MemCut introduces a benchmark to test user-requested forgetting in large language model agents, featuring 456 hand-designed secrets embedded in long-horizon conversations with fully annotated evidence trails. The benchmark defines two recovery metrics—native and residual—to assess how well secrets can be retrieved after deletion. Experiments across five memory designs and four LLMs show that deletion reduces native recovery more than residual recovery, highlighting candidate coverage as a key bottleneck.
By Cheng Zhang, Xuanning Cao, Zhicheng Zheng, Hui Xiong
The paper introduces a tool‑augmented large language model framework that improves the reliability of medical calculations derived from free‑form patient notes. It separates clinical language understanding from rule execution by retrieving calculator specifications, generating executable tools, and using a constrained extractor to map notes to calculator inputs. On the MedCalc‑Bench dataset, the method achieves 86.73% final‑answer accuracy, outperforming one‑shot prompting and fine‑tuned baselines by 24.0 and 17.27 percentage points.
By Yang Zhong, Ling Luo, Xuanlyu Zhu, Yanhua Wang, Yuanyuan Sun, Jian Wang, Hongfei Lin
The paper presents a privacy‑preserving text inference framework using CKKS homomorphic encryption for health‑misinformation classification and semantic retrieval. It combines BioBERT, MPNet, and TF‑IDF features for classification, achieving an accuracy of 0.971 and macro‑F1 of 0.950, while a late‑fusion similarity model reaches a Pearson correlation of 0.8691. Optimized ciphertext retrieval reduces latency from 222.028 s to 5.168 s and communication overhead from 6.11 GB to 2.13 MB over 7,128 items.
By Jasleen Kaur, Chlo\'e Cachet, Koray Karabina, Jean Belo Klamti, Philippe Lamontagne
RSIFlow is a recursive self‑improvement framework that combines adaptive workflow orchestration with a meta‑agent that evolves its own harness. It diagnoses performance bottlenecks from task‑agent trajectories and environmental feedback, then dynamically selects component‑specific improvement workflows. Experiments on six benchmarks and two backbone models show that RSIFlow achieves state‑of‑the‑art average performance, with ablation studies confirming the benefits of adaptive orchestration and meta‑harness evolution.
By Huan Zhang, Mingju Chen, Dongxu Zhou, Can Lv, Sen Cui, Faguo Wu, Shiji Zhou
Co2In is a bi‑level memory incremental learning framework that mimics human memory mechanisms. It uses a working memory to encode new knowledge and a long‑term memory to retain it, with a consolidation step that extracts patterns and compresses them into a compact representation. Experiments on CIFAR‑10, CIFAR‑100, and Tiny‑ImageNet demonstrate state‑of‑the‑art performance while keeping memory usage efficient.
By Wenju Sun, Qingyong Li, Boyang Li, Wen Wang, Siyu Zhang, Yi Wei, Yangliao Geng
arXiv:2610. 21808v1 Announce Type: new Abstract: Weight quantization below four bits shrinks memory and bandwidth at a perplexity that looks acceptable, but often at the cost of factual knowledge and task accuracy.
By Johannes Knittel, Alec Helbling, Hanspeter Pfister
CIM-Embed is a framework that learns instruction‑conditioned multimodal embeddings for clustering. It uses supervised contrastive learning with same‑group positives and negatives from other groups, and handles tasks with varying group counts via heterogeneous structured batches. At inference, the model maps an object’s image, text, and instruction to an embedding, and HDBSCAN clusters these embeddings without requiring a pre‑specified cluster count. The authors introduce ProductMIC, a benchmark of 5,373 Amazon products and 449 grouping tasks, and report a mean per‑task V‑measure of 95.01, outperforming the best general‑purpose baseline.
By Yinghua Xu, Pengyang Xie, Yanxiang Zhou, Chuanzhi Tong, Xiao Xiao, Chaoqun Feng, Xiaoqing Wang
OntoAct is a hypothesis‑based neuro‑symbolic framework that uses ontological abstraction to convert individual experiences into reusable axioms about entity types, relations, and conditions. A large language model proposes hypotheses and actions, while a symbolic verifier checks these against current observations, identifying missing evidence and updating assessments based on new data. The framework is evaluated on Tacit‑FLE, Mars, and ALFWorld, achieving leading task performance with efficient interaction and inference.
By Donghyeok Lee, Bong Gyun Kang, Junyong Ahn, Sungroh Yoon
The paper introduces PRM4RL, a framework that combines task planning and dense reward generation using large language models for reinforcement learning. It decomposes complex tasks into manageable subtasks and employs a fine‑grained progress function to monitor execution, which is then used to construct a Progress Reward Model for potential‑based reward shaping. Experiments on robotics control tasks show that PRM4RL outperforms existing LLM‑based planning and reward methods, achieving state‑of‑the‑art performance.
By Xiuhui Zhang, Ning Gao, Xingyu Jiang, Yihui Chen, Yuheng Pan, Mohan Zhang, Yue Deng
The paper introduces SAFE-PRO, a method for enhancing safety in multimodal large language models by proactively asking clarifying questions when cross‑modal ambiguity poses intent‑level risks. SAFE‑PRO employs a Risk‑Conditioned Uncertainty signal that blends predictive entropy with a learned cross‑modal risk score to guide both question‑asking and path optimization, prioritizing safety while reducing over‑refusal. The authors also present the SSUI‑Dialog dataset and demonstrate that SAFE‑PRO outperforms existing approaches, including GPT‑4o and Claude‑3.5‑Sonnet, across multiple benchmarks.
By Wei Cai, Jian Zhao, Yuchen Yuan, Xuelong Li
The paper introduces Shared Residual Distillation (SRD), an on‑policy self‑distillation technique that learns from a language model’s own unverified peer solutions without requiring external supervision. SRD compares how peers alter the prediction distribution relative to a control, keeping only adjustments that move in the same direction and are conservative in magnitude. Experiments show that SRD significantly boosts performance on the AIME24–26 benchmark for Qwen3 models while preserving general‑capability performance.
By Feiyang Chen, Runze Ma, Zichen Wang, Jiong Lou, Zihan Wang, Shuangjia Zheng, Jie Li
The paper introduces CART, a method for Multimodal Continual Instruction Tuning that coordinates adaptation between a projector and a language model by using task‑specific experts and weighted mixtures. CART jointly infers expert weights based on relevance and cross‑module compatibility, and it controls model growth through relevance‑guided expert consolidation. Experiments show that CART achieves state‑of‑the‑art performance on multimodal continual learning tasks.
By Ling-Xiang Zhang, Tao Hu, Da-Wei Zhou
The paper introduces the Language Diffusion Transformer (LangDiT), a new transformer architecture tailored for continuous diffusion language models (CDLMs). By incorporating feature and query-key normalization, wider noise conditioning, likelihood-shortcut gradient routing, calibration, and confidence-gated training self-conditioning, LangDiT achieves lower generation perplexity than existing models across various CDLM formulations. Experiments on the OpenWebText dataset show state-of-the-art perplexity scores, demonstrating the architecture’s effectiveness while preserving the underlying model representations and training objectives.
By Weijian Luo, Yujian Chen, Hao Jiang, Guang Cheng, Tianyang Hu
KULTURE Bench is a new evaluation framework created to assess language models within the Korean cultural context. It includes datasets of cultural news, idioms, and poetry and tests models at word, sentence, and paragraph levels for cultural comprehension and reasoning. Experiments with models trained on different corpora show that current models still lack deep understanding of Korean cultural texts.
By Xiaonan Wang, Jinyoung Yeo, Joon-Ho Lim, Hansaem Kim
CoTS: Competitive Token Selection for Latent Reasoning in Large Language Models proposes a training‑free framework that replaces binary latent reasoning decisions with continuous token‑level allocation. By using the probability competition between the two most likely tokens, CoTS dynamically balances discrete commitment and soft exploration at each reasoning step, eliminating the need for manually tuned thresholds or additional forward passes. Experiments on six reasoning benchmarks show that CoTS consistently outperforms standard Chain‑of‑Thought and other state‑of‑the‑art training‑free latent reasoning methods.
By Jie Yu, Yuxi Sun, Zhaoxu Li, Chunxiao Liu, Xudong Jiang
INFUSE is a new benchmark that tests indirect prompt injection (IPI) in large‑language‑model agents performing long‑horizon workflows. It contains 40 realistic tasks and 56 attack scenarios across five threat categories, with attacks injected through realistic interaction surfaces rather than just initial prompts. The study evaluated 19 LLMs, generating over 41,000 adversarial trajectories and found mean attack success rates of 19.6% (LLM evaluation) and 15.4% (heuristic verification), while prompt‑level hardening reduced success from 58.2% to 14.3%.
By Thuy Dung Nguyen, Jingyi Guo, Rana Seda Dubauskas, Sreepriya Damuluru, Taylor T. Johnson, Kevin Leach
CoTranslator is a method that converts reverse chain-of-thought (CoT) responses into verified forward CoT responses, enabling efficient large-scale generation and verification of training data for reasoning distillation. Trained on a small set of paired reverse and rejection-sampled forward responses, it improves student model accuracy on mathematics benchmarks by up to 5.82 percentage points and retains 95.0% of rejection sampling’s accuracy while using only 35.9% of its compute. The translator also transfers to new datasets, teachers, and STEM problems, boosting average STEM accuracy for 4B and 8B students by 4.88 and 4.31 percentage points, respectively.
By Zinan Tang, Zhuoshi Pan, Kairui Chen, Ze Yan, Shaomian Zheng, Yukun Zhang, Gang Wang, Biqing Huang
ReasonUni introduces a reasoning-augmented reconstructive post‑training framework for unified multimodal models (UMMs) that couples image‑to‑text understanding and text‑to‑image generation. The framework first reasons about the desired edit and its target state, then generates the target image conditioned on this reasoning output. A tree‑structured reinforcement learning method, TreeGRPO, jointly optimizes the reasoning and generation modules, yielding consistent performance gains across multiple benchmarks.
By Feize Wu, Zhiyuan Yan, Yuhuan Hong, Yuwei Niu, Zongjian Li, Zheyuan Liu, Bin Lin, Jinbo Yan, Jie Qin, Xudong Mao, Limeng Qiao, Guanglu Wan, Li Yuan
SHRED is a framework that trains vehicles, pedestrians, and cyclists through large‑scale heterogeneous self‑play from random initialization, without using expert demonstrations or logged trajectories. By combining type‑specific dynamics, procedural scenario generation, and per‑agent randomization, SHRED allows diverse, realistic behaviors to emerge through closed‑loop interaction. In benchmarks such as the Waymo Open Sim Agents Challenge and InterPlan, SHRED outperforms demonstration‑free baselines and prior self‑play methods, showing that heterogeneity in self‑play improves both traffic realism and planning performance.
By Wei-Jer Chang, Zhouchonghao Wu, Akshay Rangesh, Weixin Li, Masayoshi Tomizuka, Wei Zhan
TGrasp6D is a vision‑language method that generates task‑oriented 6DoF grasps in a zero‑shot setting. It uses a single RGB‑D image and a task description to infer a task‑informed prior, then refines this prior with pose optimization to avoid collisions while maintaining task consistency. The method achieves a 94% success rate on the TaskGrasp dataset, outperforming the best baseline of 74%, and is demonstrated on a Franka Research 3 robot for real‑world zero‑shot generalization.
By Manav Kulshrestha, Apoorva Vashisth, Damon Conover, Aniket Bera