‑rank adaptation (LoRA) adapter on a shared 25.5‑hour corpus, the authors improve Khmer’s mean opinion score from 3.85 to 4.23 with a rank‑64 adapter, while Korean shows no significant gain. The study highlights that adaptation benefits mainly when the base model is weak and that training loss does not always align with human ratings.
By Phannet Pov, Hyun Woo Park, Voneat Pen, Sovandara Chhoun, Wan-Sup Cho, Saksonita Khoeurn
The paper introduces an evaluation framework for structured audio captions that separates acoustic and semantic aspects, such as timestamped sound event descriptions. It covers five axes—tag sets, descriptions, reasoning, numeric measurements, and spectral profiles—using large language model judges for semantics and deterministic metrics for temporal and acoustic features. Controlled perturbations validate that the metrics are robust to paraphrases but sensitive to real semantic and acoustic errors.
By Liang-Yuan Wu, Sripathi Sridhar, Mark Cartwright, Magdalena Fuentes
The paper investigates how trigger-based backdoors function in large language models by using sparse autoencoders (SAEs) to identify feature directions across layers and transformer components. In a controlled experiment, 1B and 8B models were made to continue English prompts in French or German when presented with fixed trigger sequences. The study finds that different SAE feature directions correspond to trigger detection, residual-stream propagation, and language tracking, with residual-stream features being most effective for controlling the backdoor behavior.
By Wissam Antoun, Francis Kulumba, Th\'eo Lasnier, Beno\^it Sagot, Djam\'e Seddah
MEVL-STP introduces a two‑stage pipeline for spotting arbitrarily shaped scene text. The detection stage fuses features from six frozen vision encoders via a hierarchical Feature Pyramid Network and a Progressive Scale Expansion network to produce precise polygon masks. The recognition stage then crops these masks and feeds them to a fine‑tuned Qwen3‑VL‑8B‑Instruct model, achieving state‑of‑the‑art detection and end‑to‑end performance on CTW1500, Total‑Text, and ICDAR 2015 without synthetic pretraining.
By Aman Anand, Partha Pratim Roy, Shivakumara Palaiahnakote
The paper introduces MK‑FSS, a few‑shot segmentation framework that leverages Multimodal Large Language Models (MLLMs) to extract spatial and semantic target knowledge from query images. Spatial knowledge is encoded into a memory representation and fused with support‑guided memory via a dual‑memory debate‑fusion module, while semantic knowledge is turned into a textual feature and combined with multi‑scale query features through a progressive cross‑modal prompt generator. Together, these components produce a robust target representation that improves segmentation performance over existing methods.
By Yijun Hu, Heng Fan, Libo Zhang
MoVISA introduces Multi-Token Reasoning for Video Object Segmentation, using multiple segmentation tokens (e.g., SEG0, SEG1) instead of a single token to better localize multiple objects over time. This approach enhances fine-grained alignment between language prompts and spatio-temporal mask predictions, leading to improved performance and interpretability. On benchmarks such as MeViS, DAVIS17, ReVOS, and Ref-Youtube-VOS, MoVISA achieves significant gains, notably a 13.2% J and F improvement on MeViS and an 8.4% J and F improvement on ReVOS.
By Ruining Zhao, Ho Kei Cheng, Alexander G Schwing
The paper introduces a tool‑augmented framework that enhances a small Vision‑Language Model (Qwen3.5‑4B) with geometric tools—3D object detection, metric depth estimation, and deterministic solvers for distance, size, and bearing—to improve metric spatial reasoning. By moving metric computation from the model’s weights into explicit solvers, the approach achieves significant gains on ReVSI‑Bench tasks, notably increasing absolute distance accuracy from 0.46 to 0.74 MRA and relative direction accuracy from 25.9% to 73.4%. The modular design allows swapping in different detectors, enabling a clear separation between perception and reasoning errors, and the model can autonomously sequence the tools to match a scripted pipeline on most tasks.
By Kai Glantz, Clemens Grange
The paper introduces ImCorr, a method for sub‑pixel semantic correspondence that uses an implicit feature field to eliminate grid‑based quantization errors inherent in patch‑based vision transformers. By training a FiLM‑conditioned decoder to embed precise positional information, ImCorr queries continuous coordinates on the source side and decodes onto a denser grid on the target side, thereby reducing representation‑level errors. Experiments on SPair‑71k and AP‑10K show significant gains at fine‑grained thresholds, with a 6.2‑point improvement at PCK@0.01 over the previous state of the art.
By Yusung Choi