The paper investigates whether in-context learning (ICL) in large language model agents reflects genuine recursive reasoning or simply statistical extrapolation. By testing LLM agents in a public goods game with manipulated historical feedback, the authors compare decision quality to a rational expectations equilibrium benchmark. They find that disrupting historical patterns eliminates the benefits of longer context, especially in highly interdependent settings, indicating that ICL behavior aligns more with statistical extrapolation than strategic reasoning.
By Yu Liu, Wenwen Li, Yifan Dou, Guangnan Ye
The paper investigates whether knowledge editing truly erases original facts from language models. Using a linear trace probe, the authors find that after editing a fact in GPT‑2‑XL, the original object remains highly decodable from hidden states across three different editing methods, even when the model behaves correctly on edited prompts. This suggests that editing suppresses rather than removes the original association in representational space.
By Priyansh Srivastava, Romit Chatterjee
The paper introduces a five-task diagnostic experiment that separates perceptual and reasoning failures in multimodal large language models on physics and geometry benchmarks. It finds that misinterpreting diagrams hurts performance even on text-only solvable problems, and that accuracy improves when models receive human-authored captions. The study shows that correcting captions can recover many errors, revealing distinct reasoning bottlenecks that differ by domain, while a heavily pretrained model still underperforms and often truncates reasoning traces.
By Raj Jaiswal, Sree Krishna Uppalapati, Dhruvkumar Patel, Ria Khatoniar, Tanuja Ganu, Rajiv Ratn Shah
The paper investigates whether domain-specific fine‑tuning benefits open‑ended scientific reasoning in astronomy. Using a curated 300‑question QA benchmark from 2017–2026 Olympiad‑style materials, the authors compare open‑weight, API‑served general‑purpose, multimodal, and astronomy‑specialized language models. Results show that strong general‑purpose models set the highest correctness baseline, but variations in metric agreement, judge sensitivity, benchmark composition, and modality suggest that domain specialization is task‑ and deployment‑dependent and that domain‑specific evaluation is crucial for scientific workflows.
By Vanessa Lama, Sanjay Das, Emily Herron, Yuan-Sen Ting, Tijmen de Haan, Junqi Yin, Tirthankar Ghosal, Feiyi Wang
EDCT-Bench is a benchmark that evaluates the faithfulness of Vision‑Language Models (VLMs) by using Explanation‑Driven Counterfactual Testing (EDCT). EDCT extracts visual concepts from a model’s natural language explanation, applies minimal verified edits to those concepts, and checks whether the model’s answer and explanation remain consistent with the edited image. The benchmark covers three domains—knowledge‑intensive VQA, safety‑critical driving, and 3D spatial reasoning—and reveals significant faithfulness gaps in current VLMs, while also showing that EDCT‑generated counterfactuals can improve training.
By Sihao Ding, Santosh Vasa, Aditi Ramadwar, Thomas Monninger
The paper examines three recent studies that highlight distinct reliability issues in large language models (LLMs). Each study points to a missing capability—whether a tool‑reliability representation, safe generation behavior, or a consistent‑reasoning function—yet all converge on the need for calibrated abstention. The authors argue that current benchmarks fail to reward abstention, preventing the development of this function, and propose four evaluation changes to address the gap.
By Srijith Ravikumar
The paper introduces XConf, an experiential confidence estimator that augments a language model’s current inference with a record of its past graded episodes. By recalling similar past tasks and reflecting on past outcomes, XConf generates confidence scores without accessing logits or updating weights, achieving superior discrimination and calibration across diverse benchmarks. The method demonstrates significant gains in selective prediction, improving success rates on agent tasks by up to 8.7 points.
By Caiqi Zhang, Xiaochen Zhu, Chengzu Li, Yulong Chen, Dharshan Kumaran, Nigel Collier
The paper introduces MuViS-C, a multi‑domain benchmark that evaluates the robustness of learning‑based virtual sensing models against ten common sensor failure modes, ranging from subtle drifts to catastrophic dropouts. It assesses models using average error, relative degradation, and worst‑case fragility across nine datasets from six domains, comparing six architectures (gradient‑boosted trees, convolution, recurrence, attention, and MLP‑mixing). The study finds that all models degrade under corruption, gradient‑boosted trees are most robust, and targeted robustification can improve attention models at the cost of nominal performance.
By Jens U. Brandt, Noah C. Puetz, Alexander Windmann, Marc Hilbert, Elena Raponi, Thomas B\"ack, Thomas Bartz-Beielstein
MoRE: Mixture of Reused Experts is a hybrid architecture that combines Mixture-of-Experts (MoE) with weight‑sharing techniques. It shares expert pools across adjacent layers while each layer keeps its own router, and introduces lightweight depth embeddings to help shared experts differentiate layer contexts. Experiments on models ranging from 114 M to 1.15 B parameters show MoRE achieves lower perplexity and better downstream performance than standard MoEs and other weight‑sharing models, with only minimal changes to existing MoE implementations.
By Eric S. Qiu, Utku Umur Acikalin, Justin Lovelace, Christian Belardi, Arjun B. Mulchandani, Carla P. Gomes, Kilian Q. Weinberger
The paper presents CapMap-MS-TTA, a unified multimodal model that achieved third place in the MUMU track of the 8th Large-scale Video Object Segmentation Challenge. The solution tackles image tagging, open‑vocabulary object detection, and English captioning under strict resource limits, using an expanded keyword lexicon and lightweight expand‑hints for tagging, and Florence‑2 with multi‑scale and horizontal‑flip TTA plus label‑aware NMS for detection. The approach improves the reproduced Florence‑2 baseline from 15.16 to a best public score of 16.4815 without fine‑tuning.
By Chengfeng Qiu, Kaifeng Wei
The paper compares Knowledge-Graph Based Augmentation (Graph-RAG) with Retrieval-Augmented Generation (RAG) for answering culturally specific questions. Using the LatamQA dataset, Graph-RAG, built automatically from Wikipedia via KGGen, matches RAG performance and reduces the base LLM’s error by 72% with a standard KG and 78% with a benchmark-aware variant. The approach also transfers zero‑shot to Portuguese, showing multilingual applicability.
By Pablo Poulenard, Yannis Karmim, Valentin Barri\`ere
The paper introduces DualViewEval, a benchmark compression technique for agent evaluation that jointly uses outcome and process signals to learn an exact-size miniset and predict full-benchmark scores. By analyzing large-scale trajectories, the authors identify six complementary process signals linked to final agent performance. Across five agent benchmarks and five baselines, DualViewEval achieves superior compression (24×–40×) and lower error rates, while also revealing capability differences among agents for efficient development.
By Xinshuai Guo, Junjie Wu, Dolly Deng, Yinghui Li, Hai-Tao Zheng, Suncong Zheng, Maxm Pan
The paper introduces Dose-Aware Cold Diffusion (DACD), a physics-consistent framework that treats radiation dose as a continuous latent variable in a cold diffusion process for low-dose CT reconstruction. DACD combines dose-aware perception, multi-scale structural priors, and dose-calibrated step allocation to guide denoising, and adds an iterative forward-backprojection correction to enforce projection-domain consistency. Experiments on Mayo-2020, Mayo-2016, and LoDoPaB-CT show DACD outperforms existing diffusion-based and physics-guided methods, especially at ultra-low doses, and generalizes robustly across continuous, unseen dose levels.
By Md Imam Ahasan, Guangchao Yang, A F M Abdun Noor, S M Hasan Mahmud, Md Mahfuzur Rahman
MiST (Mid-trained Security Transformer) is a suite of 8B and 32B language models tailored for cybersecurity, achieving strong performance on public benchmarks. The approach uses a mid-training stage that adapts general pre-trained models to the domain by curating a compact, expert-vetted seed corpus and generating high-quality synthetic training data, rather than continual pre-training on large raw text. MiST checkpoints improve mean cybersecurity accuracy by +13.1 and +8.6 absolute percentage points over Qwen baselines for 8B and 32B models, respectively, and provide a stronger initialization for downstream task-specific fine-tuning and reinforcement learning.
By Oded Ovadia, Elad Ben Zaken, Elad Guttman, Orly Moreno Kadosh
arXiv:2602. 02888v2 Announce Type: replace-cross Abstract: Hallucinations remain a major obstacle for large language models (LLMs), especially in safety-critical domains.
By Ahmad Shapiro, Karan Taneja, Ashok Goel
LightSleepX is a lightweight, inception‑based dual‑modal network designed for sleep staging in resource‑constrained environments. It uses depthwise separable convolutions, multi‑scale enhanced attention for efficient EEG/EOG feature extraction, and a Mamba encoder for long‑range temporal modeling. On public benchmarks, it achieves 85.9% accuracy on Sleep‑EDF‑20 and 81.8% on ISRUC‑S3 with only 0.049M parameters and 195.9 MFLOPs.
By Yi Wang
arXiv:2609. 19039v1 Announce Type: cross Abstract: We introduce the Long-Short-Range Neural Network (LSR-Net), a novel neural operator architecture designed for data-driven forward evolution modeling, and extends it to the prediction of nonlinear fluid dynamics.
By Qian Hou, Sutrisno, Yuqing Li, Zecheng Gan
The paper introduces DantinoX, an open‑source JAX/Flax library that unifies autoregressive decoding, discrete masked diffusion, and continuous flow‑matching language modeling under a single modular Transformer backbone. By keeping the backbone, tokenizer, initialization, and training infrastructure consistent, users can switch between generation paradigms, attention mechanisms, or hardware topologies with only a configuration change. This design enables controlled cross‑paradigm comparisons within one API for training, streaming inference, and benchmarking.
By Marco Simoni, Aleksandar Fontana, Giulio Rossolini, Andrea Saracino
The paper re‑examines the impact of removing the encoder from Action Chunking Transformers (ACT), a model used for robot manipulation learning. Contrary to the original claim that encoder removal drops success rates from 35% to 2%, the authors find no such dramatic decline in their re‑runs, though minor variations remain uncertain. They attribute the discrepancy to factors like training length and checkpoint selection, and note that the encoder’s latent variable offers little reconstruction benefit on the tested benchmark, while its removal speeds up training.
By Bo Kang
The paper investigates which historical examples are most useful for time‑series forecasting by defining predictive relevance as the expected future utility conditioned on inference‑time information. It introduces a two‑stage approach: a normalized‑pattern retriever generates a coarse candidate set, and a lightweight MLP reranks these candidates using future‑supervised relevance while keeping inference strictly past‑only. Experiments on six benchmarks show that this reranker improves pattern retrieval and outperforms a matched‑protocol baseline, revealing that historical relevance is structured, domain‑dependent, and not governed by a single universal retrieval rule.
By Yong-Hoon Choi, Kwang-Hyun Park, Youngjin Cho