FinPILOT is a plugin inference‑time policy optimization framework that enhances pretrained reinforcement‑learning trading agents by leveraging a multi‑step price forecaster to generate imagined market trajectories. At each decision point, it optimizes the pretrained policy using the forecaster’s predicted price path before executing a trade, without altering the original training pipeline. Across five pretrained policies and multiple market settings, FinPILOT consistently improves total return and Sharpe ratio, with gains increasing as forecast accuracy improves.
By Eun Go, Rohan Deb, Arindam Banerjee
MIRAGE is a visual indirect prompt injection framework that targets multimodal large language model (MLLM)-based web agents. It operates under realistic constraints where an attacker controls only a semantically legitimate, spatially confined region such as an ad slot or widget. Using diffusion models and a robust optimization technique, MIRAGE generates perceptually benign adversarial images confined to the attacker’s boundaries, enabling next‑action hijacking while remaining stealthy.
By Xuelong Dai, Jianyu Ma, Boyang Ma, Biwei Yan, Yijun Yang, Yue Zhang
SWE‑Nexus is a fully automated pipeline that builds repository‑level implementation tasks by analyzing test execution. It uses four construction agents to select tests, trace production function bodies, and generate behavioral specifications, retaining scopes only when all expected scoring points pass and masking the traced bodies triggers failures. The benchmark covers 600 tasks from 105 repositories in six languages, and the best configuration resolves 37.8% of tasks in a single attempt, demonstrating the system’s ability to construct tasks from validated behavior beyond individual code changes.
By Songwei Yu, Ziqi Miao, Jingxuan Xu, Junhao Fan, Haoyang Huang, Yuan Xiong, Zihang Xu, Weihao Li, Wenqiang Zhu, Yuechen Luo, Xinshun Feng, Yanan Wu, Ken Deng
DialFork is a new benchmark that evaluates procedural hallucination in conversational agents, a failure mode where agents produce fluent, factually plausible utterances but deviate from the intended dialogue process. The benchmark consists of 36,278 dialogues across 24 tasks in 12 domains, created by forking conformant dialogues and injecting one of four hallucination types, with validation by 30 human annotators. Three evaluation tasks are defined—monitoring overall deviation, procedural retrieval of a single dialogue, and real‑time detection of off‑track turns—using dialog‑flow‑aware metrics FlowKL and FlowScore, which outperform baseline semantic similarity metrics.
By Sergio Burdisso, Esa\'u Villatoro-Tello, Dairazalia S\'anchez-Cort\'es, Yanis Labrak, Jordi Luque, Milos Cernak, Petr Motlicek
ReFlow is a system‑algorithm co‑designed framework for efficient reinforcement learning (RL) post‑training of large language models (LLMs). It introduces replay‑driven decoupling, allowing recent trajectories to be reused multiple times to amortize rollout costs while limiting policy staleness. The architecture features a dual replay cycle, an elastic resource pool, and an adaptive scheduler, achieving up to 2.06× higher training throughput compared to existing synchronous and asynchronous RL systems without sacrificing model performance.
By Yukun Tian, Junwei Su, Haofeng Xu, Lansong Diao, Zhengping Qian, Weibo Zhao, Wenchao Zhou, Ruiting Zhou, Chuan Wu
The paper investigates whether a frozen transformer can use another model’s KV cache as a substitute for its own, finding that this approach fails because the reader only accepts caches within a narrow tolerable set. Experimental results show that none of eight substitution methods improve performance over a baseline, while a different strategy—adding the sender’s cache as extra context and only adapting queries—does outperform the baseline. The authors conclude that pursuing faithful cache substitution is misguided and that augmenting the reader’s own cache is the more promising direction.
By Adam Deng, Shicheng Liu, Yada Zhu
The paper investigates how to efficiently share internal states between agents built from frozen language models. It proposes a shared pool approach where each agent publishes its internal state once, allowing peers to read it via a single mechanism, thereby avoiding the need for a dedicated transfer bridge for every agent pair. The study finds that augmenting the key‑value cache and replacing activations—rather than swapping them—yields better accuracy with minimal training, and that a single adapter can generalize across multiple publishers when their internal states are similar to the receiver’s.
By Adam Deng, Shicheng Liu, Yada Zhu
Neuroscience data are fragmented across labs and formats, making reuse labor‑intensive. This study benchmarks coding agents on eight recent mouse neural recording papers, asking them to reformat data for a common decoder task. While agents handle many subtasks, they rarely produce fully error‑free solutions; the authors pair outcome metrics with a process rubric to evaluate errors and propose agent‑assisted data‑sharing practices.
By Ling-Qi Zhang, Kristin Branson
The paper introduces Mentalized Adversarial Self-Play (MASP), a framework for training proactive dialogue agents by modeling the user as a rationally resisting participant with an evolving mental state. MASP separates teacher and student mentalization, alternates self‑play updates for system and user policies, and treats robustness to rational resistance as a training objective. Experiments on proactive dialogue benchmarks show MASP improves success rate, rationality‑constrained success, reduces average turns, and performs better against hard users compared to SFT and Planner‑only approaches.
By Zanwu Liu, Dan Zhang, Xiaodi Wang, Yongbo Mao, Xiaoyang Guo, Yonglin Wu, Peiyue Jiang, Yangyang Wei, Zecheng Lin
The paper surveys memory‑centric compression techniques for large language models, framing long‑context handling as a memory‑management problem. It distinguishes compute‑active memory from compute‑inactive memory, and models operations such as writing, offloading, recall, and forgetting as state transitions within an information lifecycle. The authors propose a unified cost model and argue that future research should focus on dynamic, lifecycle‑aware memory management rather than static compression.
By Chaogui Gou
The paper investigates why large language models (LLMs) miss evidence in multi‑hop retrieval‑augmented generation (RAG). It introduces ExecutionRAG, a training‑free register that tracks open sub‑questions and their supporting evidence, allowing the authors to distinguish between true retrieval failures and search drift. Using this method on 1,000 MuSiQue questions, the study shows a significant reduction in missed evidence but also reveals that other error types, such as premature resolutions and abstentions, increase, shifting the bottleneck from search guidance to evidence verification.
By Jiaqi Wei, Wenxuan Huang, Chengdong Xu, Xiang Zhang, Wanli Ouyang
The paper introduces Temporal State Tracking Memory (TEM), a training‑free framework that stores facts as versioned records linked to their sources and retrieves the relevant versions in source order, labeling historical records before passing them to the answer model. TEM incorporates event memory and original dialogue turns to fill gaps in structured facts. Evaluations on 420 questions from 600 sessions with 60 personas show high key‑entity hit rates (75.71–97.62%) and significant improvements when historical labels are added, while keeping evidence compact (94–118 tokens per answer).
By Pengfeng Song, Yunkai Zhuang, Liye Shi
VELD-QA is a large-scale benchmark for visual evidence grounding and question answering in long documents, comprising 15,000 human-authored question-answer pairs across 5,859 PDFs with 372,445 pages. The dataset includes detailed annotations of supporting pages, visual evidence elements, and descriptions of cited visual content, enabling separate evaluation of retrieval, localization, and interpretation. Experiments on nine vision-language models and various retrieval-augmented systems reveal significant gaps in evidence localization and visual interpretation, with oracle accuracy peaking at 81.7% and enrichment of visual-element annotations narrowing performance gaps by 17–24%.
By Zhivar Sourati, Marianne Menglin Liu, Rongguang Wang, Zheng Wang, Yazhe Hu, Lan Guo, Yuying Wang, Mengqing Guo, Bhavesh Dnyaneshwar Gore, Shubham Pathak, Ramapriya Ranganath, Sandeep Sharma, Tao Sheng, Sujith Ravi, Dan Roth
The paper proposes a new paradigm for Earth observation (EO) modeling in which a general-purpose coding agent automatically builds and recursively improves models by interacting with a harness and a verifier. Any verifiable EO task can be framed with a harness that defines permissible operations and a verifier that assesses success, enabling the agent to learn from labeled benchmarks, independent sensors, physical models, or repeated observations. On six diverse EO benchmarks—including optical, SAR, and multispectral time-series imagery—the agent-built models outperform state‑of‑the‑art methods on every task, achieving superior land‑cover mapping using only 11% of the training labels.
By Sijie Zhao, Zeqi Zhou, Heli Qi, Xuanchao Fu, Zhuo Zheng, Junshi Xia, Junjue Wang, Weihao Xuan, Naoto Yokoya
The paper introduces DeSpatial, a framework that separates spatial understanding into perception and reasoning. It uses a shared 3D perception tool with joint segmentation–depth co-refinement and cross-view instance fusion to produce a unified, metric, gravity-aligned 3D scene. DeSpatial achieves state‑of‑the‑art results on VSI‑Bench, MindCube‑Tiny, ViewSpatial, and VSI‑SUPER, demonstrating that the main bottleneck in spatial reasoning is precise 3D perception.
By Yifei Liu, Xiaolong Zhou, Ziyang Gong, Yifan Zhou, Yuanyuan Gao, Yuning Gong, Yuchen Yang, Hongjie Zhang, Xue Yang, Xiao Sun, Zhihang Zhong
VisualTrust introduces a counterfactual benchmark to evaluate how multimodal large language models handle on‑screen claims that may conflict with visual evidence. The benchmark contains 5,063 cases across 14 claim‑fact classes, each featuring a question paired with three video versions: no claim, accurate claim, and misleading claim. Results show a drop in accuracy from 90.1% to 51.9% when a misleading claim is present, and a new training‑free agent, DEAR, improves accuracy under misleading claims to 81.8–85.2% while reducing trap‑answer selection.
By Huachao Zhu, Junjielong Xu, Sicheng Li, Liang Yang, Mingtao Chen, Zhichao Hu, Li Jiang, Pinjia He
SpaceCache is a training‑free feature cache that reuses intermediate features from point‑cloud transformers across consecutive frames, allowing tokens of previously seen regions to skip transformer blocks and only compute new tokens. The method limits cache entry age and evicts low‑confidence entries to maintain reliability, and it operates once per encoder stage at negligible cost. Experiments on outdoor and indoor datasets (nuScenes, Waymo, ScanNet) with Utonia and LitePT‑B backbones show up to 1.65× speedup with less than 1‑point mIoU loss, outperforming token‑merging approaches.
By Dongheon Yang, Beomseok Kang, Jae-Joon Kim
The paper presents a method for synthesizing contrast‑enhanced MRI (CE‑MRI) from non‑contrast sequences using a novel Mixture of Feature Difference (MoFD) module. MoFD dynamically weights and aggregates features from individual sequences and their explicit differences, improving image fidelity and perceptual quality over baseline approaches. Experiments on UCSF‑PDGM and UPenn‑GBM datasets show that the synthesized images reconstruct tumor regions well and yield segmentation performance comparable to real CE‑MRI, with regression‑based training achieving higher structural fidelity.
By Jonghun Kim, Junmo Kwon, Hyunjin Park
The paper introduces Co‑VLN, a minimalist, model‑agnostic framework that lets agents in Vision‑Language Navigation share structured perceptual memory when they independently visit the same locations. By exchanging observations, each agent expands its receptive field without extra exploration. Experiments on the R2R benchmark with DUET and MapGPT show that vision‑sharing significantly improves performance, suggesting a promising direction for collaborative embodied navigation.
By Qunchao Jin, Yiliao Song, Qi Wu
PhyMix introduces a Physics Evaluator that checks geometric priors, contact, stability, and deployability through nine sub-constraints, establishing a benchmark for physical consistency in single-image 3D indoor scene generation. The framework uses implicit alignment via Scene-GRPO and explicit refinement through verifier-guided agentic repair to incorporate evaluator feedback during training and inference, improving physical plausibility while preserving scene content. Experiments demonstrate enhanced reconstruction quality and physical plausibility, with qualitative results on stylized and real-world images showing the method’s broader applicability.
By Dongli Wu, Jingyu Hu, Ka-Hei Hui, Xiaobao Wei, Chengwen Luo, Jianqiang Li, Zhengzhe Liu