A focused look at this week’s most significant advances in llms & foundation models \u2014 83 papers surveyed from arXiv and leading AI labs.<\/p>\n
—<\/p>\n
This week in foundation models, researchers pushed further into understanding how these systems learn, what they represent internally, and how far we can scale them. The papers span training dynamics, evaluation methodology, and architectural innovations.<\/p>\n
Prompt Injection in Automated R\u00e9sum\u00e9 Screening with Large Language Models: Single and Multi-Injection Settings<\/strong> \u2014 Preet Baxi, Jiannan Xu, Jane Yi Jiang, Stefanus Jasin<\/em><\/p>\n Large language models (LLMs) are increasingly used to screen and rank job applicants, creating incentives for candidates to strategically manipulate algorithmic hiring systems. We study prompt injecti…<\/p>\n arXiv<\/a><\/p>\n TOPS: First-Principles Visual Token Pruning via Constructing Token Optimal Preservation Sets for Efficient MLLM Inference<\/strong> \u2014 Tinghao Wang, Yichen Guo, Rui Huang, Zheng Lu, Qizhe Zhang, Chenxi Li, Yuan Zhang, Jiajun Cao, Zhirong Shen, Yaosong Du,<\/em><\/p>\n Multimodal large language models (MLLMs) have achieved strong multimodal reasoning capabilities, but their efficiency is limited by the large number of visual tokens, which introduces substantial comp…<\/p>\n arXiv<\/a><\/p>\n NuclearQAv2: A Structured Benchmark for Evaluating Domain-Science Competence in Large Language Models<\/strong> \u2014 Henry Shaowu Yuchi, Michal Kucer, Benjamin H. Sims, Selma Peterson, Emily Taylor<\/em><\/p>\n Large language models (LLMs) have demonstrated strong performance across a wide range of tasks, but ensuring their reliability in highly technical domains remains a significant challenge. In nuclear e…<\/p>\n arXiv<\/a><\/p>\n The Riddle Riddle: Testing Flexible Reasoning in Large Language Models and Humans<\/strong><\/p>\n Bella Fascendini, Kathryn McGregor, Max D. Gupta, Thomas L. Griffiths<\/em><\/p>\n Humans flexibly adapt their reasoning strategies to the requirements of a given problem. Large language models (LLMs) have performed well on many cognitive tasks, however, it is unclear whether this accuracy is a result of pattern matching from…<\/p>\n arXiv<\/a><\/p>\n RelAfford6D: Relational 6D Affordance Graphs for Constraint-Driven Robotic Manipulation<\/strong><\/p>\n Guodong Zhang, Qichen He, Wenyuan Xie, Shaokai Wu, Yanbiao Ji, Qiuchang Li, Bayram Bayramli, Yue Din<\/em><\/p>\n Bridging abstract semantics and precise physical control remains a fundamental challenge in open-world robotic manipulation. While recent data-driven policies show promise, their reliance on isolated contact points or latent affordance embeddings…<\/p>\n arXiv<\/a><\/p>\n Data-Free Reservoir Features for Efficient Long-Horizon Cold-Start Continual Learning<\/strong><\/p>\n Augustinas Ju\u010das, Yangchen Pan<\/em><\/p>\n Cold-start exemplar-free class-incremental learning requires learning a growing set of classes without replay, external pretraining, or a large initial task. Existing cold-start methods typically either train the backbone throughout the stream and…<\/p>\n arXiv<\/a><\/p>\n Inherited Circuits, Learned Semantics: How Fine-Tuning Creates Evasion Vulnerabilities Invisible to Standard Evaluation<\/strong><\/p>\n Ryan Fetterman<\/em><\/p>\n LLMs fine-tuned for security classification are usually evaluated on held-out examples from the same distribution as their training data. We show that this can miss vulnerabilities introduced by fine-tuning itself: models can learn token-level…<\/p>\n arXiv<\/a><\/p>\n A Generalization Theory for JEPA-Based World Models<\/strong><\/p>\n Jingyi Cui, Qi Zhang, Hongwei Wen, Yisen Wang<\/em><\/p>\n Joint Embedding Predictive Architectures (JEPAs) have recently emerged as a promising paradigm for world modeling by learning predictive dynamics in a latent space rather than generating future observations at the input level. Despite their…<\/p>\n arXiv<\/a><\/p>\n KARLA: Knowledge-base Augmented Retrieval for Language Models<\/strong><\/p>\n Francois Crespin, Fabian M. Suchanek, Nils Holzenberger<\/em><\/p>\n We propose a new method that allows an LLM to automatically pull in factual knowledge from a knowledge base during token generation. This means that (1)~factual knowledge in the LLM output can be updated without retraining the LLM, (2)~facts in the…<\/p>\n arXiv<\/a><\/p>\n Escaping Iterative Parameter-Space Noise: Differentially Private Learning with a Hypernetwork<\/strong><\/p>\n Naoki Nishikawa, Shokichi Takakura, Satoshi Hasegawa<\/em><\/p>\n Differentially private (DP) training of neural networks is often hindered by the large amount of noise required by gradient-based methods such as DP-SGD, which repeatedly inject high-dimensional noise in parameter space throughout training. In this…<\/p>\n arXiv<\/a><\/p>\n EEG Benchmarking Needs a Task Specification Layer: NeuroDoc for Rulebook-Guided, Executable Benchmark Construction<\/strong><\/p>\n Chengxuan Qin, Zhige Chen, Shu Peng, Rui Yang, Jiping Cui, Yikai Dong, Jun Li, Liu Peng, Zhida Shang<\/em><\/p>\n Electroencephalography (EEG) foundation models increasingly rely on multi-dataset training and evaluation, yet public EEG datasets still lack a shared task specification layer that can turn heterogeneous recordings into reusable benchmark units….<\/p>\n arXiv<\/a><\/p>\n DMuon: Efficient Distributed Muon Training with Near-Adam Overhead<\/strong><\/p>\n Vincent Chen, Starrick Liu, Regis Cheng, Dance Yang, Shalfun Li, Ryan Yu, Lucy Liang, Hang Su, Roy G<\/em><\/p>\n Matrix-orthogonalization-based optimizers, exemplified by Muon, have demonstrated strong convergence behavior across a wide range of modern deep learning workloads. The matrix-aware updates offer a compelling alternative to conventional…<\/p>\n arXiv<\/a><\/p>\n Just how sure are you? Improving Verbalized Uncertainty Calibration in Medical VQA<\/strong><\/p>\n Eren Senoglu, Federico Toschi, Nicolo Brunello, Andrea Sassella, Mark James Carman<\/em><\/p>\n Multimodal large language models (MLLMs) applied to Medical Visual Question Answering (VQA) tend to produce overconfident outputs regardless of actual correctness, and existing verbalized confidence calibration methods, developed primarily for text…<\/p>\n arXiv<\/a><\/p>\n Enabling self-supervised learned primal dual with Noise2Inverse<\/strong><\/p>\n Antti S\u00e4llinen, Siiri Rautio, Santeri Kaupinm\u00e4ki, Andreas Hauptmann<\/em><\/p>\n X-ray computed tomography reconstruction is an ill-posed inverse problem, particularly in low-dose and sparse-angle settings where measurements are noisy and incomplete. While learned reconstruction methods such as the Learned Primal-Dual algorithm…<\/p>\n arXiv<\/a><\/p>\n How Surprising Is Historical Italian to Language Models? Tokenization Tax, Comprehension Tax, and a Simple Mitigation<\/strong><\/p>\n Maria Levchenko<\/em><\/p>\n Large language models (LLMs) are increasingly critical to digital library workflows, yet their ability to process historical language remains poorly understood. Historical difficulty is typically treated as a monolithic barrier, conflating…<\/p>\n arXiv<\/a><\/p>\n Structure Before Collapse: Transient semantic geometry in next-token prediction<\/strong><\/p>\n Yize Zhao, Isabel Papadimitriou, Christos Thrampoulidis<\/em><\/p>\n Neural Collapse predicts that balanced one-hot classification pushes model representations to be equally far from each other; a symmetric configuration that depends only on the output label and ignores any semantic similarity in the inputs. This…<\/p>\n arXiv<\/a><\/p>\n Training for the Model You Return: Improving Optimization for Iterate-Averaged Language Models<\/strong><\/p>\n Kwok Chun Au, Adam Block<\/em><\/p>\n Many modern Language Model (LM) pipelines return an averaged model, such as an exponential moving average of the training iterates, rather than the final iterate itself. This raises a fundamental question: given that we will return an iterate…<\/p>\n arXiv<\/a><\/p>\n A probabilistic framework for online test-time adaptation<\/strong><\/p>\n Daniel Corrales, David R\u00edos Insua<\/em><\/p>\n This paper presents a probabilistic framework for online test-time adaptation problems. In them, a model is trained on labeled data but must adapt to unlabeled data at test time under the assumption that training and test distributions potentially…<\/p>\n arXiv<\/a><\/p>\n CORTEX: A Structured Reasoning Benchmark for Trustworthy 3D Chest CT MLLMs<\/strong><\/p>\n Hashmat Shadab Malik, Anees Ur Rehman Hashmi, Numan Saeed, Muzammal Naseer, Salman Khan, Christoph L<\/em><\/p>\n Reasoning in multimodal large language models (MLLMs) has shown strong promise in medical imaging. However, this reasoning is usually free-form text judged only by its final answer, making it hard to interpret and verify, especially in 3D…<\/p>\n arXiv<\/a><\/p>\n Confidence-Aware Tool Orchestration for Robust Video Understanding<\/strong><\/p>\n Yangfan He, Yujin Choi, Jaehong Yoon<\/em><\/p>\n Video reasoning language models implicitly assume that every input frame is equally reliable. This leads to what we term the Blind Trust Problem: under realistic perturbations such as motion blur, glare, or occlusion, frontier video reasoning…<\/p>\n arXiv<\/a><\/p>\n ReaORE: Reasoning-Guided Progressive Open Relation Extraction Empowered by Large Reasoning Models<\/strong><\/p>\n Xin Lin, Liang Zhang, Guoqi Ma, Hongyao Tu, Jinsong Su<\/em><\/p>\n Open Relation Extraction (OpenRE) requires a model to extract unseen relations between head and tail entities from unstructured text for real-world applications. The core challenge of OpenRE lies in achieving reliable generalization to unseen…<\/p>\n arXiv<\/a><\/p>\n Information-Aware KV Cache Compression for Long Reasoning<\/strong><\/p>\n Jushi Kai, Zhuiri Xiao, Alexandra Birch, Zhouhan Lin<\/em><\/p>\n Reasoning capability has advanced rapidly in large language models (LLMs), leading to an increasing size of key-value (KV) cache in both prefilling and decoding stages. Existing KV cache compression methods mainly rely on attention weights to…<\/p>\n arXiv<\/a><\/p>\n Beyond Logical Forms: LLM-Extracted Patterns for Fallacy Classification<\/strong><\/p>\n Eleni Papadopulos, Firoj Alam, Giovanni Da San Martino<\/em><\/p>\n In today’s fast-paced information era, logical fallacies, defined as defective patterns of reasoning, inevitably contribute to the growth of information disorder. However, often fallacies appear in nuanced forms that complicate automated…<\/p>\n arXiv<\/a><\/p>\n Statistically Valid Hyperparameter Selection: From Tuning to Guarantees<\/strong><\/p>\n Amirmohammad Farzaneh, Osvaldo Simeone<\/em><\/p>\n Hyperparameter selection is a critical step in the deployment of modern artificial intelligence systems, given the need to tune degrees of freedom such as inference-time parameters, implementation-level settings, and thresholds driving decision…<\/p>\n arXiv<\/a><\/p>\n RSPC: A Benchmark for Modeling Stress and Psychiatric Conditions in Digitally Mediated Relationships using Psychiatrist Annotations<\/strong><\/p>\n Parmitha Vangapandu, Sai Ganesh Mokkapati, Sathwik Narkedimilli, MSVPJ Sathvik, Timothy Liu, Simon S<\/em><\/p>\n In NLP, mental health conditions are often modeled as isolated phenomena, without interpersonal context. We use Reddit posts about long-distance relationships to capture both mental health distress and associated relational triggers. We introduce…<\/p>\n arXiv<\/a><\/p>\n SamaVaani: Auditing and Debiasing Multilingual Clinical ASR for Indian Languages<\/strong><\/p>\n Subham Kumar, Prakrithi Shivaprakash, Abhishek Manoharan, Astut Kurariya, Diptadhi Mukherjee, Prabha<\/em><\/p>\n Automatic Speech Recognition (ASR) is increasingly used to document clinical encounters, yet its reliability in multilingual and demographically diverse Indian healthcare context remains largely unknown. In this study, we first conduct the…<\/p>\n arXiv<\/a><\/p>\n Reproducibility Study of “AlphaEdit: Null-Space Constrained Knowledge Editing for Language Models”<\/strong><\/p>\n Ananth K S, Arya Hariharan<\/em><\/p>\n Fang et al. (2025) introduced a null-space constrained projection, named AlphaEdit, for locate-then-edit knowledge editing methods, theoretically guaranteeing that edits do not disrupt previously preserved knowledge, and reports substantial gains…<\/p>\n arXiv<\/a><\/p>\n SocialPersona: Benchmarking Personalized Profiling and Response with Multimodal Social-Media Context<\/strong><\/p>\n Qinkai Zhang, Yanyan Zhao, Xin Lu, Yulin Hu, Pengtao Han, Bing Qin<\/em><\/p>\n Personalized language-model assistants are often evaluated through a memory lens: can a model recall preferences users have explicitly stated in dialogue? More comprehensive personalization demands a harder capability — inferring what users care…<\/p>\n arXiv<\/a><\/p>\n Peek2: Regex-free Byte-level Byte-Pair Encoding Pretokenizer for LLM Inference on Edge Devices<\/strong><\/p>\n Liu Zai, Iraklis Klampanos<\/em><\/p>\n Pretokenization is a crucial, sequential pass in Byte-level BPE tokenizers, yet little work has been done to optimize it for edge-side inference. Our proposed new implementation, Peek2, serves as a drop-in replacement for cl100k-like pretokenizers…<\/p>\n arXiv<\/a><\/p>\n Application of LLMs to Threat Assessment of Foreign Peacekeeping Missions<\/strong><\/p>\n Gerhard Backfried, Christian Schmidt, Diego Pilutti, Michael Suker<\/em><\/p>\n We present a novel approach for applying Large Language Models (LLMs) to threat assessment in the context of foreign peacekeeping missions. Building on the PINPOINT project and its use case, the EU Monitoring Mission in Georgia, we combine an…<\/p>\n arXiv<\/a><\/p>\n Decision-Aligned Evaluation of Uncertainty Quantification<\/strong><\/p>\n Annika Schneider, Tommy Rochussen, Joshua Stiller, Vincent Fortuin<\/em><\/p>\n Uncertainty estimates in machine learning are typically evaluated using generic metrics such as the negative log-likelihood and expected calibration error, yet good performance on such metrics does not necessarily imply high utility in downstream…<\/p>\n arXiv<\/a><\/p>\n XMSE-Aware Adaptive Empirical Bayes Estimation<\/strong><\/p>\n Minghao Chen, Jiale Zheng<\/em><\/p>\n Empirical Bayes (EB) estimators can match the first-order asymptotic risk of maximum likelihood (ML) while behaving very differently at second order: recent excess mean squared error (XMSE) analysis shows that kernel-based EB estimation may be…<\/p>\n arXiv<\/a><\/p>\n Generative Models on Analog Hardware with Dynamics<\/strong><\/p>\n Yu-Neng Wang, Sara Achour<\/em><\/p>\n Analog hardware platforms such as coupled oscillators and Analog Ising Machines naturally solve differential equations at a fraction of the energy cost of digital computation, making them attractive for low-power generative modeling, yet a…<\/p>\n arXiv<\/a><\/p>\n Ribbon: Scalable Approximation and Robust Uncertainty Quantification<\/strong><\/p>\n Graham Gibson, John Tipton, Kellin Rumsey, Natalie Klein<\/em><\/p>\n Reliably quantifying predictive uncertainty is difficult for complex, high-dimensional, or misspecified models. Both fully Bayesian and bootstrap resampling methods provide principled uncertainty estimates but are often too expensive for modern…<\/p>\n arXiv<\/a><\/p>\n GAVEL: Grounded Caption Error Verification and Localization<\/strong><\/p>\n Zixian Gao, Atsushi Hashimoto, Kuniaki Saito<\/em><\/p>\n Vision-language models (VLMs) often produce hallucinated or inconsistent outputs, where text and images are not properly aligned. Addressing this issue requires not only detecting misalignment but also explaining the discrepancy and localizing its…<\/p>\n arXiv<\/a><\/p>\n FBK’s Long-form SpeechLLMs for IWSLT 2026 Instruction Following<\/strong><\/p>\n Zhihang Xie, Marco Gaido, Sara Papi, Matteo Negri, Luisa Bentivogli<\/em><\/p>\n This paper describes our submission to the IWSLT 2026 Instruction Following shared task. SpeechLLMs are developed for both short-form and long-form speech instruction following under constrained settings. For the short track, strong performance is…<\/p>\n arXiv<\/a><\/p>\n DnA: Denoising Attention for Visual Tasks<\/strong><\/p>\n Ron Campos, Subhajit Maity, Xin Li, Srijan Das, Aritra Dutta<\/em><\/p>\n The softmax activation in multihead attention (MHA) is the de facto standard for attention-based models in visual perception tasks. However, standard softmax can produce noisy attention patterns that dilute relevant features and degrade its…<\/p>\n arXiv<\/a><\/p>\n RoPEMover: Depth-Aware Object Relocation via Positional Embeddings<\/strong><\/p>\n Ipek Oztas, Duygu Ceylan, Aybars Bugra Aksoy, Aysegul Dundar<\/em><\/p>\n Moving an object in a single image requires geometry-consistent spatial rearrangement, including handling occlusions, revealing previously unseen regions, and maintaining coherent shadows and reflections. Existing approaches are not well suited to…<\/p>\n arXiv<\/a><\/p>\n Learning Probabilistic Filters with Strictly Proper Scoring Rules<\/strong><\/p>\n Eviatar Bach, Ricardo Baptista, Jochen Br\u00f6cker, Bohan Chen, Andrew Stuart<\/em><\/p>\n Bayesian filtering of partially and noisily observed dynamical systems seeks to infer the evolving conditional distribution of the state of a dynamical system, given observations, in an online fashion. This Bayesian filtering distribution is the…<\/p>\n arXiv<\/a><\/p>\n Stabilizing black-box algorithms through task-oriented randomization<\/strong><\/p>\n Yali Wang, Zhaojun Wang<\/em><\/p>\n As black-box models become foundational to modern research, ensuring their stability is paramount for the realization of trustworthy artificial intelligence. The inherent diversity of inputs – ranging from structured Gaussian distributions to…<\/p>\n arXiv<\/a><\/p>\n Efficient Adaptive Data Acquisition via Pretrained Belief Representations<\/strong><\/p>\n Daolang Huang, Zhuoyue Huang, Conor Hassan, Luigi Acerbi, Samuel Kaski, Tom Rainforth<\/em><\/p>\n Learning effective policies for adaptive data acquisition remains challenging: posterior-based methods rely on surrogate models and posterior approximations that can be misspecified or biased, while direct policy-learning methods map from…<\/p>\nTraining & Scaling<\/h2>\n
Reasoning & Inference<\/h2>\n
Safety & Alignment<\/h2>\n
Additional Research<\/h2>\n