FloorSAV is a framework that enhances audio‑visual large language models (AV‑LLMs) by embedding a dynamic 2D floormap derived from 3D point clouds, camera trajectories, spatial audio cues, and object landmarks. The floormap is fed into the AV‑LLM as a synchronized stream with egocentric video, enabling the model to jointly reason over visual, auditory, and geometric information in a single inference. The authors also introduce SAVED‑Bench, a benchmark for spatial reasoning tasks, and show that FloorSAV improves AV‑LLM performance on both SAVED‑Bench and SAVVY‑Bench, with ground‑truth floormaps highlighting its potential.
By Kyeong-Rae Kim, Sungnyun Kim, Tae-Hyun Oh
The paper introduces QGMem, a memory framework for conversational agents that organizes dialogue history into event-indexed atomic units and consolidates related units into dynamic memory traces. When a query is posed, QGMem retrieves complementary candidate memories, reranks them to form a compact working memory, and represents this working memory as a local graph to expose relational dependencies. Experiments on six benchmarks demonstrate consistent improvements in retrieval, multi‑hop evidence composition, conflict resolution, and ultra‑long dialogue reasoning while keeping context size and inference cost moderate.
By Yichen Liu, Chunfeng Yuan, Haowei Liu, Wenjuan Li, Zefeng Lin, Bing Li, Xu Chen, Weiming Hu
The paper evaluates how well voice user simulators generate realistic speech naturalness behaviors such as disfluency, interruption, and backchanneling. By auditing the tau-Voice LLM-based prompting approach and a rule-based injection algorithm, the authors find that prompting is unreliable and produces inconsistent, less natural speech, whereas the rule-based method yields controllable and diverse behaviors that better align with natural speech patterns. The study highlights the need for linguistically informed deterministic or specialized models and stresses the importance of auditing realized behaviors rather than relying solely on configured settings.
By Riqiang Wang, Elena Khasanova, Harsh Saini, Lex Konnelly, Parsa Kavehzadeh, Matthias Lee, Mohamed Attia
The paper introduces CogMem, a cognitive memory architecture for long‑term dialogue agents that uses a PEC$^2$F graph schema to separate claims, facts, and events. Dialogue turns are incrementally turned into provenance‑aware graph records, consolidated into higher‑level facts, and reconciled into temporally scoped claim views when conflicting updates arise. Retrieval is performed by a rule‑based controller that applies deterministic graph operators—anchoring, traversal, intersection, and evidence grounding—guided by LLM intent parsing, achieving strong performance on multi‑hop, temporal, and knowledge‑update tasks in LoCoMo and LongMemEval benchmarks.
By Zirui Liao, Zhengxian Wu, Zhuohong Chen, Yunyao Yu, Xiaoyu Liu, Yifan Xu, Haoqian Wang
The study investigates how generative agents in simulated societies revise their stances and reorganise interactions. By combining stance-change metrics, source evaluations, and temporal network analysis across multiple GPT‑4o runs, the authors find that environmental rational persuasion leads to the largest overall stance shift, while economic emotional persuasion produces the highest rate of low‑trust stance changes. In both experimental settings, agents often partially accept new positions yet maintain low trust in the source, and sustained coordination can coexist with ongoing disagreement.
By Hanzhong Zhang, Siyang Song, Jindong Wang
The paper introduces WorkForge, a scalable synthesis framework that builds verifiable work‑agent environments from real‑world resources. By extracting factual anchors from expert workflows and associated files, WorkForge generates task instructions, solution plans, and verifiers that remain traceable to observable workspace evidence. The authors constructed 16.7K environments across 40 professional domains, demonstrating significant performance gains for large language models such as Qwen3.5-35B-A3B-Base and Qwen3.5-27B.
By Jiazheng Zhang, Long Ma, Yunxian Yang, Zhiheng Xi, Zhikai Lei, Yajie Yang, Chenyang Liao, Enyu Zhou, Yang Nan, Yuchen Tian, Senjie Jin, Yibo Wang, Wei He, Boyang Liu, Jixuan Huang, Xin Guo, Zhezheng Hao, Xinbing Liang, Zhihao Zhang, Changzhi Zhou, Wiggin Zhou, Tao Gui, Qi Zhang, Xuanjing Huang, Clarenceai, Aiden Adams
NavGPT-3 is a hierarchical navigation runtime that links a large language model trained with long‑horizon reinforcement learning to a low‑latency action policy, enabling an embodied agent to reason, act, and monitor in parallel threads. The system’s OS‑like scheduler selects which thread controls the robot, allowing rapid interruption and reaction to real‑world events. With the NavGPT VLA action policy, the complete harness achieves state‑of‑the‑art performance on R2R‑CE and matches human performance on RxR‑CE, demonstrating a significant advance in autonomous navigation.
By Gengze Zhou, Yicong Hong, Jiazhao Zhang, Xunyi Zhao, Jian Zhou, Zixing Lei, Zun Wang, Chongyang Zhao, Xionghui Chen, Stephen Gould, Anton van den Hengel, Qi Wu
NOMOS is a four‑pass compiler that transforms natural‑language policies into deterministic, statically verified tool‑call gates for large‑language‑model agents. By performing tool‑schema‑level checks without relying on provers, solvers, or additional LLM calls, it repairs or rejects 37% of airline and 13% of retail policy candidates, enabling most shipped rules to become operable. In benchmark tests, the compiled gates reduce policy violations from 66.3% to 2.6% in airline scenarios and from 30.8% to 6.9% in retail, while achieving a zero attack success rate on banking and low rates on other suites, all with microsecond decision times.
By Min-Young Yu, Tony Kim, Jang Won Choi
The paper introduces a 4‑tensor attention model that predicts the next semantic state of a scene, useful for video generation and robot planning. The model represents a window of states with positions (x, t) and two fibers—a semantic fiber and a temporal‑context fiber—and applies a joint softmax attention over the window. Experiments on ROCStories show that, with comparable parameter counts, the 4‑tensor model achieves lower cross‑entropy on the next‑sentence task and trains significantly faster than a baseline one‑dimensional transformer.
By Jongwook Kim, Sangheon Yun
The paper introduces a formal framework for analyzing how agentic AI systems can improve themselves through longer searches, additional support, or changes in output verification. It defines bounded verification stages, alternating protocols, and terminal checkers, and shows how these mechanisms affect language classes and verification complexity. The authors prove that majority amplification preserves verification classes while existential acceptance can introduce errors, and they explore the implications for recursive self‑improvement and error budgeting.
By Chien-Ping Lu
arXiv:2605. 27381v3 Announce Type: replace-cross Abstract: Agents increasingly modify the procedures by which they solve tasks and improve themselves.
By Chien-Ping Lu
SWE‑Duel proposes a self‑scaling benchmark for coding agents, where tasks and oracles are generated at evaluation time to prevent data leakage and adapt difficulty to the agents’ capabilities. In the arena, two agents take turns modifying a real repository: the Red agent adds a feature, introduces a subtle bug, and the Blue agent must fix it without knowing the bug exists. A tournament of 11 model‑harness participants on 15 repositories in five languages showed GPT‑5.5 as the best solver and Claude Opus 4.8 as the best task author, with newer model versions producing harder tasks for older ones.
AI agents have been shown to write CUDA kernels that outperform PyTorch, but validating these speedups is challenging. The author tested the agents on an NVIDIA DGX Spark and discovered that benchmark design is as crucial as the code itself. The article highlights the need for careful evaluation when comparing AI-generated performance claims.
By Chien Vu Minh
The paper proposes study contracts that bind experimental choices, run obligations, and claim scope to recorded execution evidence, aiming to distinguish contract-relative verification from scientific truth. A diagnostic with eight clean/mutated pairs shows that a deterministic checker can detect all registered mutations, while a broader test on 18 judge aliases flagged 104 of 144 mutated cases, with remaining cases abstaining or failing. The results highlight an information-asymmetric development setting and suggest further tests—full-information comparisons, legitimate-adaptation controls, and closed-loop agent evaluations—to assess whether contract checks improve compliant completion under optimization.
DeltaReplay introduces a step-level memory reuse framework for mobile GUI agents that selectively reuses parts of stored execution trajectories without modifying them. By representing trajectories as paths in a transition graph and separating actions into task-independent operations and task-specific parameters, the system compares each recorded step with the new task and current screen to decide whether to follow, adapt, or ignore it. Experiments on AndroidWorld and SPA-Bench show that this selective reuse can boost task success rates by up to 10.3 and 25.0 percentage points over a baseline agent.
The paper introduces an Evidence-Traceable Dynamic Interviewer Architecture that uses a locally hosted Large Language Model to adapt interview questions in real time based on participants’ expertise. It features a five‑module prompt‑driven design, including an Expertise Profiling module that achieved 78.9% exact agreement with self‑reported expertise and a Generate Iterative Questions module that showed a strong expertise‑complexity association. In a study with 246 participants, the system received high ratings for relevance, engagement, and satisfaction, indicating that its adaptive components functioned as intended.
The paper introduces EMFA, a method for pinpointing the decisive error in large‑language‑model‑based multi‑agent systems. EMFA builds a structured representation of a failed execution, models how errors cascade and loop, and then screens candidates with propagation awareness before verifying them counterfactually. On the Who&When benchmark, EMFA outperforms prior methods, improving step‑level attribution accuracy by 3.45 and 4.40 percentage points on two subsets.
LLM-IDEA is an identifiability‑driven experimental agent that guides large language model scientists in autonomous discovery of mechanistic world models. It evaluates whether a plateau in learning is due to the agent’s capability limits, the model’s resolvability within the design class, or a certified exhaustion of the experiment space. The system demonstrates its effectiveness across benchmarks such as ODEBench, DiscoverPhysics, and a custom Alien Universe testbed, consistently identifying when additional experiments are futile and when new experimental designs are required.
The Neuro‑Memory Fuzzy Inference System (NeMeFIS) is a hierarchical machine learning framework that models car‑following acceleration and deceleration by integrating five human memory types—procedural, working, episodic, semantic, and declarative—through metaheuristics and statistical validation. Trained on 54 models across arterial, collector, and rural highways, NeMeFIS outperforms traditional statistical and conventional machine learning approaches, revealing that declarative memory dominates braking decisions while procedural memory drives acceleration. The system’s robust performance on heterogeneous and homogeneous datasets supports its use in adaptive, human‑like decision systems for Connected and Autonomous Vehicles, potentially improving traffic safety.
arXiv:2610.08862v1 Announce Type: cross
Abstract: Household robots must accommodate new user instructions while executing ongoing tasks. Existing agents often regenerate or extensively revise the rem...
By Ci Zhang, Enfu Nan, Arman Akbari, Lin Zhao, Li Wang, Chen Wang, Weiwei Chen, Yanzhi Wang, Geng Yuan