The paper surveys the emerging field of self‑improving AI systems that refine their behavior during deployment. It introduces a unified framework called feedback‑driven Test‑Time Intelligence (TTI) to connect two previously separate research directions: model state modification via test‑time signals and prediction enhancement through additional inference resources. The survey reviews key methods, applications across vision, language, multimodal learning, generative models, robotics, and healthcare, and outlines open challenges and a research roadmap.
By Shuaicheng Niu, Guohao Chen, Yaofo Chen, Zhiquan Wen, Jinwu Hu, Zeshuai Deng, Deyu Chen, Shuhai Zhang, Renjie Chen, Zihao Lian, Shoukai Xu, Gang Dai, Yunbei Zhang, Wei Luo, Yifan Zhang, Mingkui Tan, Cheng Deng
The paper introduces Athena for Cloud Knowledge Base, a fully offline multi‑agent Retrieval Augmented Generation (RAG) system tailored for Vodafone Idea’s Telecom SNOC environment. It combines dense E5 Large V2 embeddings, BM25 sparse retrieval, and Knowledge Graph expansion, fusing results with Weighted CombSUM, cross‑encoder reranking, and MMR to produce a diverse evidence set. The framework further verifies each chunk with an LLM before response generation, discarding weak evidence and falling back to multi‑chunk evaluation when necessary, achieving an MRR@10 of 0.910 and an EM of 78.4% on a 4,200‑document corpus.
By Harish Saragadam, Sudhanshu Sharma, Ipsha Routray
TAPVid-MV is a new benchmark for tracking any point in 3D across multiple synchronized camera views. It comprises 284 sequences, 1,142 calibrated camera streams, and 109,769 point tracks, covering indoor and outdoor domains and derived from various modalities such as depth, LiDAR, SLAM, and simulation. The dataset is visually verified, and evaluation shows that current multi‑view trackers do not consistently outperform monocular trackers, highlighting geometry recovery as a key bottleneck.
By Skanda Koppula, Frano Rajic, Abdullah Faiz Ur Rahman, Yi Yang, Ignacio Rocco, Jeet Thakwani, Rishabh Kabra, Andrew Zisserman, Joao Carreira, Siyu Tang, Carl Doersch, Gabriel Brostow
The paper investigates how Vision‑Language‑Action (VLA) models can generalise across different driving environments and camera setups. It introduces a multi‑dataset training strategy and an auxiliary objective called BEV‑Forcing, which injects bird‑eye‑view spatial information into the VLA backbone to improve both in‑distribution and out‑of‑distribution performance on a limited number of camera rigs. The authors observe that while BEV‑Forcing helps when training data is scarce, its advantage diminishes as the number of training embodiments grows, suggesting that scaling diversity may reduce the impact of such auxiliary tasks.
By Caio Azevedo, Stefano Sabatini, Sascha Hornauer, Fabien Moutarde
The paper introduces a method for monitoring web agents without relying on internal signals such as token logits. It proposes two observable trajectory representations—Macro features that capture cross‑step agent–environment interactions, and Micro features that assess consistency of intention, action, and expected state change via repeated black‑box queries. By labeling the first uncorrected critical error that leads to final failure as a key‑step boundary, the approach preserves valid early prefixes of failed trajectories and achieves risk prediction performance competitive with internal‑signal baselines across WebArena‑Lite and Online Mind2Web benchmarks.
By Sitong Pan, Yipeng Shen, Yilin Lu, Caiwen Ding, Lu Cheng, Qianwen Wang
DiffuSearch is a hybrid trajectory planner for autonomous driving that unifies objectives across both generation and refinement stages. It first uses a guided diffusion model to produce scene-consistent joint trajectories, then refines them with a Monte Carlo Tree Search that shares the same driving goals—collision avoidance, drivable area compliance, comfort, and progress. Experiments on nuPlan and interPlan benchmarks show that this synergy reduces collisions and improves comfort, especially in complex interactive scenarios.
By Steffen Hagedorn, Aron Distelzweig, Alexandru P. Condurache
CrashDiffuser is a closed-loop VLM‑guided diffusion framework designed for fine‑grained safety‑critical traffic scenario generation. It separates semantic collision reasoning from trajectory synthesis via a hierarchical collision‑intent interface that specifies target contact regions (head, rear, or side). The system uses a vision‑language model to extract scene context and predict structured action tuples, which condition a diffusion model to produce executable adversarial trajectories, achieving high target‑collision and contact‑region control rates on WOMD‑derived scenarios.
By Shucheng Zhang, Yuang Zhang, Bingzhang Wang, Muhammad Monjurul Karim, Kehua Chen, Yinhai Wang
PaperCompiler is a framework that translates research papers into repository-level code by compiling paper-grounded evidence into explicit implementation specifications. It preserves source provenance, distinguishes between paper-supported, inferred, externally delegated, and unresolved information, and encodes requirements such as non-degradation, ownership, cross-file dependencies, and file-level constraints. The framework improves reference-based fidelity by 13.8% and cuts high-severity evaluator critiques from 13.2% to 6.1% on Paper2CodeBench.
By Yunhao Liu, Hong Phuc Pham, Jaehong Yoon
The paper introduces Representational Empowerment (RepEmp), a metric for selecting representational elements that enhance an agent’s future modeling and planning abilities. Using a hierarchical Curator-Actor architecture, RepEmp is tested in three experiments: a closed‑vocabulary causal‑learning task, simulations showing improved structure recovery and transfer, and an open‑vocabulary planning domain where an LLM‑augmented Curator builds compact symbolic libraries. Ablation studies confirm that RepEmp is essential for these gains, positioning it as a key principle for continual model construction under resource constraints.
By Fei Dai, Hanqi Zhou, Alison Gopnik, Charley Wu
The paper argues that in multi‑turn agentic reinforcement learning, credit assignment should be viewed as a coverage problem rather than a targeting problem. It introduces verifier information density (V_d) as a structural metric, showing that terminal‑state verifiers operate in a low‑V_d regime where targeting fails. Experiments on tau^2‑bench, BFCL, and ToolACE‑2‑8B demonstrate that uniformly distributing reward across all turns outperforms sparse, targeted rewards, and that full chain coverage is necessary for optimal performance.
By Chenyu Zhou, Qiliang Jiang, Shuning Wu, Xu Zhou
The paper introduces TRACE, a Transparent Reasoning Architecture for Credible Execution, which provides an explainable AI-based decision framework for autonomous robots. TRACE structures decision-making into four auditable layers—Semantic Perception, Belief Reasoning, Action Synthesis, and Execution Verification—to ensure every action can be traced back to sensor evidence through documented causal chains. Experimental results on warehouse robot navigation show high evidence traceability (98.6%), temporal continuity (99.0%), and decision reconstructability (98.1%) across 500 simulated decision cycles.
By Cagri Temel
ToolGate is an executable acceptance pipeline designed to streamline the creation of scientific benchmarks that rely on specialist software. It evaluates each model-generated item through three gates: (1) an executable solution script must reproduce the proposed answer, (2) a randomized no‑tool screen rejects items solvable without the software, and (3) a tool‑using agent must solve the item within a time limit. In a FEniCSx instantiation, 500 generation attempts produced 128 unique, verified benchmark items after successive filtering.
By Ke Zhang, Yankang Liu, Roya Zandi, Maziar Raissi
CHIME introduces a credit‑aware hierarchical memory evolution framework that separates planning and execution experiences into distinct memory banks. By attributing each task outcome to the plan, execution, both, or neither before memorization, CHIME mitigates bias from noisy final outcomes and improves long‑horizon agent planning. Experiments on four benchmarks demonstrate that CHIME outperforms existing training‑based and self‑evolving memory methods, requires fewer memory items, and transfers effectively across backbone models.
By Yongshi Ye, Tian Lan, Feihu Jiang, Muyang Ye, Bin Zhu, Qianghuai Jia, Longyue Wang, Zhao Xu, Weihua Luo, Xiaodong Shi
The paper introduces SCOPED‑Hiring, a process‑aware fairness diagnosis pipeline for large language model (LLM) based multi‑agent hiring systems. It generates controlled resume variants, runs role‑based hiring committees, and logs over 311,000 structured decision trajectories, converting them into quantitative fairness signals across six diagnostic lenses: final outcome, counterfactual, process, pathway, dynamic, and design effects. The study finds that balanced final hire rates can conceal hidden trajectory unfairness—such as career gaps, proxy cues, and identity cues—and demonstrates that targeted repairs guided by these diagnoses can reduce the total layered burden by 72.3% while only slightly altering the hire rate.
By Yiran Zhao, Lu Zhou, Liming Fang, Yufei Chen, Jiafei Wu, Zhe Liu, Xiaogang Xu
MASkills is a continual learning framework designed to enhance multi‑agent large language model (LLM) systems by optimizing their agent skills. It introduces a new agent‑optimization pipeline that combines skill‑conditioned credit assignment, hierarchical credit aggregation, and momentum‑smoothed optimization, allowing skill libraries to evolve through refinement, induction, consolidation, and pruning. Experiments on HotpotQA, LoCoMo, and GAIA demonstrate its effectiveness across multiple agentic tasks.
By Huaiyuan Yao, Xiaoou Liu, Charles Fleming, Tianlong Chen, Hua Wei
The paper introduces persistent discovery context, a lightweight memory layer that stores prior intent-to-object mappings for data‑centric agents. By reusing these mappings, the system improves retrieval quality over metadata‑only search across three structured data environments, remains effective with automatically generated memories, and reveals a reproducible interference failure mode. In lexically sparse domains, memory‑only retrieval can even outperform metadata‑based retrieval.
By Jalal Mahmud
The paper investigates how human interventions at specific fault points—moments when an AI agent’s reasoning is most vulnerable—affect the diagnostic accuracy of multi‑agent medical systems. Using the MedQA dataset, the authors found that correct interventions can boost baseline accuracy by up to 40%, whereas incorrect or bias‑related interventions can reduce performance by up to 6% and increase diagnostic drift and uncertainty. The study also highlights behavioral parallels between cognitive biases observed in simulated agent conversations and real‑world clinical practice, such as premature closure and susceptibility to misleading cues.
By Benjamin C Liu, Dillon Mehta, Rishi Malhotra, Adam Zobian, Yong Ying Tan, Samir Chopra, Daniella Rand, Natalie Pang, Abhiram Gudimella, Kevin Zhu
The paper introduces Potential-Guided Policy Optimization (PGPO), a method for multi-turn agentic tasks that improves credit assignment by estimating empirical state potentials from anchor-state-group return statistics. PGPO derives action advantages from potential differences between adjacent states, enabling cross-trajectory credit propagation and finer-grained step-level credit assignment, especially within failed trajectories. Experiments on ALFWorld and WebShop demonstrate strong performance compared to recent group-based reinforcement learning methods, with negligible training overhead.
By Yuyao Zheng, Haipeng Sun, Junwei Bao, Lemao Liu, Hongfei Jiang, Yang Song, Dejing Dou
The paper introduces two methods to make alignment evaluations more realistic: critique refinement, which adds inference-time compute to generate and refine candidate actions, and DISH, a deployment-imitating harness that narrows the gap between simulation and real deployment. Experiments on multiple target models show that combining both techniques yields greater realism improvements than using either alone. The study demonstrates that automated approaches can enhance evaluation realism more efficiently than simply extending audit duration.
By Axel Ahlqvist, Richard Guan, Juan-Pablo Rivera, Adeline Kassler, Dmitrii Troitskii, Alexandra Souly, Kai Fronsdal, Robert Kirk, John Hughes
The article discusses how generative AI is reshaping the creation and circulation of cultural artifacts, prompting debate over whether these tools enrich or impoverish culture. It distinguishes between novelty—an attribute of individual artifacts—and diversity—an attribute of populations—arguing that creativity should be viewed as a property of hybrid collectives composed of people and algorithms. The authors find that AI-assisted ideation increases the novelty of individual outputs while potentially narrowing overall diversity, but that mixed human–machine groups can outperform single-type groups and sustain machine-generated solutions within human culture, depending on the composition and connectivity of the agents involved.
By Mason Youngblood, Katie Mudd, Manuel Anglada-Tort, Cameron Jones, Elena Miu, Diana Omigie, Margaret Schedel