ThinkSafe is a method for aligning large reasoning models with safety objectives without relying on external teachers. It uses a frozen student model to generate safety-alignment data through refusal steering and a safety guard, then fine‑tunes offline. Experiments on DeepSeek‑R1‑Distill and Qwen3 demonstrate significant safety gains while maintaining competitive reasoning performance, and on Qwen3‑0.6B it outperforms GRPO in roughly one‑seventh the wall‑clock time.
By Seanie Lee, Sangwoo Park, Yumin Choi, Gyeongman Kim, Minki Kang, Jihun Yun, Sung Ju Hwang, Dongmin Park, Jongho Park
The paper presents InstruBPM, an instruction‑tuned open‑weight large language model that generates BPMN process models from natural language descriptions. By preparing paired text‑diagram data and applying parameter‑efficient fine‑tuning with quantization, the authors demonstrate that the tuned model outperforms both untuned open‑weight baselines and strong proprietary models in sequence, structural, and guideline‑conformance metrics. The approach is cost‑effective, privacy‑preserving, and supports on‑prem deployment, with publicly released models and scripts for reproducibility.
By G\"okberk \c{C}elikmasat, Atay \"Ozg\"ovde, Fatma Ba\c{s}ak Aydemir
The study compares chemical foundation models (ChemBERTa, MoLFormer) with general‑purpose models (GPT‑6 Astra, Claude Opus 5.5) on chemical property prediction tasks. On five physicochemical and quantum‑chemical regression tasks with only 60 experimental datapoints, GPT‑6 Astra outperforms the specialized chemical models. When the chemical models receive ten times more data (1,000–10,000 points), they match GPT‑6 Astra’s performance at 60–1,000 points, indicating that general‑purpose models trained on broad knowledge can be more effective in data‑scarce chemical research.
By Stepan N. Ostarkov, Gregorii A. Semakin, Ivan A. Bespalov, Michael G. Medvedev
The paper investigates how large language models (LLMs) incorrectly use post‑cutoff information when asked to answer from an earlier time, a problem termed temporal leakage. It shows that explicit cutoff statements and prefix constraints reduce leakage more effectively than implicit framing or suffix constraints, but that prompting alone cannot give models the ability to verify temporal admissibility. To overcome this, the authors introduce Temporal Critique Fine‑Tuning (TCFT), which trains models to detect and explain post‑cutoff leakage and judge whether a response is temporally admissible, achieving significant reductions in leakage on Qwen2.5 models.
By Chenlu Ding, Jiancan Wu, Yanchen Luo, Zheyuan Liu, Yancheng Yuan, Xiang Wang
EgoActionBench is a new benchmark for dense action captioning in egocentric videos, featuring 1,407 episodes and 41,132 fine‑grained action annotations that specify which hand (left, right, or both) performs each temporally localized atomic action. The benchmark introduces an evaluation protocol that assesses semantic fidelity, temporal overlap, action precision, and recall, thereby measuring what happens, when it happens, and which hand is involved. Twenty baseline models were evaluated, and two backbones (Qwen3‑VL‑4B and LLaVA‑OneVision‑2‑8B) were fine‑tuned and reinforced, achieving significant score improvements and outperforming all open‑source baselines.
The paper investigates how the optimizer Muon compares to AdamW during fine‑tuning of large language models. While Muon speeds up pretraining loss minimization, the study finds that when pretraining data is sufficiently large, AdamW actually outperforms Muon after fine‑tuning. The authors explain this reversal by introducing the margin‑to‑spread ratio, showing that Muon has a higher ratio at small data budgets but AdamW surpasses it as data grows, a trend confirmed both analytically and in simulations.
The paper investigates how a chess transformer, Maia‑2, simulates players of different skill levels by conditioning on quantitative skill ratings. By fixing board positions and model weights, the authors show that the model’s internal knowledge of 172 chess concepts remains stable across skill levels, yet lower‑skill predictions can be dramatically improved by fine‑tuning only the final output layer. This demonstrates that weaker behavior in generative models may stem from how retained knowledge is applied rather than from a lack of knowledge itself.
The paper examines how fine‑tuned Transformer detectors (DistilBERT, RoBERTa) perform on an out‑of‑distribution dataset of hyper‑local news generated by LLMs. Using SHAP‑based explainable AI, it identifies three phenomena: a reliance on a “Call‑to‑Action” syndrome, a “Specificity Bias” where hallucinated named entities evade detection, and a “Wikipedia Paradox” where formal human text is misclassified as AI. These results highlight systematic shortcomings of purely statistical detection methods in structured human domains.
The paper demonstrates that a frozen vision‑language model (VLM) can serve as a zero‑shot judge to rerank candidate lists for cloth‑changing person re‑identification (CC‑ReID). Using the VLM to rerank two expert models on the CHIRLA benchmark improves rank‑1 performance by an average of 12.9 points over the best expert and 9.4 over a naive ensemble, without any fine‑tuning on target data. By gating the VLM’s use to cases where the experts disagree at rank‑1, the authors reduce compute by 40% with minimal impact on rank‑1 accuracy.
BRIDGE (Balancing Representations by Identifying and Generating Under-represented Data) is a label‑free feedback loop that augments a long‑tailed unlabeled corpus by selecting under‑represented images based on mean kNN distance and generating local variants with a frozen image‑to‑image diffusion model. The method reallocates a finite budget to a sparse band of the score distribution, avoiding the extreme tail, and continues pretraining on the enlarged set. Across five long‑tailed regimes, BRIDGE achieves the highest linear‑probe and kNN performance, improves ViT‑S under contrastive and self‑distillation objectives, and benefits low‑label fine‑tuning and semantic segmentation, while matching an oracle that restores withheld real images.
The paper introduces a control foundation model that treats control-experiment data as dynamic graphs, allowing both response prediction and control-input generation to be handled as a unified masked prediction problem. It defines a data space of dynamic graphs with connections, signal values, and missing values, and uses the Block Diagram Transformer (BDT) to combine temporal attention with graph neural networks. Trained on mixed experimental settings, the BDT successfully stabilized an unstable system via feedback control, outperforming fine‑tuned time‑series foundation models.
The paper introduces D2MoE, a Mixture of Task Adapters designed for vision foundation models to handle multi-task dense prediction. It separates adaptation into a task‑generic low‑rank channel (TG‑LoRA) that preserves pretrained priors and a task‑adaptive sparse expert module (TA‑MoE) that resolves task‑specific conflicts. A Semantic Prior Injector further aligns multi‑level task features with frozen foundation knowledge, achieving robust multi‑task performance with a small trainable footprint.
AffToM is a new benchmark for affective Theory of Mind (ToM) in dynamic multi‑party interactions, comprising 1,922 test pairs and 19,759 training pairs. It evaluates Individual Affective Reasoning (IAR) and Multi‑party Affective Reasoning (MPAR), the latter tracking nested beliefs and emotions as events unfold. The authors introduce an event‑driven multi‑agent generation pipeline and a two‑stage post‑training framework, AffToM‑R1, which boosts MPAR accuracy by 13.38 percentage points over baseline models.
The paper investigates catastrophic forgetting in the scientific foundation model Walrus-1.3B during a two‑stage sim‑to‑real adaptation. Finetuning first on simulated flow around a cylinder quickly degrades performance on the original pretraining data, with one‑step errors rising up to 35× after 20 epochs, and further adaptation to experimental data worsens this degradation. The authors show that replaying as little as 1% of the simulated data during real‑data finetuning largely restores simulation performance while still adapting effectively to real dynamics.
By Haaris Mian, Miles Cranmer, Alessandro Favero, Payel Mukhopadhyay
The paper introduces Muon-RT, a variant of the Muon self‑distillation method that eliminates the need for a separate momentum buffer by reusing the teacher‑student displacement. By removing the radial component, normalizing spectrally, and mixing with the current gradient after Newton‑Schulz orthogonalization, Muon‑RT achieves linear contraction of the objective gap under certain assumptions. Experiments on Qwen2.5‑7B‑Instruct and Olmo‑3‑7B‑Instruct show that Muon‑RT outperforms AdamW and momentum Muon while reducing persistent memory usage by up to 42%.
By Aryan Das, Vinay Kumar Verma, Aakansha Mishra
AquaProbe is a method that efficiently adapts frozen vision models to underwater imagery by measuring how the domain gap varies with water type and imaging range. It uses paired clean and physically rendered images to separate new divergence from accumulated encoder divergence, then allocates a fixed LoRA budget based on this profile. The resulting proportional LoRA rank allocation achieves a mean Intersection-over-Union of 70.1, outperforming uniform allocation by 1.9 points while using the same 0.59 million adapter parameters, and a three‑way router further improves performance to 70.6 mIoU on the SUIM split.
By Md Shihabul Islam Shovo, Nishi Kanta Paul, Trung Dong, Xiaomin Lin
The paper introduces Pancreas‑MedNeXt, a unified MRI segmentation model that simultaneously delineates the entire pancreas, its subregions, pancreatic tumors, and surrounding organs at risk (OARs) from a single MRI volume. Using a shared MedNeXt encoder‑decoder and an annotation‑aware loss, the model is trained on partially labeled data and fine‑tuned for multitask output heads. In cross‑validation on 142 PANTHER patients and external test sets, Pancreas‑MedNeXt outperforms state‑of‑the‑art methods, achieving higher Dice similarity coefficients for pancreas, tumor, and OARs.
"whyItMatters":"The model’s superior performance and unified approach could streamline contouring workflows for pancreatic cancer radiotherapy, potentially improving treatment planning efficiency."
By Zach Eidex, Mojtaba Safari, Kundan Thind, Parag J. Parikh, Zhen Tian, Xiaofeng Yang
The paper investigates the faithfulness of visual information in multimodal chain-of-thought (MCoT) reasoning by large vision-language models (LVLMs). It shows that visual cues in MCoT are often inaccurate and largely ignored, as predictions change little when visual input is altered but significantly when textual input is changed. To improve visual faithfulness, the authors propose an annotation‑free Sufficient‑Component Cause Model (SCCM) that encourages the generation of minimal yet sufficient visual components, and demonstrate its effectiveness across fine‑grained perception and reasoning benchmarks.
By Zujing Liu, Junwen Pan, Qi She, Yuan Gao, Guisong Xia
The paper critiques current LLM safety evaluations that rely solely on severity taxonomies, arguing they ignore the probability that a threat will actually be executed. It introduces Expected Harm, a metric that multiplies severity by execution likelihood—modeled via execution cost—to better gauge real-world risk. Empirical studies show that models over‑refuse low‑likelihood, high‑cost threats while remaining vulnerable to high‑likelihood, low‑cost ones, a miscalibration traced to the distribution of safety examples in supervised fine‑tuning. This misalignment can be exploited to double the success rate of jailbreak attacks.
By Yen-Shan Chen, Zhi Rui Tam, Cheng-Kuang Wu, Yun-Nung Chen
The paper proposes pre-training uncertainty estimators on large, diverse materials science datasets to improve uncertainty quantification (UQ) for machine‑learned interatomic potentials (MLIPs). By evaluating these pre‑trained models on both in‑distribution and out‑of‑domain adsorption‑energy predictions, the authors demonstrate that zero‑shot UQ is viable and that fine‑tuning yields strong correlations between predicted uncertainty and actual prediction error, even on challenging datasets. This approach aims to provide foundational UQ models that enable broader, more reliable deployment of MLIPs across chemical space.
By Stefan P. Schmid, Tobias Setzer, Maximilian Scheurer