The paper introduces Dr. Post-Training, a framework that treats general training data as a data‑induced regularizer to prevent overfitting to scarce target objectives, rather than merely a pool for selection. It constructs a feasible set of model update directions from general data and projects target‑driven updates onto this set, unifying standard training and existing data‑selection methods as special cases along a bias‑variance spectrum. Experiments on SFT, RLHF, and RLVR demonstrate that these methods outperform both standard training and state‑of‑the‑art data‑selection baselines, especially when general and target data are highly misaligned, while system optimizations keep overhead minimal.
By Pingbang Hu, Xueshen Liu, Z. Morley Mao, Jiaqi W. Ma
The paper presents a Physics Informed Neural Network (PINN) approach to solve the drift‑diffusion equations that govern semiconductor devices. By applying advanced training techniques, the authors achieve errors comparable to commercial TCAD software while handling extreme gradients and discontinuities in a PN‑junction. The method also enables inverse design, retrieving doping profiles from desired electric field or potential distributions under non‑equilibrium bias conditions.
By Luke Ragan, Matteo G. C. Alasio, Enrico Bellotti, Luca Dal Negro
InterQ‑RL is a new method for multi‑turn agent reinforcement learning that addresses execution mismatch caused by low‑precision quantized rollouts. It corrects the mismatch by applying bounded execution‑prefix adjustments over the policy history, allocating correction strength across action segments within a fixed displacement budget, and does so without extra model forward passes. Experiments on ALFWorld, WebShop, and ScienceWorld with 4B‑30B backbones show that InterQ‑RL achieves the best quantized performance, matching BF16 within 0.7 points and surpassing it in two settings while maintaining or improving learning speed.
By Yingbo Hao, Song Liu, Guanghua Yu, Zhaojing Zhou, Jianchen Zhu, Yi Zou
The paper introduces StressSim, an intent‑preserving adversarial user simulator designed to stress‑test multi‑turn agents by generating challenging yet valid user interactions. StressSim optimises high‑level adversarial plans for each task instance using a genetic prompt‑optimisation loop that leverages rollout traces, task‑completion signals, and validity feedback to reduce task success while keeping the user’s goal unchanged. Experiments on MultiWOZ, retail, and airline domains show that StressSim reveals significant robustness gaps in leading LLM agents—reducing tau‑bench pass^4 scores by up to 69%—while maintaining realistic conversation quality and goal preservation with a modest optimisation budget.
By Atri Vivek Sharma, Brian Formento, Alessio Lomuscio
The paper introduces Dual-Path Spatiotemporal Contrastive Learning (DP-STCL) for cross-subject EEG emotion recognition, using two complementary encoders—Temporal-to-Spatial and Spatial-to-Temporal—to model EEG data under different integration orders. It enforces consistency of high-level emotion representations across these paths and across subjects through three complementary relations, enhancing stability and discriminability. A LoRA-based few-shot adaptation strategy is added to handle limited labeled data from target subjects, and experiments on SEED, SEED-IV, SEED-V, and FACED datasets show DP-STCL outperforming existing methods.
By Zehui Xiao, Yi Zheng, Guixing Fu, Wenkai Lu, Lingfei He, Tao Tan, Yue Liu, Silun Wang, Junhai Xu, Mengting Liu
The paper introduces MUNKEY, a memory‑augmented transformer that enables machine unlearning by design. By decoupling instance‑specific memorization from model weights, unlearning is achieved simply by deleting an instance‑identifying key, allowing zero‑shot forgetting without retraining or accessing original data. Experiments on natural image, fine‑grained visual recognition, and medical datasets show that MUNKEY outperforms existing post‑hoc unlearning methods while maintaining predictive performance.
By Sonia Laguna, Jorge da Silva Goncalves, Moritz Vandenhirtz, Alain Ryser, Irene Cannistraci, Julia E. Vogt
Markovian Circuit Tracing (MCT) is a benchmark that tests whether transformer activations encode a compact state abstraction with recoverable transition structure. In experiments with small causal transformers trained on sequences from four‑state hidden Markov models, models achieve near‑optimal validation loss and recover transition matrices accurately under easy and medium observability, but performance collapses to control levels under hard observability. Various controls—including shuffled data, untrained activations, and sparse autoencoders—demonstrate that the recovered transition structure is genuinely causally useful rather than a trivial artifact.
By Abdullah X
The paper introduces a harness designed to curb over‑defensive writing by agents, which occurs when agents add stronger claims than the source material and then retract them. The harness separates the reviewer, filter, and writer components, allowing the writer to focus only on problems selected by an independent filter. Evaluation on 25 business‑project drafts shows a 96.61% drop in defensive propositions and an increase in source‑claim retention from 80.35% to 93.12%.
By Xiaochen Chang
PrinGS introduces a privacy‑preserving framework for 3D Gaussian Splatting, enabling post‑training anonymization of sensitive content such as faces, license plates, documents, and computer screens. By localizing and directly anonymizing regions in the 3D representation, the method removes sensitive information from all rendered novel views while maintaining scene geometry and visual fidelity, achieving this in under one second. The approach also offers a privacy‑by‑design training strategy that prevents sensitive data from being encoded during optimization, yielding splats with low face similarity and unreadable text.
By Gaurav Behera, Abhinav Raundhal, Archisha Panda, Saumya Balina, Gargi Shroff, Ponnurangam Kumaraguru
The paper investigates the safety implications of few-step distilled text-to-image models, finding that faster generation often leads to higher unsafe content rates and that existing concept-erasure methods are less effective on these models. The authors diagnose why erasure fails, attributing it to the fragile generation process and the way guidance is distilled. They propose DICE, a post-distillation erasure technique that aligns with the distillation process, achieving a significant reduction in unsafe-generation rates while preserving quality.
By Lior Ziv, Mahmood Sharif
The paper introduces a dynamic conceptual representation framework for video recognition, extending Concept Bottleneck Models from static images to procedural videos. It encodes a multi‑step video as a temporal sequence of segment‑level concept score vectors and processes this sequence with a Transformer encoder to predict the video’s class label. Experiments on cooking and household‑activity videos show that coarser‑grained concepts are sufficient for accurate classification, suggesting that fine‑grained textual descriptions may be unnecessary.
By Ryoya Yoshiyama, Bowen Wang, Yuta Nakashima
The article introduces deep Bayesian filtering techniques for nonlinear structural system identification, emphasizing how learnable dynamics can be integrated with physical inductive biases to achieve uncertainty-aware inference. It reviews key models such as the Deep Markov Model, Physics-guided Deep Markov Model, and Neural Extended Kalman Filter, and evaluates them on a three‑dimensional shear‑frame benchmark with hysteretic Bouc‑Wen connections. The study compares measurement‑conditioned reconstruction versus open‑loop forecasting, explores the impact of physical structure on latent‑state interpretability, and discusses robustness in response and parameter estimation, offering practical guidelines for model construction and deployment.
By Wei Liu, Konstantinos Vlachas, Eleni Chatzi
The paper introduces a lightweight neural network inspired by insect olfaction that classifies raw gas sensor readings using a sparse, high‑dimensional binary representation. It operates online, handling random sequences of single gases and binary mixtures, and dynamically incorporates new samples while providing class‑level frequency estimates to differentiate familiar from novel stimuli. The architecture, built with binary weights and Hebbian local learning, outperforms a linear SVM on mixtures, matches an optimized nonlinear SVM, and was demonstrated on the TEXEL neuromorphic chip with 74.6% accuracy, 482 nJ per inference, and 27 µW static power.
By Ton Juny Pina, Hugh Greatorex, Elisabetta Chicca
ToPO: Token-Oriented Preference Optimization for Text-to-Image Diffusion Models proposes a method that builds a token-conditioned spatial-temporal weighting map for each image pair. By using a frozen reference denoiser to derive spatial and temporal weights from local noise-prediction error differences and linking spatial evidence to content tokens via preferred-branch cross-attention, ToPO reweights Diffusion-DPO and incorporates an auxiliary pixel-midpoint ordering loss. The approach requires no local preference annotations or a separately trained reward model, and achieves state-of-the-art performance with only 500 training updates across multiple preference and compositional evaluations.
By Juntao Xu, Shihong Li, Hoi Fan Au, Ning Zhu
The paper introduces an ontology‑guided, multi‑agent framework for differential diagnosis that mirrors physicians’ iterative reasoning. By grounding inference in a structured, expert‑curated medical knowledge base, the system produces explicit, auditable reasoning steps, reduces hallucination, and improves transparency. It also identifies key missing symptoms and generates targeted follow‑up questions to lower diagnostic uncertainty.
By Roxana Ramona Szomiu, George Ursu, Adrian Groza
The paper investigates whether large language models (LLMs) can reliably grade Olympiad-level mathematical proofs. It introduces a human-curated dataset of USAMO 2025 candidate solutions with expert grading and evaluates 20 LLM verifiers across multiple metrics, prompt styles, and inference methods, finding that current models still have significant shortcomings. To improve performance, the authors propose TROJAN, an iterative self-critique pipeline for generating adversarial proofs, and MABGrader, a Multi-Armed Bandit framework that outperforms the state-of-the-art ProofGrader by 26.4% in Quadratic Weighted Kappa and reduces grading error by 14%.
By Azim Ospanov, Zijin Feng, Ding Ding, Chengwu Liu, Haoli Bai, Jiacheng Sun, Lifeng Shang, Farzan Farnia
The paper introduces Wiener-informed Decoding (WiD), a method that leverages frozen Wiener estimates and statistical knowledge to improve the training of pixel-space flow models. By integrating a trainable Wiener prediction path with coefficient interfaces that inject features into the native decoder, WiD accelerates training, enhances generation quality, and yields smoother gradient trajectories across various model scales and decoder architectures. Experiments on ImageNet 256×256 show that PixelDiT-XL with WiD achieves an FID of 1.53 after 500 epochs, matching the quality of the baseline while requiring 37.5% fewer training epochs.
By Ziqiao Wang, Yang You
The paper introduces AdvGuard, a model‑agnostic framework designed to detect and mitigate query‑based system prompt leakage attacks against large language models. AdvGuard constructs a bounded semantic interaction graph to capture cross‑query relationships, applies statistical filtering, and uses graph convolutional networks for classification, enabling it to identify coordinated extraction attempts across multiple turns. In evaluations, AdvGuard achieves significant improvements in recall and F1 scores over baselines, with negligible false positives, and effectively limits prompt leakage in adaptive attack scenarios.
By Tushar Nayan, Pratik Poudel, Qian Lou, Yuede Ji, Jason Liu, Ruimin Sun
The paper introduces Closed-Loop Aware Planning for Autonomous Driving (CLAP‑AD), a new planning head that learns a driving path via imitation while independently scoring candidate speeds using a learned function guided by safety constraints. Unlike traditional open‑loop imitation learning, CLAP‑AD allows explicit test‑time control over speed selection without requiring speed‑conditioned training. Experiments on Bench2Drive‑Speed, Bench2Drive, and NAVSIM v2 demonstrate significant improvements in speed‑adherence and overall driving performance compared to state‑of‑the‑art baselines.
By Emad Aghajanzadeh, Thomas Schwarzfischer, Marc Kaufeld, Moritz Thoma, Ryan P\'egoud, Manoj Rohit Vemparala, Pierpaolo Mori, Nael Fasfous, Johannes Betz, Daniel Mueller-Gritschneder, Ulf Schlichtmann
HOLogicBench is a 1,379-instance benchmark designed to evaluate higher-order logical reasoning in large language models, focusing on operations over rules and policies rather than just object-level inference. It includes Isabelle/HOL formalizations and proof‑verified answers, and provides solver‑verified reference reasoning paths for fine‑grained evaluation. Across 16 model configurations, average accuracy is 53.86%, with the best reaching 65.99%, and analyses show limited alignment with valid derivations, vulnerability to minimal counterfactual perturbations, and distinct failure patterns across reasoning types.
By Yucheng Wu, Jundong Xu, Mingzhen Ju, Yue Yu, Chenpeng Wang, Haoxuan Li, Liangming Pan