The paper reports an entry to the EgoLongQA track of the Wearable‑AI Challenge at ECCV 2026 that achieved first place in the ≤2 B parameter division with a score of 0.8279 on the held‑out test set. The system is a single 2 B vision‑language model that answers multiple‑choice questions about ten‑minute egocentric videos in one greedy forward pass, obtained by distilling the junior perception module of a tool‑using agentic pipeline into a small student using teacher traces that answered correctly. By pruning the multilingual embedding table from 248,320 to 143,469 rows, the authors reduce the model to 1.9985 B parameters while maintaining identical logits on retained rows, enabling the model to reach 89 % of the accuracy of the large agentic pipeline with only 1.1 % of its parameters.
whyItMatters":"The work demonstrates that a highly efficient, sub‑2 B parameter vision‑language model can match the performance of a much larger agentic pipeline on long‑video perception tasks, showing the feasibility of deploying such models in resource‑constrained settings."
By Logesh Kumar Umapathi
The paper introduces NetArtifactBench, a benchmark designed to evaluate whether AI agents can detect and repair inconsistencies in network experiment records while preserving supported claims. It tests 23 agent configurations on 52 instances with injected inconsistencies, finding an average pass rate of 65.3 % but no runtime exceeding 30 % for complex repairs that require recovering implicit relations and propagating changes across artifacts. The results highlight a clear distinction between local corrections and full record-level repair, leading the authors to argue that artifact integrity should be a primary design and evaluation criterion for AI agents in network systems.
By Tianzhu Zhang, Weichen Tao, Changgang Zheng, Yusheng Zheng, Long Chen, Xiaoyi Fan, Meikang Qiu
The paper introduces CEAE, a hierarchical agentic framework that transforms natural-language instructions into machine-executable GeoJSON edits for urban renewal. It breaks down instructions into hierarchical geometric intents and applies edits from coarse to fine, using a self-reflective execution-validation loop to maintain spatial consistency. Experiments demonstrate that CEAE achieves higher execution validity, robustness, and geometric accuracy compared to baseline methods.
By Rui Liu, Steven Jige Quan, Zhong-Ren Peng, Zijun Yao, Han Wang, Pengyang Wang, Kunpeng Liu, Yanjie Fu, Dongjie Wang
MC-DeTra is a reimplementation of the DeTra model that jointly performs object detection and socially-aware trajectory forecasting in bird's-eye-view images. It introduces motion-consistency mechanisms that add supervision from each actor’s past motion, surrounding traffic occupancy, and a consistency constraint aligning predicted heading with motion direction. The added losses are train‑only and inference‑safe, improving dynamic trajectory forecasting on the Waymo Open Dataset while maintaining or enhancing detection accuracy.
By Vladislav Diuzhev, Dmitry Yudin
The paper examines whether internal representations of agentic systems can better indicate task success than traditional confidence measures. It introduces two methods—Latent Trajectory Dynamics (LTD) and Action Representation Probe (ARP)—that analyze changes in residual-stream representations and action-level representations, respectively. Experiments on Bash, SQL, and Python benchmarks with Qwen and DeepSeek models show these methods outperform conventional surface-level and sequence-based calibration baselines, offering a zero‑overhead reliability monitor without prompt changes or multiple rollouts.
By Priyanka Mary Mammen, Emil Joswin, Srujananjali Medicherla
The paper introduces a four‑layer architecture for Cognitive Digital Twins (CDTs) that extends traditional Digital Twin systems by embedding cognitive capabilities. It outlines a self‑evolving closed loop across physical, digital‑twin, cognitive, and task layers, enabling task‑specific cognitive models and decision generation that are refined through operational feedback. Two operation modes—user‑request‑driven and self‑driven cognition—are described, along with key enabling mechanisms and a lightweight simulation demonstrating closed‑loop feasibility and efficiency gains.
By Haoran Gao, An Li, Zhen Li, Jun Cai
The paper investigates how procedural memory behaves when its stored routines become mismatched with changing interfaces. Using a BrowserGym TimeWarp case and controlled synthetic shopping tasks, the authors found that procedural memory can be mismatched without causing behavioral disruption, as none of the predefined interference signatures appeared. The study highlights a specific region of non‑interference but leaves open which conditions would turn mismatches into observable errors.
By Yanze Cao
AgentAudit is an open, extensible framework that evaluates the full lifecycle of AI agents, assessing planning, tool selection, execution, memory, and reasoning across ten dimensions such as instruction integrity, security, and alignment. Unlike existing benchmarks that focus on single aspects, AgentAudit analyzes the entire execution trace to attribute failures to specific stages. The framework was tested on five large language models, revealing significant differences in trustworthiness even among models with similar task‑completion performance.
By Shrey Nag, Sachita, Abhishek Kumar Singh, Lipi Goel, Rajeshwar Singh Janwar
Auto-RecSys is an autonomous research system designed to scale long-horizon experimentation for industry‑scale recommendation models. It tackles long feedback loops and system complexity by enabling distributed asynchronous execution, centralized cross‑server memory, and a cognitive‑procedural separation that combines natural‑language skill files with deterministic scripts. The system incorporates a dual‑loop self‑evolving architecture—Execution Evolution and Idea Evolution loops—to refine operational playbooks and guide future experiments, thereby reducing human effort per cycle and improving reliability as playbooks mature.
By Ming Li, Dai Li, Xuying Ning, Bo Sun, Rui Li, Yi Zhang, Silvia Gong, Xuan Cao, Rui Li, Cornelia Carapcea, Qunshu Zhang, Zhigang Wang, Yinglong Xia, Andy Wang
The paper investigates whether a language‑model agent can autonomously study an unfamiliar environment without prior task instructions or examples, and decide how to prepare for future tasks. It formalizes task‑agnostic environment preprocessing, where a studying system explores under a budget to produce reusable artifacts for a later solver. Experiments on six diverse benchmarks show that a meta‑agent variant often outperforms fixed methods, though larger budgets do not consistently boost downstream reward, yet the artifacts still reduce test‑time sampling needed to achieve a target score.
By Vinay Samuel, Varun Ursekar, Vijay S. Kalmath, Apaar Shanker, Veronica Chatrath, Yuan Xue
The paper introduces SurgicalRoomAgent, a voice‑interactive multi‑agent system for smart operating rooms that leverages large language models to understand natural language, control devices, record intraoperative events, and generate surgical reports. Its layered architecture includes a voice interaction pipeline (wake, ASR, turn detection, agent reasoning, TTS) and an agent core (skill registry, task planner, device manager). Three key technologies—KV Cache prefix warming, streaming partial JSON parsing with early parallel task execution, and progressive skill prompt disclosure—reduce latency and maximize context efficiency, enabling real‑time operation within a 16,384‑token limit.
By Tianxiang Zhou
The paper reports a data‑efficient language modeling study conducted by Qiushi Engine on the BabyLM 2026 Strict‑Small benchmark, using only 10 million corpus words and 100 million cumulative presentations. It describes a three‑stage research program: Stage I built a frontier model via compact restatements and incremental learning; Stage II identified that exact repetition versus aligned restatement affect context use and proposed a principle for organizing experience around contextual dependencies; Stage III applied selective supervision and preservation techniques, achieving a modest overall score increase from 42.02 to 42.25 and the highest public Strict‑Small result as of 8 September 2026. The work also discusses further studies on compression, relational anchors, shared representations, and measurement, and makes models and code publicly available.
By Shuxing Yang, Kaihao Zhu, Junjie Yang, Rui Zhao, Junyao Wu, Yize Wang, Wenhao Li, Fujia Chen, Taowen Deng, Shenzhan Hong, Yaqi Li, Zichen Li, Jincheng Mi, Yuang Pan, Hongsheng Chen, Yihao Yang
The paper announces VectraYX‑Vision‑1B, a Spanish/LATAM cybersecurity vision‑language model that uses a natively‑trained visual tower (Qwen2‑VL) and a frozen decoder. The authors report that transplanting the visual tower improves the failing 8‑nibble address field from 0.00 to 0.81 exact, using a coarser token budget than 2×2 tiling, and that a second pre‑registered field shows 63% contamination and is demoted. The model’s B6/B7 tool‑id remains at floor, and code/checkpoints are available on Hugging Face.
By Juan S. Santillana
MindTopo is a benchmark that tests foundation models on topological reasoning, covering five cognitive properties—continuity, separation, order, enclosure, and knots—across two cognitive levels: reasoning and planning. It contains 11,030 instances from 13 procedurally generated task types, and evaluates 14 multimodal large language models, including agent configurations with image and video generation. Results show that models perform better on reasoning than planning, and even the best model lags far behind human performance, with fine‑tuning and reinforcement learning improving reasoning more than planning.
By Yunfei Ge, Anbang Liu, Qineng Wang, Johnalbert Garnica, Jianwen Lyu, Zihan Wang, Reuben Tan, Jianfeng Gao, Ruohan Zhang, Yining Hong, Jiajun Wu, Manling Li
The paper introduces T1, a 122‑billion‑parameter Mixture‑of‑Experts model trained with reinforcement learning to perform long‑horizon terminal tasks such as coding and scientific discovery. T1 operates a real shell in a cloud sandbox, making over 300 tool‑call turns per task and receiving rewards from task‑specific verifiers. The authors detail a training recipe that includes aggressive warm‑starting, TITO construction with drift repair, and rollout‑routing replay, achieving significant performance gains on Terminal‑Bench 2.1 and surpassing GPT‑5.4 and GLM‑5.1 on the Long‑Horizon Terminal Bench.
By Junyao Yang, Yucheng Shi, Zhongzhi Li, Ruhan Wang, Zongxia Li, Haitao Mi, Leowei Liang
VikingRAG is a directory‑aware semantic data management system that reduces token usage in retrieval‑augmented generation by tightly integrating semantic and structural access. It employs multi‑round retrieval traces as reusable experience edges and an adaptive escalation strategy to avoid unnecessary multi‑round exploration. Experiments show that VikingRAG achieves comparable accuracy to state‑of‑the‑art methods while using only 11.6%–51.9% of their tokens, and further reductions to 5.1%–32.5% with trace reuse and escalation.
By Peiyuan Gao, Gaoyuan Zhang, Haojie Qin, Yahui Sun, Qianyi Zhang, Yunhao Zhang, Zeyu Wang, Wei Lu
OpenResearcher is a fully open, reproducible pipeline for generating long‑horizon deep research trajectories that interleave search, evidence aggregation, and multi‑step reasoning. It decouples corpus bootstrapping from trajectory synthesis and runs the search‑and‑browse loop offline using three browser primitives over a 15M‑document corpus. Using GPT‑OSS‑120B as a teacher, the pipeline produced over 97K trajectories, enabling a 30B‑A3B model to achieve 54.8% accuracy on BrowseComp‑Plus and providing insights into pipeline design through controlled analysis.
By Zhuofeng Li, Dongfu Jiang, Xueguang Ma, Haoxiang Zhang, Ping Nie, Yuyu Zhang, Kai Zou, Jianwen Xie, Yu Zhang, Wenhu Chen
Valerant is a training‑free framework that turns a pretrained action‑conditioned world model into a world action model (WAM) capable of exploring and building 3D game maps. By combining predictive visual rollouts with SLAM‑based spatial reconstruction and exploration‑driven action selection, it converts a single image into a persistent, navigable 3D geometry. This approach extends WAM‑based interaction beyond 2D visual simulation and reduces manual effort in 3D game‑map creation.
By Yiran Qiao, Feng Wang, Jing Ma
RobustSGPO extends semantic‑gradient‑based prompt optimization (SGPO) by explicitly defining the edit scope, constructing and validating patches, and allowing search to resume from either the current or retained snapshots. In the AgentX brainstorming workflow, experiments on 120 tasks and 95 runs show that periodic permission scheduling improves test scores, while RobustSGPO raises completion rates on held‑out tasks from 60.0% to 80.0% and test quality from 3.77 to 4.14 within a 20‑million‑token budget. The approach also demonstrates that category‑based retention mitigates source‑task degradation after a shift, and that search‑space control enhances quality through executable edits and alternative starting points.
By Zibo Zhao, Jijun Shi, Mo Zhou, Zhongyuan Wang, Shifu Bie, Yunfei Zhang, Xuanting Zhou, Xiangyu Wu, Bin Liu, Ruiming Tang, Wenwu Ou, Kun Gai
Recursive Code World Models (RCWM) present a framework that reconstructs complex 3D worlds from a single reference image by combining a Recursive Scene Program (RSP) representation with a construction solver that recursively calls itself. The RSP encodes the world as compositional scene code, while each solver invocation follows a global‑local‑global recursion: it first establishes the whole scene, then recursively reconstructs unresolved parts, and finally revisits the whole to refine composition. This approach preserves scene‑wide geometry and relationships, propagates a shared camera projection across levels, and uses a vision‑language coding agent to compare reference images with scene renders, enabling fine‑scale structure refinement and improved reconstruction over prior methods.
By Zhiqi Li, Yuxuan Liao, Bo Zhu