The paper evaluates legal text classification models for Korean sexual offense cases, comparing traditional machine learning, large language models, and fine‑tuned domain models. Fine‑tuned KLUE‑BERT achieved the highest accuracy of 99.3%, outperforming GPT‑3.5, GPT‑4.0, and other traditional approaches. Explainable AI techniques were used to analyze predictions, revealing linguistic features that influence decisions and highlighting limitations in capturing subtle textual cues, especially in real‑world KICS data.
By Jeongmin Lee
Tokenized Key-Gated Adapter Routing (Locket) is a framework that embeds fine‑grained, policy‑driven access control into large language models by training lightweight LoRA adapters for different privacy policies. A gating module associates a learned keyed entry token with a specific adapter, allowing authorized tokens to unlock private knowledge while invalid or missing tokens trigger privacy‑preserving adapters that redact or sanitize sensitive content. Experiments on datasets such as Enron, ECHR, and Yelp with models like Qwen3, Llama‑3.2, and Gemma‑2‑2B show that Locket maintains perplexity comparable to fine‑tuning when the correct token is provided, and significantly reduces PII leakage when the token is absent or invalid, without sacrificing utility.
By Mohamed Shaaban, Mohamed Elmahallawy
The paper introduces Kinematic MeanFlow (K-MF), a one‑step action generation policy for Robotic Foundation Models that addresses instability in the MeanFlow framework. By decoupling the time derivative into two sub‑interval terms, K-MF captures early and late denoising dynamics separately, reducing error amplification. Experiments show K-MF achieves faster inference—reducing action‑head latency by 67.5%–74.4% and overall end‑to‑end latency by 30.3%–54.9%—while outperforming multi‑step flow matching on various tasks.
By Jiawei Fan, Sifeng Wang, Yuqing Hou, Anbang Yao
The paper reports that applying reinforcement learning (RL) to the Kimi K2.7 Code model on 1,700 agentic coding tasks improves its performance on six external benchmarks. After a single epoch of GSPO training on a rank‑32 LoRA adapter, pass‑@1 scores increased across all benchmarks, with significant gains even on data released after training. The trained model also reduces agent steps and avoids common failure modes such as dropping requirements or breaking existing behavior.
By Sushant Mehta, Logan Ritchie, Edwin Chen
MWOP (Modality-aware Width-wise Operation Pruning) is a method that independently prunes visual‑to‑visual, text‑to‑visual, and text‑to‑text attention paths within each layer of multimodal large language models, and separately selects feed‑forward network channels for visual and textual inputs. It uses a first‑order Taylor criterion to guide pruning, re‑evaluates FFN importance after attention pruning, and applies LoRA‑based recovery training. The approach is paired with path‑sparse Triton attention kernels and compact visual‑side FFN execution to achieve practical acceleration, preserving token sequences while reducing computation.
"whyItMatters":"MWOP achieves a 1.6× prefill speedup on LLaVA‑OneVision‑7B while retaining 99.7% performance, and further boosts token‑compression methods to 2.9× and 2.7× speedups, demonstrating its effectiveness across architectures."
By Xudong Wang, Hao Wu, Haozhe Hu, Peiran Yin, Xinghao Chen, Yunpu Ma, Wei Zhang, Xiaoyu Shen
The paper introduces a Markov chain Monte Carlo (MCMC) sampling algorithm that transforms off‑policy traces into more on‑policy data for supervised finetuning (SFT). By tailoring the data distribution rather than changing the learning objective, the method allows SFT to match or surpass traditional post‑training techniques such as reinforcement learning across tasks like scientific skill acquisition, mathematical reasoning, and open‑ended expertise. The resulting models generalize better, forget less, and exhibit strong distributional performance, demonstrating that sampling can serve as a model‑native operator for broader post‑training applications.
By Aayush Karan, Sitan Chen, Yilun Du
KaliBench is a fine‑grained benchmark that evaluates large language models’ ability to translate natural‑language queries into executable command‑line instructions for Kali Linux. It contains 8,504 query–command pairs covering 1,642 tools across 23 capability dimensions and 5 security phases, and uses a deterministic, alias‑aware evaluation pipeline with multi‑stage verification (LLM validation, sandboxed execution, and human refinement). The benchmark also supports runtime‑free verifiable rewards, enabling training improvements that bring smaller models close to the performance of a 685B MoE model.
By Pengfei Li, Naufal Suryanto, Sicheng Zhang, Muzammal Naseer
HyperGuide introduces a hyperbolic geometric signal to guide multi-step reasoning in large language models, addressing the trade-off between efficient single-pass generation and computationally heavy tree-search methods. By projecting LLM hidden states into hyperbolic space, the approach leverages the space’s asymmetry—compact near the origin and exponentially expanding toward the boundary—to encode solution proximity and branch differentiation. A lightweight head and a fine-tuned low-rank adapter use this signal to improve reasoning accuracy, especially on deeper reasoning chains, with consistent gains across multiple benchmarks.
By Yuyu Liu, Haotian Xu, Yanan He, Sarang Rajendra Patil, Mengjia Xu, Tengfei Ma
The paper introduces masked self‑distillation, a post‑training framework that trains a language model to internalize portions of its own intermediate reasoning traces. By varying the fraction of trace internalized, the authors demonstrate that models can achieve higher inference efficiency and improved task performance on math and graph‑coloring problems. Experiments on Qwen3‑4B and Qwen3‑8B show that the method generalizes well to in‑domain out‑of‑distribution cases without catastrophic forgetting, and that supervised fine‑tuning alone can reduce trace length at the expense of generalization.
By Durgesh Kalwar, Vardhan Palod, Jaya Adithya Pavuluri, Subbarao Kambhampati
UrbanVLA is a Vision‑Language‑Action framework designed to enable delivery robots to navigate large‑scale urban environments using long‑horizon route instructions. The model aligns noisy route waypoints with visual observations and plans trajectories, trained through a two‑stage pipeline of supervised fine‑tuning on simulated data and reinforcement fine‑tuning on mixed simulation and real‑world data. Experiments show UrbanVLA outperforms strong baselines by over 55% on the SocialNav task and demonstrates reliable real‑world navigation in large urban settings.
By Anqi Li, Zhiyong Wang, Jiazhao Zhang, Minghan Li, Yunpeng Qi, Zhibo Chen, Zhizheng Zhang, He Wang
The paper introduces a new diffusion-based approach for probabilistic weather forecasting that incorporates auxiliary conditional denoising tasks. By predicting future states from both clean and corrupted contexts, the model reduces ambiguity and learns conditional distributions across noise levels using a proper scoring rule. Experiments demonstrate improved one‑step forecasting, especially at longer horizons, and better calibration in high‑dimensional global weather scenarios.
By Joonhyeong Park, Giung Nam, Hyungi Lee, Kyunghyun Cho, Byoungwoo Park, Juho Lee
The paper presents a mathematical framework that treats a deep classifier as a finite, nonautonomous discrete dynamical system, with each layer acting as a time step. It introduces the Finite-Time Maximum Lyapunov Exponent (FTMLE) to quantify how nearby data trajectories converge or diverge as they propagate through the network, and shows that at the probability level the FTMLE’s dominant direction aligns with the decision boundary normal for Gaussian classes. The authors further analyze how this relationship changes when moving backward to logits and hidden representations, identify conditions under which boundary-related FTMLE persists, and propose geometry-aware fine‑tuning to concentrate high hidden FTMLE near the decision boundary.
By Shirin Panahi, Amirhossein Nazerian, Ali Pezeshki
The paper introduces the Retain-and-Repair Neural Operator (R$^2$NO), a method that adapts neural operators pretrained on numerical simulations to real-world data while preserving useful pretrained structure. R$^2$NO first fine‑tunes the pretrained operator on real data, freezes it to serve as a source prediction, and then employs a shared repair module that learns refinements via orthogonal Fourier projections and a spectral ensemble of ridge regressions. The approach allows refinement depth to vary across frequency magnitude, orientation, and channels, and consistently outperforms full fine‑tuning and iterative refinement on RealPDEBench systems across six backbones.
By Woojin Cho, Junghwan Park
Hyperbolic Prototype Routing (HyPro) is a rehearsal‑free framework for class‑incremental learning that assigns a dedicated LoRA‑Expert module to each incremental task, enabling isolated representation learning. It projects routing features onto a Poincaré ball and uses geodesic nearest‑prototype matching for reliable task‑level discrimination. Experiments on standard and few‑shot CIL benchmarks demonstrate that HyPro consistently improves both average and final‑stage accuracy compared to strong baselines.
By HongWei Zhao (Beihang University), Rui Liu (Beihang University), Yong Chen (Beijing University of Posts,Telecommunications)
The paper introduces CG-OMatG, an equivariant Riemannian flow-based generative model that predicts molecular crystal structures using a coarse-grained, hierarchical representation. It treats molecules as rigid bodies, performs inter- and intra-molecular message passing, and learns to reconstruct molecule centroids, orientations, and lattice parameters conditioned on chemical species and conformer geometry. The model is trained on OMC25 and CSD datasets, fine-tuned with policy gradient reinforcement learning to favor low-energy structures, and validated against a CSP blind test benchmark using COMPACK packing-similarity analysis.
By Thomas Egg, Harry Winston Sullivan, Maya M. Martirossyan, Philipp H\"ollmer, Cheng Zeng, Adrian Roitberg, Mingjie Liu, Richard Hennig, Sapna Sarupria, Ellad B. Tadmor, Stefano Martiniani
The paper introduces a preemptive unlearning technique called gradient sealing to prevent large language models (LLMs) from acquiring forbidden capabilities during downstream fine‑tuning. It shows that existing retrospective unlearning methods fail to block hidden gradient pathways that can later enable illicit knowledge. By pushing relevant pre‑activations into the negative region, gradient sealing effectively blocks these pathways and demonstrates stronger resistance to downstream acquisition across multiple LLM families.
By Kemou Li, Qizhou Wang, Yue Wang, Fengpeng Li, Zhuan Shi, Negar Rostamzadeh, Golnoosh Farnadi, Masashi Sugiyama, Jiantao Zhou
The paper establishes a link between the adjoint equations of the Itô and rough path formulations of the Pontryagin Maximum Principle (PMP) for stochastic optimal control, showing that the Itô adjoint equals the conditional expectation of the rough adjoint. It introduces a rough stochastic PMP for adapted controls without using forward–backward SDEs, and then derives a unified PMP that connects the two frameworks via Itô–Stratonovich conversion and duality identities. Applications include rederiving the adjoint matching method for fine‑tuning generative models and proposing an indirect shooting method for certain feedback problems.
By Thomas Lew
The paper presents C2C (From Codebase to Culprit), a bug‑localization framework that narrows the debugging search space through semantic retrieval and hierarchical reinforcement learning. It first retrieves likely buggy files using contrastively fine‑tuned CodeBERT embeddings and then progressively localizes bugs from files to functions to lines of code. Experiments on Java and Python datasets show that this multi‑resolution approach improves both retrieval precision and localization accuracy, with ablation studies underscoring the importance of hierarchical decomposition and reward shaping.
By Ankur Garg, Corey Yang-Smith, Rishav Rishav, Ahmad Abdellatif, Samira Ebrahimi Kahou
The paper investigates a specific failure mode in vision‑language‑action models, termed instruction‑action binding, where models respond to language and vision separately but fail to combine them to select the correct action under counterfactual changes. Through behavioral analyses of fine‑tuned policies, the authors show that failed rollouts often preserve source behavior or switch to other demonstrated tasks, indicating that language is not ignored but mis‑bound. They propose Equivariant Counterfactual Training (ECT), which supplies counterfactual demonstrations and a paired loss to enforce correct action selection, achieving significant performance gains across simulated and real‑world benchmarks.
By Hung-Jen Chen, Yu-Hsun Hou, Yan-Hong Chen, Yan-Fu Chen, Binghua Cai, Min Sun, Chun-Yi Lee
The article outlines a Ph.D. research agenda aimed at creating provably efficient and practical algorithms for data‑driven sequential decision‑making under uncertainty. It focuses on reinforcement learning and multi‑armed bandits, targeting applications such as recommendation systems, computer networks, video analytics, and large language models. The work seeks to overcome limitations of existing methods—such as reliance on idealized models, lack of robustness to adversarial perturbations, and poor instance‑dependent performance—by developing algorithms that are more efficient, robust, instance‑adaptive, and generalizable to new environments.
By Zhiyong Wang