arXiv subjects

Hao Fu

Publications and source records attributed to Hao Fu.

At least 19 recordsLinked to original sources

arXiv2026

Dark Matter and MOND successes and degeneracies in reproducing SPARC rotation curves and weak lensing data

The observed rotation velocities of spiral galaxies at large scales exceed those predicted by their baryonic components, motivating the need for dark matter (DM) or Modified Newtonian Dynamics (MOND). We fit the rotation curves of 147 late-type galaxies from the SPARC database, using a suite of models including Cold Dark Matter (CDM) models with and without baryonic feedback, Warm DM, Fuzzy DM, and MOND. We also combine SPARC kinematics with weak-lensing data to extend the mean rotation curves to ~1 Mpc. On a galaxy-by-galaxy basis, the inner SPARC rotation curves marginally favour CDM models with baryonic feedback, although accounting for the uncertainties in distances and inclinations increases the degeneracy among DM-based models and MOND. All DM models predict host halo masses larger than abundance-matching expectations, with only the CDM model with baryonic feedback reducing the scatter around the mean luminosity-halo mass relation. MOND is able to reproduce the flatness of the mean galactic rotation curves on scales up to 1 Mpc, providing stellar mass-to-light ratios within observational uncertainties. In contrast, all DM models, without the contribution of mass external to the bound halo, struggle to explain the flat rotation curves. The inclusion of the current weak-lensing data tends to favour MOND over DM models, at least at the largest scales. Future observations from BIG-SPARC, the Square Kilometre Array, and most importantly, upcoming weak-lensing surveys such as EUCLID and Rubin will provide more stringent and independent tests of these competing frameworks

astro-ph.GA↗

arXiv2026

Rubin J122659.4+090236: An Extremely Low Surface Brightness Galaxy Candidate Discovered in the Rubin LSST Early Data Preview 2

We report the serendipitous discovery of an exceptionally low surface brightness galaxy (LSBG) candidate, Rubin J122659.4+090236, in Rubin Observatory imaging of the interacting NGC 4410 system, identified in the Cosmic Treasure Chest public release. 2D Sérsic modelling of the Rubin g, r, and i images reveals a nearly round system with a shallow profile (n ~ 0.4), an effective radius of R_e ~ 6'', and central surface brightnesses of $μ_{0,g}=27.52\pm0.04$, $μ_{0,r}=27.62\pm0.07$, and $μ_{0,i}=27.04\pm0.08$ mag arcsec$^{-2}$. EAZY photo-z fitting favours an intermediate-z solution at z~0.3, while a low-redshift solution at z~0.028, consistent with the NGC 4410 system, is also permitted by a restricted EAZY fit over 0<z<0.1 without imposing a redshift prior. These alternatives imply substantially different physical interpretations, ranging from a diffuse dwarf-like system to an exceptionally extended background LSBG. This discovery demonstrates Rubin's sensitivity to extremely diffuse galaxies and highlights the potential of the LSST survey to uncover large samples of such elusive systems across wide areas, enabling systematic studies of the LSBG population and its role in galaxy evolution.

astro-ph.GA↗

arXiv2026

Mergers may not be the main channel for galaxy stellar mass growth at redshift $z\gtrsim5$: Predictions from cosmological hierarchical models and simulations

Mergers are a crucial channel of galaxy growth, but their role relative to star formation remains uncertain at high redshift. Here we investigate merger rates and merger-driven stellar mass assembly for galaxies with stellar mass $M_\star \gtrsim 10^8 \, {\rm M_{\odot}}$ at $z\simeq 5-11$, by combining complementary theoretical approaches within $Λ$CDM. These include the DECODE semi-empirical model, the TNG100 and EAGLE hydrodynamical simulations, and the \textsc{L-Galaxies} semi-analytical model, together with the Millennium and TNG100-Dark simulations used to examine dark matter subhalo statistics and merger timescales. This compares models with different treatments of baryonic physics and galaxy-halo connections. We find that dark matter-only simulations predict merger timescales $\sim 30\%$ longer than those inferred from cosmological simulations including baryons. Galaxy major merger ($M_{\rm \star,sat} / M_{\rm \star,cen} > 0.25$) rates are high at cosmic dawn and increase towards earlier epochs, reaching $\sim 1/{\rm Gyr}$ at $z\sim 5$ and peaking at $\sim 3/{\rm Gyr}$ at $z\sim 8$. Despite these high rates, mergers contribute only $\lesssim 10\%$ of the total cosmic stellar mass density evolution. The ex-situ stellar mass fraction ranges from a few percent to $\sim 20\%$ for all galaxies at $z\gtrsim 5$, across different models and simulations. Their broad agreement implies that stellar mass assembly at cosmic dawn is dominated by in-situ star formation rather than mergers. These findings suggest that the high pair fractions identified by JWST may not directly trace efficient ex-situ stellar mass growth, highlighting the need for physically motivated merger timescales to interpret galaxy assembly in the early Universe.

astro-ph.GA↗

arXiv2026

NetAgent: Multi-Task Agentic Network Traffic Analysis Made Practical

Network traffic analysis is central to network security, spanning tasks from intrusion detection to encrypted traffic classification. Existing approaches either train task-specific models that generalize poorly or rely on costly traffic foundation models that still struggle under distribution shift. We present NetAgent, the first agentic framework for multi-task traffic analysis. Through a carefully designed agent loop, NetAgent supports complex task understanding, on-the-fly decomposition and orchestration, dynamic replanning, and long-horizon analysis, without task-specific training. It introduces five key designs: (1) knowledge-augmented workflow planning that maps attack knowledge to traffic features to bridge the semantic gap; (2) a comprehensive tool action space with 150+ verified tools extracted from 50+ published systems; (3) a unified code execution space for flexible action composition; (4) a three-tier memory for long-term knowledge consolidation; and (5) sandboxing and runtime repair for reliable execution. Across 9 major benchmarks, NetAgent outperforms all baselines (23 single-task and 5 multi-task) on nearly all tasks and generalizes substantially better to unseen traffic distribution (90.04% F1 vs. 2.74% and 3.04% for the best single-task and multi-task baselines) and under realistic background shift (4.85-point F1 drop vs. 74.88-point and 74.80-point drop for the best single-task and multi-task baselines). These results reveal that existing methods owe much of their reported success to overfitting dataset-specific patterns and degrade sharply in realistic network environments, while NetAgent's agentic design remains accurate, generalizable, and robust.

cs.CR↗

arXiv2026

Verify, Don't Trust: Agentic Model Development for Video Discovery Retrieval at Scale

Large language model (LLM) agents can propose, implement, and evaluate model changes. Autoresearch loops demonstrate this capability through minutes-scale iterations on a self-contained program. Online autoresearch instead spans asynchronous systems, hours-long variants, and weeks-long campaigns that can influence a product. A completed run can still support an invalid conclusion when a code change is a no-op, data windows leak, evaluator semantics drift, or the two arms traverse different serving funnels. We present EvoPilot, a human-gated method for long-horizon online autoresearch. Role-specific agents execute each round through a versioned domain skill and typed adapter. Durable records preserve experiments and failures; deterministic checks enforce recorded lessons. We study a 37-day campaign for the retrieval system that powers Video Deep Dive (VDD), an online experience for discovering follow-on videos after a user opens a seed video. The campaign covered seven directions and used an hourly refreshed index of hundreds of millions of videos. Earlier manual experiments had not established a benefit from an interaction head. A primitive autoresearch attempt revisited the direction but incorrectly attributed an offline hit-rate decline of 22 percentage points to the head. We then introduced EvoPilot. Its human-gated verification traced the drop to a pre-existing evaluation defect that produced output depths of 3,000 and 600. After repair, a matched comparison measured an offline improvement of 3.20 percentage points. Post-study replay and mutation tests rejected invalid comparisons while admitting valid counterparts. Durable state recovered an interrupted round, and artifact reuse avoided approximately five GPU-hours. Separately, a seven-day randomized online evaluation estimated a 0.66% relative increase in the VDD slice of Good Search Result Rate for Retention (GSRR).

cs.IR↗

arXiv2026

Hybrid GPU-CPU Retrieval for Personalized Search at Ultra-Large Scale

Embedding-based retrieval on user-generated content at the trillion-document scale exposes a sharp conflict between two production demands: deep, expressive personalization for queries with rich user intent, and broad coverage of a massive inventory under fixed latency and resource budgets. We characterize this as the personalization-scale paradox: hosting the full serving inventory in GPU memory is too resource intensive, while CPU compute cannot execute the same interaction-heavy model on the latency-critical path. We present a hybrid GPU-CPU co-serving system that resolves the paradox through orchestration rather than a new model class. A high-depth GPU pathway fuses retrieval and interaction pre-ranking over a curated online pool on the order of a billion documents, while a high-breadth CPU pathway searches an independently selected online inventory roughly twenty times larger with lightweight personalized scoring. Either or both pathways can run per request; candidates are deduplicated before shared downstream ranking. The system is deployed in production. A full-system A/B test against the legacy CPU-only configuration improves model-scored relevance and substantive engagement, while separate pathway experiments show positive value at their own deployment scopes. Retrieval logs show that the pathways contribute structurally distinct candidates, production serving measurements characterize their latency, and a matched capacity plan quantifies the economic rationale for assigning modeling depth to GPUs and inventory breadth to CPUs. Together, these results validate a practical, independently evolvable depth-breadth architecture for ultra-large-scale personalized search.

cs.IR↗

arXiv2026

Massive Galaxy Halos Contain Less Inner Dark Matter Than Predicted

The mass profiles of galaxy halos encode how baryons reshape dark matter distribution, yet direct observational constraints across the full radial range remain scarce. Here we combine stellar kinematics from MaNGA, H I dynamical measurements from ALFALFA, and independently calibrated halo masses of SDSS groups to statistically reconstruct the mass distribution of central galaxies over nearly two orders of magnitude in radius. We demonstrate that H I data alone do not provide reliable total halo mass estimates, necessitating an independent group-based halo-mass scale. Compared to the IllustrisTNG and EAGLE simulations, the observational profiles of low-mass halos are broadly consistent; in contrast, massive observed halos exhibit systematically lower dynamical masses at the H I radius, lower inner dark-matter masses, and lower central dark-matter fractions (about 4$σ$ difference in units of population scatter) at fixed total halo mass. After subtracting baryonic contributions, the inferred dark-matter profiles remain broadly consistent with an NFW form, but with lower effective concentrations than predicted for massive halos. These results suggest that the inner dark-matter content of massive halos has been reduced more significantly than predicted by current hydrodynamical simulations, plausibly due to long-term baryonic halo heating in massive systems.

astro-ph.GA↗

arXiv2026

Forged in Quenching: Morphological Transformation across Star-forming and Quiescent Galaxies in EAGLE

The connection between morphology and quenching in central galaxies is well established, but its physical origin remains widely debated. We address this by tracing the main progenitor branches of $z=0$ star-forming and quiescent central galaxies in the EAGLE cosmological simulation from $z\gtrsim4$. Their disc-to-total ratio and triaxiality tracks are indistinguishable until $z\approx 1$-$2$, when both diverge concurrently with the onset of quenching, whereas the size and supermassive black hole (SMBH) mass differences are established earlier. We identify four physically distinct channels linking galaxy morphology and quenching. First, mergers cause size growth, rotation suppression, triaxiality increase, and SMBH growth, with the accumulated SMBH mass subsequently causes the quenching of galaxies. Second, with merger history controlled, galaxy morphology modulates SMBH growth throughout the star-forming phase: compact, dispersion-dominated galaxies grow their SMBHs faster and are preferentially quenched, producing the size and morphology differences between star-forming and quiescent galaxies. Third, at fixed stellar mass and SMBH mass, compactness further facilitates the quenching of galaxies. Fourth, disc instability transforms compact oblate discs into prolate systems, with substantial size growth and suppressed rotation but negligible stellar mass growth. This secular channel contributes about half of the prolate galaxy population around $M_{\rm star}\approx 10^{10.6}\,\rm M_\odot$. Prior to quenching, the progenitors of quiescent galaxies already have smaller sizes, lower disc-to-total ratios, and more massive SMBHs than star-forming galaxies at the same epoch, by amounts comparable to their differences at $z=0$. Morphology therefore plays an active role in growing the SMBH and quenching the galaxy, rather than being passively inherited through progenitor bias.

astro-ph.GA↗

arXiv2026

Revolutionizing Turn-by-Turn Navigation with Cloud-Edge Deep Learning

Turn-by-turn (TBT) navigation systems are integral to modern driving experiences, providing real-time audio instructions to guide drivers safely to destinations. However, existing audio instruction policy often relies on rule-based approaches that struggle to balance informational content with cognitive load, potentially leading to driver confusion or missed turns in complex environments. To overcome these difficulties, we first model the generation of navigation instructions as a multi-task learning problem by decomposing the audio content into combinations of modular elements. Then, we propose a novel deep learning framework that leverages the powerful spatiotemporal information processing capabilities of Transformers and the strong multi-task learning abilities of Mixture of Experts (MoE) to generate real-time, context-aware audio instructions for TBT driving navigation. A cloud-edge collaborative architecture is implemented to handle the computational demands of the model, ensuring scalability and real-time performance for practical applications. Experimental results in the real world demonstrate that the proposed method significantly reduces the yaw rate (the proportion of vehicles deviating from navigation routes) compared to traditional methods, delivering clearer and more effective audio instructions. This is the first large-scale application of deep learning in driving audio navigation, marking a substantial advancement in intelligent transportation and driving assistance technologies.

cs.AI↗

arXiv2026

DIANOIA: Diagnostic Decomposition and Joint Optimization for Multi-Agent Reasoning

Multi-agent LLM systems consistently outperform single-agent baselines, yet practitioners still cannot predict which design works for a new task or diagnose why one fails. We argue this gap persists largely because the field lacks a diagnostic framework with measurable primitives and testable predictions. We introduce \textbf{DIANOIA}, a three-channel decomposition of multi-agent reasoning gain into coverage, fidelity, and synthesis, each of which is empirically measurable. From this decomposition, we derive a diagnostic protocol that identifies the bottleneck channels for any given task. We instantiate the protocol as a multi-agent system whose three components mirror the channels: role-diverse proposers for coverage, execution-grounded verification for fidelity, and iterative synthesis. On GSM8K, AIME-2025, MBPP, and BFCL-SP, our method outperforms strong multi-agent baselines under matched token budgets, dominating the Pareto frontier on MBPP at $\sim$$5{\times}$ token savings and reaching $+4.6$pp at matched cost. On every benchmark, the protocol picks the right bottleneck channels; the system we built around it leads across models. We release code, adapters, diagnostic metrics, and a Claude Code skill at https://anonymous.4open.science/r/DIANOIA4MAS. DIANOIA reframes multi-agent design as channel-aware resource allocation: diagnose which channel is the bottleneck for your task, then invest tokens accordingly.

cs.AI↗

arXiv2026

Generalized Mermin Inequalities for Benchmarking Large-Scale GHZ States

Multipartite Bell tests provide a correlation-only route to benchmarking quantum processors, but their application at large scales is hindered by the rapid decay of many-body correlators under noise and exponentially many terms in conventional Bell expressions. Here we address these scalability obstacles by introducing a finite-setting generalized Mermin family of state-tailored Bell inequalities with analytic certification bounds, in which the measurement-setting number $m$ provides an additional certification dimension complementary to the system size $n$. We show that, for the powers-of-two setting choices considered here, increasing $m$ leaves the ideal normalized multipartite quantum value unchanged while lowering the relevant classical bounds, thereby strengthening the Bell-violation ratios and yielding an improved noise-robustness scaling compared to the standard Mermin inequality. We test this construction experimentally on a programmable superconducting processor by preparing Greenberger-Horne-Zeilinger (GHZ) states of up to 80 qubits. Using randomized sampling for direct Bell-operator estimation, we observe Bell ratios that grow exponentially with system size, certify a nonlocality depth of 14, and show that increasing $m$ strengthens both the Bell ratio and depth certification. All results are obtained solely from measured correlators and analytical bounds, without readout correction, tomography, or model-based mitigation. Generalized Mermin inequalities therefore provide a sharper Bell benchmark for noisy large-scale GHZ states.

quant-ph↗

arXiv2026

How galaxies acquire their stellar mass at high redshift: High star formation efficiencies and the relative roles of dust and initial mass function

JWST has measured an unprecedented abundance of galaxies above $z\gtrsim 4-5$, whose formation and evolution are still difficult to reconcile within traditional galaxy evolution models in a $Λ$CDM framework. Here, we present a study on the star formation histories of these high-redshift galaxies between $z\simeq5-12$ via a data-driven semi-empirical model that uses the observed UV LFs as input to retrieve SFRs, naturally bypassing any uncertain modelling of cooling, feedback and/or stochastic processes. Galaxy stellar masses are progressively built in time by integrating their SFRs assigned along their progenitor haloes via the SFR-halo accretion rate relation, derived from abundance matching between the input observed UV LFs with the dark matter halo accretion rate distributions at each redshift. This makes the SFEs a full prediction of the model rather than a tuned input, serving as a natural baseline to test burstiness, dust attenuation, or IMF variations. Our approach naturally reproduces the total stellar mass function, the large-scale clustering, and the star-forming main sequence. We find that massive galaxies grew their stellar mass with a bursty star formation at $z\sim9-10$, broadly in agreement with the star formation histories inferred from spectral energy distribution fitting, with the SFE reaching high peaks of $0.8-0.9$ at $z>9$ and lowering to standard values of $0.2-0.3$ below $z\lesssim9$. We find that the presence of dust could enhance the predicted SFRs at $z\lesssim8$, better reproducing the observed SFRs of massive dusty galaxies, and increase the SFEs to values close to or even above unity at $z \gtrsim 8$. Finally, switching to top-heavy IMFs reduces the SFEs by a factor of $2-3$, highlighting the need for a variable IMF as an inevitable ingredient in the evolution of galaxies at high redshifts to avoid unphysical SFEs, especially in the presence of dust.

astro-ph.GA↗

arXiv2026

SPLC: Social Preference Learning for Crowd Robot Navigation

Offline reinforcement learning (RL) holds significant potential for crowd robot navigation in human-robot coexistence applications. However, the inherent complexity of pedestrian motion renders the design of effective reward functions for promoting socially compliant robot behaviors a persistent challenge. This paper proposes a Social Preference Learning for Crowd Robot Navigation (SPLC) algorithm to eliminate the need for detailed reward design. Its core innovation lies in the introduction of a social preference feedback mechanism to automatically generate preference data through principled preference evaluation criteria. By explicitly accounting for the intricacies of pedestrian dynamics, the pipeline mitigates the reward bias and facilitates the systematic quantification of broad social norms, thereby fostering socially compliant behaviors. Extensive experiments integrating SPLC with offline RL methods demonstrate consistent improvements over state-of-the-art baselines across standard performance metrics. Furthermore, real-world experiments on the TurtleBot4 further validate the effectiveness of SPLC in practical human-robot coexistence settings. Our code and video demos are available at https://github.com/sklus949/SPLC.

cs.RO↗

arXiv2026

HyBIRD: Hyperbolic Bridge Retrieval and Diagnosis for Methodology Inspiration Retrieval

Methodology Inspiration Retrieval (MIR) asks a system to retrieve prior papers whose methods can inspire a new research proposal. Unlike general scientific retrieval, the central challenge is not topical similarity but whether a candidate paper provides concrete mechanisms that can instantiate an abstract methodological need. Existing MIR dense retrievers provide strong paper-level rankings, but the returned lists do not expose how proposal needs are bridged by retrieved methods, where evidence is weak, or which complementary snippets may help. We propose HyBIRD, a frozen-anchor framework that treats MIR as hyperbolic bridge retrieval and post-hoc method diagnosis. HyBIRD keeps a strong MIR dense retriever fixed, learns lightweight point, cone, and factorized hyperbolic bridge variants, and uses LLM-assisted method blocks for post-hoc explanation and evidence selection. On the MIR benchmark, the factorized bridge reaches 59.034 mAP while preserving the dense anchor's strong retrieval behavior. More importantly, HyBIRD converts ranked papers into inspectable query need profiles, factor coverage, maturity views, and complementary evidence bundles. The results suggest that hyperbolic geometry is most useful as calibrated structure over a dense anchor, rather than as a standalone replacement for dense retrieval.

cs.IR↗

arXiv2026

Macaron-A2UI: A Model for Generative UI in Personal Agents

As personal agents evolve to handle complex, user-centric tasks, static plain-text chat is rapidly becoming a bottleneck. Generative UI emerges as the necessary new interface layer, dynamically synthesizing the right controls, options, and state from the interaction context in real time. We present Macaron-A2UI, a model for Generative UI in personal agents. Our goal is to move beyond text-only interaction by enabling agents to generate natural language together with lightweight, executable UI actions for information collection, preference refinement, confirmation, and multi-goal organization. We build a large-scale Generative UI corpus from heterogeneous dialogue sources, introduce A2UI-Bench for controlled evaluation, and train 30B, 235B and 754B models with parameter-efficient LoRA-based supervised fine-tuning followed by reward-driven reinforcement learning. The best Macaron-A2UI model reaches 75.6 overall on A2UI-Bench without explicit schema hints, surpassing the strongest full-schema frontier baseline. We release the models, benchmark, and evaluation protocol to support future work on Generative UI for personal agents.

cs.HC↗

arXiv2026

When Retrieval Hurts Code Completion: A Diagnostic Study of Stale Repository Context

Context: Retrieval-augmented code generation relies on cross-file repository context, but retrieved snippets may come from obsolete project states. Objectives: We study whether temporally stale repository snippets act as harmless noise or actively induce current-state-incompatible code. Methods: We conduct a controlled diagnostic study on a curated 17-sample set of production-helper signature changes from five Python repositories. For each sample, we compare current-only, stale-only, no-retrieval, and mixed current/stale retrieval conditions under prompts that hide commit freshness and expected current signatures. Results: Under neutralized prompts, stale-only retrieval induces stale helper references on 15/17 Qwen2.5-Coder-7B-Instruct samples and 13/17 gpt-4.1-mini samples, corresponding to 88.2 and 76.5 percentage-point increases over current-only retrieval. No retrieval produces zero stale references but only 1/17 passing completions. The two models share 75.0% Jaccard overlap among stale-triggering samples, and mixed conditions show that adding valid current evidence largely rescues stale-only failures. Conclusion: Temporal validity of retrieved repository context is a distinct diagnostic variable for Code RAG robustness: stale context can actively bias models toward obsolete repository state rather than merely removing useful evidence.

cs.SE↗

arXiv2026

Learning Agile Striker Skills for Humanoid Soccer Robots from Noisy Sensory Input

Learning fast and robust ball-kicking skills is a critical capability for humanoid soccer robots, yet it remains a challenging problem due to the need for rapid leg swings, postural stability on a single support foot, and robustness under noisy sensory input and external perturbations (e.g., opponents). This paper presents a reinforcement learning (RL)-based system that enables humanoid robots to execute robust continual ball-kicking with adaptability to different ball-goal configurations. The system extends a typical teacher-student training framework -- in which a "teacher" policy is trained with ground truth state information and the "student" learns to mimic it with noisy, imperfect sensing -- by including four training stages: (1) long-distance ball chasing (teacher); (2) directional kicking (teacher); (3) teacher policy distillation (student); and (4) student adaptation and refinement (student). Key design elements -- including tailored reward functions, realistic noise modeling, and online constrained RL for adaptation and refinement -- are critical for closing the sim-to-real gap and sustaining performance under perceptual uncertainty. Extensive evaluations in both simulation and on a real robot demonstrate strong kicking accuracy and goal-scoring success across diverse ball-goal configurations. Ablation studies further highlight the necessity of the constrained RL, noise modeling, and the adaptation stage. This work presents a system for learning robust continual humanoid ball-kicking under imperfect perception, establishing a benchmark task for visuomotor skill learning in humanoid whole-body control.

cs.RO↗

arXiv2026

Benchmarking Real-Time Question Answering via Executable Code Workflows

Retrieving real-time information is a fundamental capability for search-integrated agents in real-world applications. However, existing benchmarks are predominantly static and therefore fail to capture the temporal dynamics of information and the continuously evolving nature of real-world knowledge. To address this limitation, we propose RT-QA, a dynamic evaluation framework that leverages executable code workflows to retrieve up-to-date answers at evaluation time. Specifically, we construct an agent-driven pipeline that autonomously generates code for web crawling and DOM-based answer extraction to produce real-time ground truth. To ensure robust evaluation over time, the pipeline further incorporates a self-repair mechanism to adapt to changes in web page structures. RT-QA spans 12 domains (e.g., Finance, Sports) with 320 Chinese questions categorized into three difficulty levels. Extensive evaluations of state-of-the-art models (e.g., GPT-5.2, GLM-4.7) reveal significant limitations in real-time adaptability: even the best models achieve only 46% accuracy. Our analysis highlights two primary failure modes: (1) Lazy Retrieval, where agents rely on search snippets instead of deeply scanning specific websites for information (20% of failures); and (2) Temporal Confusion, a cognitive error where agents retrieve a historical date (e.g., an event in 2024) and fail to re-anchor to the current time (2026) for subsequent reasoning. These findings suggest that future agents require not just better retrieval strategies, but robust temporal state management.

cs.IR↗