Single-Agent Baselines

Air-ground collaborative Vision-and-Language Navigation (VLN) pairs an unmanned aerial vehicle (UAV) with a global bird's-eye view and an unmanned ground vehicle (UGV) with a local first-person view, yet the setting remains largely unexplored: existing training-free methods solve single-agent tasks but offer no collaboration mechanism, and a recent CARLA-Air evaluation found no stable cooperative behavior across five state-of-the-art VLA models; naive semantic communication or bidirectional coupling even degrades performance. We establish AGC-VLN (Air-Ground Collaborative VLN), the first training-free baseline for air-ground collaborative VLN. The key insight is that training-free methods decompose navigation into VLM-based semantic reasoning and deterministic geometric execution, exposing a collaboration interface: the UAV's global view, over which it renders the UGV's reported pose and the VLM-anchored target as CAR/GOAL markers with distance labels, yielding a shared bird's-eye map. From this map, the UGV acquires global spatial context its first-person view cannot provide, plans a road-following path with a frozen VLM, and executes it under closed-loop control; in parallel, the UAV runs 3D-SPF, a spatial-search upgrade of SPF that localizes the target in the downward view and flies toward it. On 100 closed-loop episodes in CARLA-Air's Town10HD scene, AGC-VLN reaches a 77.0% joint success rate, a collaboration gain of +27.0% over the weaker individual agent (the UAV, 50.0%), and exceeds the strongest published single-agent baseline (Travel UAV, 53.0%) by 24.0 points, stemming from the complementarity of the UAV's global view and the UGV's road-following execution. Project page: https://github.com/ZSN2024/AGC-VLN.

Modern image generation and editing systems can produce photorealistic, prompt-aligned images, but still often render familiar objects at implausible relative sizes. To measure this failure mode, we introduce GenScale, a benchmark and evaluation protocol for real-world relative object scale in image generation and editing. GenScale contains 900 image-level entries and 1,643 pairwise anchor-target scale relations across common-object generation, human-product generation with metric dimensions, and scale correction from failed generations. We further design a human-calibrated ordinal judge for scalable pairwise scale evaluation. Last but not the least, we introduce Rescale, a model-agnostic post-processing agent for localized scale correction without modifying the source generator. Experiments reveal that state-of-the-art image generators and editors cannot reliably observe relative scale yet, while Rescale consistently improves scale plausibility across generated and edited images. Together, GenScale establishes relative object scale as a distinct, measurable, and actionable capability for image generation systems.

Agent performance depends jointly on the model parameters and the executable harness code that manages context and control flow. Optimizing either component in isolation can leave the system bottlenecked by its frozen counterpart: weight updates can change which harness is effective, while harness updates can change which model capabilities are exposed. Existing joint-adaptation methods optimize weights and textual prompts but leave the broader harness fixed. We propose Weight-Harness Alternating LEarning (WHALE), a simple recipe that alternates two phases: updating the model under the current harness, then searching for a better harness under the updated model. We instantiate these two phases with online rejection-sampling fine-tuning and Meta-Harness, respectively. When to switch is a key design choice: to separate real improvements from noise without over-optimizing against a changing counterpart, WHALE uses either fixed phase durations or an adaptive patience rule over training signals. Using Qwen3.5-2B/4B agents across three domains (search question answering, mathematical reasoning, and chess puzzles), WHALE outperforms weight-only, harness-only, and Fast-Slow Training by 4.15-24.38 percentage points in best mean@8 accuracy. Either component can be the bottleneck: harness search matches peak weight-only accuracy with far fewer rollouts in SearchQA, but improves math accuracy only after a weight update. Small interleaved updates also outperform stagewise weight-then-harness optimization in accuracy and rollout cost. The code is available at https://github.com/krafton-ai/WHALE.

Maintaining price consistency and executing an Every Day Low Price strategy is critical for global retailers. However, with catalogs spanning millions of active items, manual governance of price relationships is infeasible. Inconsistent pricing across item variants distorts customer value perception and cannibalizes sales. To address this, we present a scalable, context-aware Multi-Agent Framework designed to automate the construction of "Lines and Ladders" pricing taxonomies. Our framework employs specialized LLM agents to construct these coherent pricing structures by identifying key attributes, extracting multi-modal values, and applying hierarchical grouping logic. Evaluated on real-world enterprise data and deployed in production, our 3-Agent system achieves an F1-score of 0.83 for Lines, outperforming single-agent baselines by mitigating cognitive overload. The system achieves >90% precision and >75% recall in Food & Consumables, and 80.2% assignment accuracy in the unstructured General Merchandise catalog.

Medical diagnostic reasoning is a high-impact use case for LLMs that carries significant implications for the health and wellbeing of users. When OpenAI (2026) reports that more than 5% of ChatGPT messages globally are healthcare-related, the transparency of these systems becomes a serious design concern. This is especially true for complex cases, where differential diagnosis often requires integrating multiple forms of specialist reasoning. Existing work has proposed multi-agent approaches to medical diagnosis, but it remains unclear when such systems are needed, why they help, and where they outperform monolithic inference. We introduce Social Chain of Thought (SCoT),a multi-round pipeline for medical differential diagnosis that structures multi-agent interaction as a deliberative framework for collabora. tive LLM reasoning. Evaluating SCoT against single-agent baselines, one-agent pipeline ablations, and best-of-n scaling, we show that its recall advantage is not reproduced by monolithic inference alone. SCoT is most successful in the hardest diagnostic cases, where multiple rounds of specialist conversation help recover ground-truth diagnoses and converge on a higher-recall differential.

We introduce the Dark Souls Learning Environment (DSLE), a containerized platform that presents all 22 boss encounters of Dark Souls: Remastered as game-playing agent benchmarks through a Gymnasium-style interface. DSLE combines real-time combat, high-dimensional visual input, and sparse terminal rewards, with each environment step being a real action executed against the running game. To support controlled comparison, we define DSLE-5, a representative five-boss subset, spanning a melee fight, a spatially constrained arena, an environmental-hazard fight, a multi-target fight, and a fast final-boss fight, that we recommend as the starting suite for agents built on DSLE. On DSLE-5 we evaluate a random policy, an expert system, an evolutionary baseline, and PPO and DQN agents trained from visual input. The expert system and the evolutionary baseline each defeat the Asylum Demon, the game's tutorial boss (63% and 43% peak win rates), but none of the five methods defeats the other four DSLE-5 bosses; PPO and DQN show no measurable learning (at most 0.33% win rate on the tutorial boss, 0% elsewhere) within a budget that already costs tens of wall-clock hours per run. A broader study running the evolutionary baseline across all 22 encounters under advantaged all level-50 stats yields wins on only a handful of additional early-game bosses and leaves the rest unwon. The failure cases range from sub-10-second deaths in cramped, multi-target encounters to minute-long stalemates that inflict almost no damage, and we report them through survival time and damage dealt rather than win rate alone.

Aligning large language models (LLMs) with human values remains a major challenge, especially for trustworthy AI. While existing approaches such as RLHF, CAI, and their variants have achieved promising results, they often rely on a single-agent framework and a unified reward system. This limits their ability to capture ethical pluralism, adapt to diverse moral contexts, and reflect the dynamics of multi-agent moral reasoning. In this work, we propose a framework that utilizes multilayer combinatorial fusion for contextual value alignment (MCF-CVA). At the first layer of the framework, it instantiates multiple moral agents, each fine-tuned to represent a distinctive value. Their outputs are then expanded combinatorially using both score- and rank-combinations as well as average and weighted aggregations. These combined models are then reduced to the same number of initial moral agents. This expansion and reduction (EAR) process continues for multi-layers until a stopping criterion is reached. The MCF-CVA framework leverages cognitive diversity between agents to mitigate conflicts and redundancies across multiple agents, producing responses that better reflect contextual human values. The framework using the EAR algorithm is performed on the dual architecture of Euclidean score space and Kemeny rank space. Empirical evaluations demonstrated that the proposed framework outperforms single-agent baselines, multi-agent single-layer results, and previous aggregation approaches on standard metrics, showing that the MCF-CVA framework provides a robust and effective mechanism for advancing contextual value alignment in LLMs.

AI agents are moving from advisory interfaces into systems that execute consequential operations: issuing refunds, reserving scarce inventory, provisioning cloud resources, and initiating financial transfers. These workflows require governance over effects, not only over model outputs. Existing safeguards often decide whether an action is allowed from the information available when the action is requested. For stateful policies, that request-time view may be incomplete: budgets, inventory, approval status, and risk signals can change before the effect occurs, making an earlier authorization or approval stale. This paper studies stateful governance for concurrent agentic systems. We identify stale authorization as the core failure mode and define policy-state serializability, a correctness condition requiring committed effects to be explainable as authorized against the policy state immediately before they occur. We present MasuGate, a runtime architecture that keeps policies as reviewable programs while coordinating the state and effects needed to preserve their decisions. In experiments with a PostgreSQL-backed prototype of MasuGate, the system prevents stale authorizations missed by baselines that pass policy state as ordinary request context, preserves delayed approvals while unrelated work proceeds, keeps policy evolution mostly in policy text rather than trusted provider code, and avoids policy violations in a scripted, LLM-free procurement workflow where agent-governance baselines produce stale authorizations over shared budgets and inventory. More broadly, MasuGate suggests a path for integrating stateful governance boundaries into agent frameworks and provider-backed domains where agents act on shared resources.

On standard factuality tasks, frontier models now cluster near the top of the scale. The question is therefore shifting from how factual a system is toward how much compute that factuality costs. Static leaderboards score factuality in isolation and treat compute as free, so they cannot tell a genuinely better system apart from one that simply spends more. Consider a ranking reversal. A brute-force Best-of-4 agent posts the higher raw factuality score (H-Score 0.9169 vs 0.9103) and would top a static leaderboard, but once cost is counted it is the worse system, losing on Q-Score (0.5169 vs 0.5217) at roughly four times the tokens and latency, under a reported cost weight whose sensitivity we sweep. So the system that tops a static leaderboard can be the worse one to deploy. To make this trade-off visible, we introduce MAS-HQ (Multi-Agent System Hallucination Quest), a resource-aware evaluation protocol. It wraps any factuality detector and normalizes for cost, and it pits systems against each other rather than scoring them in isolation. The Q-Score measures factuality minus normalized cost under a competitive match. Across summarization and open-domain QA, single-agent baselines drift into resource-heavy over-optimization, while competition elicits more resource-efficient policies. These gains are small but consistent, and stable across 100 trials. The axis stays discriminative for frontier systems (Gemini-2.5-Pro, and GPT-5) whose raw factuality scores are already bunched near the ceiling. MAS-HQ provides a reproducible way to measure how much a factual answer costs.

Large language model agents have advanced rapidly, yet progress remains fragmented across domains, capabilities, task difficulty, and interaction settings. We frame this as full-scenario agentic scaling and present AgentOmnia, a framework coordinating task-space definition, data synthesis, post-training, evaluation, and improvement across To-Consumer (ToC), To-Business (ToB), and To-Employee (ToE) applications. An extensible Domain x Capability x Atomic Difficulty taxonomy aligns these stages and enables fine-grained diagnosis with OmniaBench. AgentOmnia combines bidirectional environment-task synthesis with tool-dependency, program-structured, and solver-based pipelines, constructing 5,018 stateful environments with 255,375 tools and 52,361 tasks. Programs, solvers, and verifiers provide correctness signals, while supervised fine-tuning, online agentic reinforcement learning, and a rollback curriculum support post-training. Evaluation failures translate into Product Requirement Documents (PRDs) for targeted self-evolution. Starting from Qwen3-30B-A3B-Thinking-2507, AgentOmnia raises the pass rate on the OmniaBench challenging subset from 9.16% to 37.11% and the macro-average across OmniaBench, $τ^2$-Bench, DeepPlanning, and VitaBench from 22.86% to 41.69%. Under a unified protocol,it leads the evaluated agentic post-trained baselines on OmniaBench and retains the highest four-benchmark macro-average. It also surpasses Qwen3-235B-A22B-Thinking-2507 on all four benchmarks and exceeds Qwen3.5-35B-A3B on the macro-average. Gains span three application splits, ten capability dimensions, eight atomic-difficulty factors, and 76 of 90 level-1 domains, indicating broad rather than category-specific improvement. A one-round study provides initial evidence for PRD-guided self-evolution, motivating validation at larger scales and in industrial settings.

This technical report presents our approach to Challenge Track~3: SeePhys Pro at the 3rd AI for Math Workshop, where the task is to answer college-level physics questions whose statement and figure may be given partly or entirely as an image. Visual physics problems become substantially harder for large language models when the decisive information resides in a figure rather than in the text, and this modality gap widens as more of the problem migrates into the image. We address the task with a two-stage framework: a visual information extraction stage that re-expresses figure content as solver-readable text to close the modality gap, and a reasoning stage that orchestrates three heterogeneous solvers through multi-agent debate. Our analysis yields two findings: the gain from orchestration comes from reliable answer selection rather than from additional debate, and the value of a figure aid scales with how much of the problem is locked inside the image. The resulting pipeline improves overall accuracy over a single-agent baseline from 0.643 to 0.802 on the public split, and won 1st place on both the public and the private leaderboard (private overall 0.743).

Self-improving AI systems typically treat the agent as the object that improves, by optimizing prompts, workflows, harnesses, or even the agent's own code. This agent-centric view can make improvements expensive to maintain and difficult to transfer, because gains become tied to a particular agent design, task distribution, or adaptation run. We study a complementary paradigm: knowledge-centric self-improvement, in which agents remain generic and disposable while the persistent object is a curated knowledge base that agents can leverage for future tasks. We conduct controlled case studies to operationalize this idea via a simple protocol. Agents attempt one task, then contribute evidence-grounded insights to a shared knowledge base via task-level and cross-task forums, followed by knowledge distillation. Because self-improvement is contained in the knowledge rather than the agent, improvement can be more inspectable, transferable, and portable. Across abstract reasoning, coding, and terminal benchmarks, this protocol improves solve rates while reducing dollar cost relative to agent-centric baselines. The resulting distilled knowledge also transfers to held-out tasks and across LLM families, indicating that the improvement is not merely an LLM- or run-specific behavior. These results support a new view of self-improving agentic systems: progress can be driven primarily by the curated persistent knowledge. Code is available at https://github.com/recursive-knowledge/KSI.

Recent autonomous penetration testing papers report high benchmark scores while adding multi-component security harnesses around frontier LLMs. Because these systems often change both architecture and backbone model, it is difficult to tell how much performance comes from the harness rather than from the underlying model. This paper presents a controlled study on the 104-task XBOW benchmark using default coding CLI agents as plain-agent baselines. We first run Codex, OpenCode, and Pi with the same GPT-5 model, budget, target interface, and scoring rule. This phase identifies the strongest same-model baseline and tests whether security-specific prompt variants improve its observed score. We then compare the default Codex scaffold with published MAPTA and PentestGPT V2 results under the closest available model matches. Finally, we repeat the plain-agent experiment with GPT-5.2 and GPT-5.5 to measure model scaling inside the same scaffold. The results show a mixed but practical picture. Specialised harnesses can add measurable benchmark lift and may improve cost efficiency, but plain coding agents already solve a large share of the benchmark; repeated plain-agent runs can match or exceed some published architecture scores in union coverage, and newer models substantially improve the same scaffold. Future evaluations should report model-matched plain-agent baselines before attributing benchmark gains to architecture design alone.

While autonomous coding agents have significantly advanced automated test generation, they remain fundamentally limited by lazy generation, a phenomenon where agents prematurely terminate tasks and systematically avoid complex programmatic logic, resulting in inadequate code coverage. Currently, mitigating this premature termination requires continuous human-in-the-loop supervision. This heavy reliance on human intuition creates a bottleneck that negates the efficiency gains of automated generation. We propose SCATE, a framework for adaptive, automated supervision of coding agents that replaces human intervention during test generation. By formulating supervision as a contextual bandit problem, SCATE learns to select the most promising testing actions based on the current coverage and class testability metrics, maximizing coverage gains while minimizing wasted generation effort. Our empirical evaluation demonstrates that SCATE integrates seamlessly with different coding agents. When applied to GEMINI-CLI, it achieves 32.3% higher line coverage and 30.9% higher branch coverage than the agent-only baseline. A comparison with CLAUDE CODE confirms the framework dynamically adapts its policy to optimize each agent's unique strengths. SCATE also consistently outperforms state-of-the-art non-agentic approaches across all metrics.

Large language models (LLMs) demonstrate remarkable reasoning capabilities, yet their stateless architecture fundamentally limits deployment in long-horizon research workflows requiring multi-session continuity and quantitative rigor. Here we present Ensemble QSP, a multi-agent framework featuring a three-layer hierarchical memory architecture that keeps injected context bounded and constant in project duration (mid-term project state: median 301 tokens, max 4,050, across 104 runs) by capping each state category and evicting completed work, enabling continuous autonomous operation without context degradation. The system orchestrates five specialist worker agents under domain-expert principal investigators, enforcing physical constraints through physics-based checklists and structured-domain knowledge. Comprehensive benchmarking demonstrates robust autonomous pharmacokinetic-pharmacodynamic model selection without human intervention, consistent result quality across both lower-cost and frontier LLMs, improved PK parameter recovery relative to single-agent baselines, and stable model selection across linguistically diverse prompts of the same task. Feature-level ablation across physiologically based pharmacokinetic (PBPK) models spanning a broad complexity range shows that PI-agent oversight improves debugging efficiency while preserving final accuracy across conditions. The architecture is structurally domain-agnostic, adding a new scientific domain requires only a new PI agent configuration.

Vision-language models (VLMs) and agentic AI have shown strong performance on semantic visual tasks, but it remains unclear whether they can handle the physics and inverse problems that underlie computational imaging. We present ImagingBench, a benchmark of 20 computational imaging tasks spanning five categories: ray and wave optics, image signal processing, inverse reconstruction, computational sensing, and calibration. ImagingBench evaluates three complementary settings: Expert, fixed expert-guided inverse reconstruction; Planner, planner-guided inverse reconstruction; and Forward, forward-system simulation for consistency checking. We benchmark leading proprietary and open-source image-centric multimodal systems, including Gemini, GPT, and Qwen, and compare them with representative task-specific non-agentic baselines. Across tasks, agentic models remain consistently weaker than specialized methods, especially on computational sensing problems such as lensless imaging, event-based reconstruction, time-of-flight imaging, and holography. Planner guidance provides only modest and inconsistent gains over the fixed-prompt Expert baseline. Although the models often generate visually plausible outputs, their reference-based fidelity remains poor, revealing a substantial gap between semantic visual competence and physically grounded imaging performance. ImagingBench provides a unified testbed for measuring this gap and tracking progress in agentic AI for computational imaging.

While current AI agents support increasingly long context windows, tool use, and skill execution for long-horizon tasks, they still require memory systems to effectively leverage historical experience. Existing memory frameworks typically rely on fixed storage, retrieval, and summarization mechanisms, which can be rigid across different tasks and often require manual tuning. To address this limitation, we propose SelfMem, a self-optimizing memory framework. Inspired by prior work on self-improving AI, we follow the principle of "teaching an agent to fish rather than giving it a fish." Instead of forcing the model to follow a predefined memory strategy or format, SelfMem provides an environment with memory tools and feedback signals that allow the agent to explore, evaluate, and refine its own memory strategy. Our results show that SelfMem consistently outperforms retrieval, compression, and agent-memory baselines on BEAM across conversation scales from 100K to 1M tokens. Compared with the strongest baseline, SelfMem improves the official score by 48.7%, 40.8%, and 41.9% at 100K, 500K, and 1M, respectively. Further question-type analysis shows broad robustness across diverse memory demands, and our optimization study shows that model-guided strategy refinement further improves performance.

Agent memory systems are increasingly evaluated against RAG and full-context baselines, but reported gains often mix changes in the memory method with changes in the language model, embedding model, or retrieval pipeline, making it unclear what is actually being measured. We present MemDelta, a controlled evaluation protocol that varies one component at a time on LongMemEval-S (500 questions, 50+ sessions, three model families). Four findings emerge: (1) verbatim RAG matches full-context GPT-4o-mini (47.2% vs. 49.8%, p = 0.34), but the ranking reverses across models: Gemini gains +14pp from full context, while Sonnet gains +31pp from RAG, partly because it refuses 63% of full-context queries; (2) swapping only the embedding model in an identical pipeline shifts accuracy by +6.2pp at n = 500 (p = 0.004), and Mem0 beats MiniLM-RAG by +11pp but loses to cloud-RAG by 1.2pp, so one variable flips the conclusion; (3) agent self-memory (42%) underperforms basic retrieval (47%); (4) on 2 of 6 question types (n = 88), Mem0 matches cloud RAG (72.7% vs. 73.9%, p = 1.0) at 50x the cost, suggesting narrow rather than general gains. We recommend memory evaluations fix embedding models across comparisons, stratify by model family, and report write-path cost before attributing gains to architecture.

Vision foundation tools such as open-vocabulary detectors, segmentation models, and post-processing operators are powerful building blocks for computer vision, but their effectiveness depends heavily on how they are orchestrated: which tools are used, in what order, with what parameters, and under what visual conditions. Existing visual-programming agents typically generate a fixed solution pipeline, making them brittle under dense objects, occlusion, small targets, and domain shift. We introduce VTOS (Vision Tools Orchestration Search), a framework for adaptive visual tool orchestration through joint solution-observer search. VTOS co-searches executable solution programs that compose vision tools such as Grounding DINO, SAM, NMS, and slice-and-detect, together with observer programs that diagnose candidate solutions, identify failure modes, and generate actionable feedback. These observations are accumulated in a shared VisionThoughts knowledge base to guide subsequent search. We evaluate VTOS through two case studies: dense object counting on LVIS-Count and zero-shot plant-disease segmentation on PlantSeg-OOD, which stress different orchestration challenges including threshold calibration, NMS, slicing, mask refinement, and domain generalization. Across both tasks, VTOS outperforms static tool pipelines and agentic visual-programming baselines, specifically in complex settings such as dense, occluded scenes and out-of-distribution segmentation where static pipelines leave measurable headroom, rather than in standard tasks where a single well-calibrated tool already approaches its ceiling.

Language models are remarkably capable at medical question answering, in some cases surpassing the accuracy of general physicians. However, answering questions about wearable health data remains challenging and understudied, as these ubiquitous sensors produce continuous, high-dimensional, and longitudinal data, which is non-trivial to align with text-centric distributions in LLM pretraining. The diversity of sensor modalities and user intents cannot be effectively handled by a fixed reasoning workflow or a single pretrained foundation model. To address these challenges, we propose WEQA, a query-adaptive agent framework that unifies LLM reasoning with specialized wearable analytical and modeling tools. An LLM controller is employed to synthesize execution plans and dynamically route each query to the appropriate combination of sensor analysis and pretrained models, and perform grounded response auditing with external knowledge. We also curate a benchmark spanning four open wearable datasets comprising analytic and predictive tasks in three different health domains. Experiments show that our framework is 24% more accurate than LLM and agentic baselines, and a blinded study with 12 medical experts and 8 users shows substantial gains in usefulness and clinical soundness.

Translating deep learning models from PyTorch's flexible, object-oriented design to JAX's functional, stateless setup is usually a manual and error-prone task. Automated migration is challenging because Large Language Models (LLMs) struggle with strict and dynamic API alignment and are prone to mistakes for exacting operations. We propose a fully autonomous system that combines In-Context Learning (ICL) with oracle-driven self-debugging. First, we curated an ICL context that serves as a strict reference for idiomatic JAX styling and test case generation. Second, instead of depending on the LLM to deduce mathematical outputs, we run the source PyTorch modules to get their actual dynamic tensor states. This creates an unchangeable execution oracle. We then use an autonomous agentic loop to synthesize tests based on the oracle data. The test cases are executed repeatedly, and the traceback is sent back to the LLM for self-correction. Ablations show that combining ICL references with oracle grounding and self-debugging greatly outperforms pure instructional and basic agentic baselines. This improvement does not add an excessive computational overhead. Our lightweight pipeline achieves 91% numerical equivalence (compared to baseline: 9%, instruction + self-debugging: 27%) on neural modules, providing a highly reliable, scalable blueprint for cross-framework migration. This has been validated across several state-of-the-art models including SAM (segment anything), T5, Code Whisper amongst others showing high numerical equivalency. Code: https://github.com/AI-Hypercomputer/accelerator-agents/tree/main/MaxCode

Agentic search over large corpora relies on retriever-mediated interfaces (e.g., BM25 or ColBERT) for scalable candidate discovery. While effective at ranking relevant documents, these interfaces expose evidence only as ranked results or bounded document views, limiting agents' ability to reorganize material and verify constraints across documents. Direct Corpus Interaction (DCI) addresses this limitation by exposing shell-executable corpus operations for flexible search, filtering, comparison, and verification. However, full-corpus terminal commands become slow and unstable as the corpus grows, degrading performance and efficiency. We introduce DR-DCI, a retriever-steered DCI framework that treats retrieval as an agent-callable action for expanding a local workspace. Rather than operating directly over the full corpus, the agent dynamically pulls relevant documents into an evolving workspace and conducts DCI operations within it. This design combines retriever-level recall with DCI-style precision: retrieval keeps exploration scalable, while DCI preserves the local operations needed for effective evidence resolution. Experiments show that DR-DCI is both effective and efficient across scales. On Browsecomp-Plus, DR-DCI reaches 71.2% accuracy, improving over raw DCI and ablated variants by up to 8.3 points while reducing tool usage, wall time, and estimated cost. With workspace-preserving context reset, accuracy further improves to 73.3%. In corpus-scaling experiments, DR-DCI remains effective from 100K to 10M documents, whereas raw DCI becomes unstable and BM25 performs substantially worse. DR-DCI also scales to a 20M-scale file-per-document Wiki-18 QA setting, achieving an average score of 63.0 across six benchmarks and outperforming retrieval-based and trained search-agent baselines. Ablation analysis further shows that ranked previews and inter-document DCI are key to performance.

A central goal of large language model (LLM) research is to build agentic systems that can plan, act, and adapt through sustained interaction with dynamic environments. While recent LLM-based agents exhibit impressive contextual reasoning, their long-horizon decision-making remains fragile, often suffering from objective drift, where goals and plans drift over extended interactions. We introduce Multi$^2$, a hierarchical multi-agent decision-making framework that explicitly decomposes agent behavior into complementary roles. A high-level agent (System 1) focuses on context-aware sub-goal generation using supervised fine-tuning (SFT), while a low-level agent (System 2) executes atomic actions through offline-to-online reinforcement learning (RL) in interactive environments. This separation enables stable long-horizon control, mitigates objective drift, and allows efficient adaptation. Across diverse interactive environments, Multi$^2$ consistently outperforms strong agentic baselines, demonstrating improved robustness and coordination in multi-turn interaction. Beyond performance, we introduce and release three hierarchical benchmark datasets, filling a long-standing gap in training and evaluating hierarchical decision-making for LLM-based agents.

Long-horizon search agents accumulate large amounts of retrieved content across many tool calls, making context-budget efficiency increasingly important. A minimal intervention is to mask stale observations from the context as the trajectory progresses, but it remains unclear when this form of context management helps and why. We study observation masking through a systematic sweep over various agent backbones (4B to 284B parameters) and three retrievers on offline and live-web agentic search benchmarks. We find that the accuracy gain from masking follows an asymmetric inverted-U shape when plotted against the model's accuracy without context management: a plateau under weak retrievers, a peak when a strong retriever meets a mid-capacity model, and a sharp collapse when the model is saturated. This pattern reflects the interaction between retriever recall and the model's implicit filtering capacity, rather than either factor in isolation. Mechanistically, masking implements a token-for-turn trade-off: it removes observations the model has largely stopped attending to and pages the agent rarely re-opens. The added turns help when they convert failures into successes, but they fail when masking removes evidence the model would otherwise have used. We therefore reframe context management as a regime-dependent intervention and provide a holistic perspective for analyzing context use in agentic deep search. We release our scaffold and trajectories here (https://github.com/i-DeepSearch/observation-masking) to support future research.

Scientific figures are among the most effective means of communicating complex research ideas, yet producing publication-quality illustrations remains one of the most labor-intensive parts of paper preparation. Existing automated systems each target a single figure type under text-only input, leaving the diversity of types and conditions researchers actually use unaddressed; their raster outputs further cannot be locally revised. Because scientific figures are structured compositions of discrete semantic components, the localized errors generators produce on such layouts demand not a stronger backbone but a harness. We instantiate this harness in two complementary systems: Crafter, a multi-agent harness for figure generation that generalizes across figure types and input conditions without architectural changes, and CraftEditor, which applies the same pattern to convert raster outputs into editable SVGs. Moreover, we introduce CraftBench, a benchmark spanning three figure types and four input conditions with human quality annotation. Experiments show that Crafter substantially outperforms both standalone generators and the agentic baseline on PaperBanana-Bench and CraftBench, with ablations confirming each component's independent contribution; CraftEditor faithfully converts outputs into editable SVGs that surpass all baselines. Our code and benchmark are available at https://github.com/HaozheZhao/Crafter.

In the era of autonomous agents, machine-actionable data is critical for data-driven workflows. For more than a decade, semantic metadata like schema.org has anchored the FAIR principles (Findable, Accessible, Interoperable, and Reusable) for machine-actionable data and enabled discovery tools like Google Dataset Search. However, the rise of Large Language Models (LLMs) capable of navigating the unstructured web raises a fundamental question: Is semantic metadata still necessary for agentic data discovery, or can agents reliably retrieve actionable data directly from the web? We present a comparative analysis of agentic data retrieval across two distinct environments: a Baseline Agent searching billions of open-web documents, and a Semantic Agent leveraging a corpus of 90 million datasets using schema.org. We deploy an "LLM-as-a-judge" evaluation pipeline, mapped directly to the FAIR principles, to assess the semantic relevance, data accessibility, and computational utility of the retrieved data. Our results reveal a clear divergence. The Semantic Agent excels at retrieving actionable data, achieving a 44.9% higher precision for metadata-rich registries and a 46.6% higher precision for pages with machine-readable downloads among its returned results. Conversely, the Baseline Agent frequently suffers "Last-Mile Utility" failures, retrieving prose-heavy pages (20.1% of results) and portal landing pages (8.5%) rather than actual data pages. While the Baseline Agent achieves higher coverage by answering 40% more questions, the Semantic Agent delivers greater accuracy, achieving 65.7% higher overall precision in retrieving FAIR-compliant datasets. We conclude that while unstructured retrieval supports broad exploratory tasks, structured ecosystems remain the indispensable foundation for reliable, execution-oriented autonomous workflows.

Audio agents extend large audio-language models (LALMs) by decomposing audio questions into tool calls, intermediate evidence, and iterative reasoning steps. However, as LALMs become stronger, the key challenge shifts from enabling tool use to determining when agentic evidence acquisition genuinely benefits audio understanding. We propose Audio-Mind, an auditable and pluggable framework for conditional evidence acquisition in audio understanding. Audio-Mind dynamically combines a strong frontend with planner-guided tool use, preserving frontend judgment when initial evidence is sufficient while acquiring bounded external evidence for questions with unresolved evidence gaps. Experiments on MMAR and MSU-Bench show that Audio-Mind outperforms prior audio-agent baselines, reaching 80.4% accuracy on MMAR and 82.8% accuracy on MSU-Bench. A matched-backbone comparison highlights why this design matters: under strong audio frontends, agentic decomposition can become an orchestration bottleneck when the workflow does not preserve the frontend's holistic audio-grounded judgment. Beyond accuracy, Audio-Mind produces higher-quality, auditable reasoning traces that expose uncertainty, tool evidence, and answer rationales, offering a potential basis for more reliable audio-QA annotation and error analysis.

Generating a game is not the same as making one that can be played. Despite advances in code generation, existing approaches treat game generation as one-shot translation from prompt to artifact, leaving interaction-level failures undetected. We argue that evaluating and improving game generation requires a player, and study two roles for graphical user interface (GUI) agents in this process: (1) as an objective evaluator, for which we introduce PlaytestArena, a new evaluation environment that pairs 200 browser-based game generation tasks across eight genres with rubrics of expected in-play behaviors, adjudicated by a GUI agent that loads each build in a browser and plays it; and (2) as a subjective playtester, for which we propose Play2Code, where a game agent and a GUI agent operate in a sustained loop with shared memory, turning game generation into a dialogue between coding and playing. Our experiments show that even frontier models struggle to generate playable games directly, while Play2Code achieves a 66.8% rubric pass-rate, improving over single-pass and agentic-coding baselines by 37.1 and 14.6 points respectively. Further analysis shows that GUI playtester feedback is more traceable than a human report, yet idiosyncratic in ways reminiscent of human testers, establishing game playtesting as a critical testbed for interactive code generation. Our project website is available at https://continual-game-generation.vercel.app/.

Empirical claims about autonomous Kubernetes operations agents are largely unfalsifiable. Published work reports observational results without controlled comparisons against an agent-disabled baseline, selection bias is endemic, pre-registered decision matrices are absent, and samples are typically too small for the noise level of the underlying scoring system. The cause is the same gap that limits the agents themselves: code agents have a verification substrate that turns "did it work" into a fast, falsifiable, ground-truth signal, and operations has nothing equivalent. We present agent-breakage, a closed-loop measurement framework that injects faults into a target Kubernetes cluster, observes how an autonomous agent responds, scores the response on four axes against ground truth, and accumulates outcome-labeled (state, action, outcome) tuples. The framework distinguishes framework error from reasoning error, supports a true off-condition control via a deterministic-embedder mechanism, and enforces pre-registered decision matrices. We use it as a case study to test whether retrieval over past postmortems compounds an agent's capability. The methodological payload is three confounds the substrate caught during that case study, each of which would have produced a wrong published claim on a less instrumented version of the same work: a pgvector index bug, a +19% selection-bias artifact, and small-sample estimates that overstated effects by roughly 3x. The retrieval result itself is a partial falsification: 1 of 3 dense-corpus scenarios significant at p<0.05, pooled effect +3.9 percentage points, not significant at n=60. A within-scenario corpus-density sweep at 360 runs shows that mechanistic alignment of near-neighbors dominates raw count. The framework is released open source.

Foundation model agents increasingly operate in multi-agent deployments where a coordinator must decide which agent's response to trust. The standard approach weights agents by their self-reported confidence, but recent evidence shows that foundation model confidence is systematically miscalibrated and, on hard tasks, inversely correlated with accuracy. Design-time calibration methods (temperature scaling, Platt scaling, histogram binning) cannot address this problem because they fit a fixed correction to held-out data and degrade under distribution shift. We present MARGIN (Multi-Agent Runtime Grading via Incremental Normalisation), an online calibration method that learns per-agent, per-confidence-band calibration factors from the task stream itself, requiring no model access, no held-out data, and no retraining. MARGIN uses symmetric exponentially weighted moving averages with Bayesian shrinkage blending, and has three hyperparameters with robust defaults. Across 18 foundation models, 8 benchmarks, and over 44,000 observations, MARGIN achieves 3-6x lower calibration error than the best design-time baseline under distribution shift. In multi-agent selection, raw verbalized confidence fails to beat random at pairwise resolution (43-50%) on hard benchmarks. MARGIN corrects this completely, raising pairwise resolution to 70-89% and closing 37-78% of the Raw-to-Oracle pass@1 gap across the five code-generation benchmarks without any oracle knowledge of which model is strongest. Six formal propositions characterize convergence, tracking speed, and the optimality of symmetric updates for non-strategic agents, with all predictions illustrated empirically.

Terminal agents extend Large Language Models with the ability to execute tasks directly in command-line environments, but their progress is bottlenecked by the scarcity of high-quality training data. Existing approaches bootstrap from partial sources such as human-defined seeds or GitHub repositories to instantiate one component and then complete the rest, producing tasks confined to narrow seed distributions, environments misaligned with task semantics, and inefficient trajectories from unguided exploration. To address these limitations, we introduce Terminal-World, a fully automated pipeline that uses agent skills as the central synthesis primitive, which jointly encode what to accomplish, when to apply (preconditions and environment state), and how to execute, enabling task instructions, environments, and teacher trajectories to be co-derived. To further broaden the synthesis space, Terminal-World composes skills into skill teams and skill graphs for multi-role and cross-domain task synthesis. Using this pipeline, we construct 5,723 training environments and train Terminal-World-8B/14B/32B, evaluated across 6 benchmarks where the Terminal-World series consistently outperforms terminal-agent baselines. Notably, using the same teacher model and only 1.2% of the training data, Terminal-World-32B surpasses Nemotron-Terminal-32B on Terminal-Bench 2.0 by +4.5 Pass@1 (31.5) and achieves 43.8 Pass@3.

Recent advances in Multimodal Large Language Models (MLLMs) and agent workflows have shown strong promise for computational pathology, yet reliable patch-level reasoning remains challenging. End-to-end pathology MLLMs often hallucinate morphological features, while recent agentic systems usually merge tool outputs and retrieved knowledge into a shared context, making decisions vulnerable to conflicting evidence and context contamination. We propose PathoSage, a three-stage framework that explicitly separates knowledge retrieval, evidence collection, and evidence adjudication for patch-level pathology multimodal reasoning. Its core component, Structured Evidence Deliberation, independently evaluates heterogeneous evidence from tools, performs conflict analysis, and generates the final judgment in a fresh context to reduce anchoring bias. We further introduce a training-free Beta-Bernoulli experience system with continuous credit assignment to model long-term tool reliability and construct similarity-weighted priors for future tool use. Experiments show that PathoSage effectively mitigates VQA hallucinations and classifier disagreement, outperforming strong pathology MLLM and agentic baselines. Our results highlight explicit evidence adjudication and reliability-aware tool modeling as key ingredients for robust pathology agents.

We introduce PopuLoRA, a population-based asymmetric self-play framework for reinforcement learning with verifiable rewards (RLVR) post-training of LLMs. Teachers and students are specialised LoRA adapters on a shared frozen base: teachers propose problems, matched students solve them under a programmatic verifier, and cross-evaluation between sub-populations replaces the self-calibration that limits single-agent self-play. A family of LoRA weight-space evolution operators (mutations and crossovers that produce same-rank population members in seconds) serves as the replacement step of a population-based training loop at 7B scale. We instantiate PopuLoRA on top of Absolute Zero Reasoner and compare it against a per-adapter compute-matched single-agent baseline. Where the single agent self-calibrates to generating easy problems it can reliably solve, the population enters a co-evolutionary arms race: teachers produce increasingly complex problems, student solve rates oscillate, and problem-space coverage keeps expanding throughout training. Despite lower training-time reward, the population mean outperforms the baseline on three code benchmarks (HumanEval+, MBPP+, LiveCodeBench) and seven math benchmarks (AIME 24/25, AMC 23, MATH-500, Minerva, GSM8K, OlympiadBench), and even the weakest member of the population beats the baseline on aggregate.

Long-term memory is crucial for agents in specialized web environments, where success depends on recalling interface affordances, state dynamics, workflows, and recurring failure modes. However, existing memory benchmarks for agents mostly focus on user histories, short traces, or downstream task success, leaving open how to directly evaluate whether memory systems effectively internalize environment-specific experience. To address this gap, we introduce LongMemEval-V2 (LME-V2), a benchmark for evaluating whether memory systems can help agents acquire the experience needed to become knowledgeable colleagues in customized environments. LME-V2 contains 451 manually curated questions covering five core memory abilities for web agents: static state recall, dynamic state tracking, workflow knowledge, environment gotchas, and premise awareness. Questions are paired with history trajectories containing up to 500 trajectories and 115M tokens. We use a context gathering formulation: memory systems consume history trajectories and return compact evidence for downstream question answering. We propose a suite of two memory methods: AgentRunbook-R, an efficient RAG-based memory with knowledge pools for raw state observations, events, and strategy notes, and AgentRunbook-C, which stores trajectories as files and invokes a coding agent to gather evidence in an augmented sandbox. Experiments show that AgentRunbook-C achieves the best performance with 72.5% average accuracy, outperforming the strongest RAG baseline (48.5%) and the off-the-shelf coding agent baseline (69.3%). Despite the strong performance gains, coding agent based methods have high latency costs. While AgentRunbook-C advances the accuracy-latency Pareto frontier, substantial room for improvement remains. Together, these results establish LME-V2 as a challenging testbed for developing long-term memory systems for environment experience.

Large language models (LLMs) are increasingly used to translate natural-language optimization problems into mathematical formulations and solver code, but matching the reference objective value is not a reliable test of correctness: an artifact may agree numerically while still changing the underlying optimization semantics. We formulate this issue as \emph{optimization-modeling hallucination detection}, namely structural consistency auditing over the problem description, symbolic model, and solver implementation. We develop, to our knowledge, the first fine-grained hallucination taxonomy specifically for optimization modeling, spanning objective, variable, constraint, and implementation failures. We use this taxonomy to design OptArgus, a multi-agent detector with conductor routing, specialist auditors, and evidence consolidation. To evaluate this setting, we introduce a three-part benchmark suite with $484$ clean artifacts, $1266$ controlled injected artifacts, and $6292$ natural LLM-generated artifacts. Against a matched single-agent baseline, OptArgus produces fewer false alarms on clean artifacts, more accurate top-ranked localization on controlled single-error cases, and stronger detection on natural model outputs. Together, these contributions turn optimization-modeling hallucination detection into a concrete empirical problem and suggest that modular, taxonomy-grounded auditing is a practical route to more reliable optimization modeling.

Large language model (LLM)-based multi-agent systems have shown strong potential on complex tasks through agent specialization, tool use, and collaborative reasoning. However, most automated multi-agent system design methods still follow a one-shot paradigm: a workflow is optimized or selected before execution and then reused unchanged throughout the task. This static coordination strategy is ill-suited for long-horizon tasks whose subgoals, intermediate evidence, and information needs evolve over multiple execution stages. We propose EvoMAS, a framework for execution-time multi-agent workflow construction. EvoMAS formulates workflow construction as a meta-level sequential decision problem along a single task trajectory. At each stage, it constructs an explicit task state through a Planner-Evaluator-Updater pipeline and uses a learned Workflow Adapter to instantiate a stage-specific layered workflow from a fixed pool of candidate agents. The adapter is trained with policy gradients using sparse, verifiable terminal task success as the main supervision signal, while evaluator-based process reward is analyzed separately under very-hard sparse-reward settings. Experiments on GAIA, HLE, and DeepResearcher show that EvoMAS outperforms single-agent baselines and recent automated multi-agent workflow design methods. Our analyses further show that explicit task-state construction and learned workflow adaptation provide complementary benefits. Additional results indicate that process reward is most useful when terminal success is extremely sparse, and qualitative case studies illustrate that EvoMAS adapts agent coordination as the task state evolves.

Retrieval-augmented agents are increasingly the interface to large knowledge bases, yet most treat retrieval as a black box: they issue exploratory queries, inspect snippets, and reformulate until evidence emerges. This resembles how a newcomer searches an unfamiliar database rather than how an expert navigates it with strong priors about terminology and likely evidence, causing extra retrieval rounds, latency, and poor recall. We introduce \textit{Superintelligent Retrieval Agent} (SIRA), which casts \emph{superintelligence} in retrieval as compressing multi-round exploratory search into a single corpus-discriminative retrieval action. SIRA does not merely ask which terms are relevant; it asks which terms separate the desired evidence from corpus-level confusers. Offline, an LLM enriches each document with missing search vocabulary; at query time, it predicts evidence vocabulary the query omits; and corpus statistics serve as tool calls that filter terms that are absent, overly common, or unlikely to create retrieval margin. The final step is a single weighted BM25 call combining the query with the validated expansion. Across ten BEIR benchmarks, SIRA achieves the strongest average retrieval performance in our comparison, beating dense retrievers, learned sparse retrievers, and LLM search-agent baselines while using no relevance labels or retriever fine-tuning. On downstream QA, its retrieval-only answer coverage exceeds recent RL-trained agentic QA systems on NQ and HotpotQA. We also introduce \textbf{BrowseComp-Wikipedia}, a hard-search benchmark of 232 BrowseComp-derived queries over a 25,587,229-document Wikipedia index. Even without index-time enrichment, using only grounded Wikipedia categories, SIRA outperforms multi-round Perplexity agents at every budget, reaching 9.70% Recall@1, 15.27% Recall@10, and 36.14% Recall@100.

Large language model (LLM)-based coding assistants have made substantial progress, yet most systems remain reactive, requiring developers to explicitly formulate their needs. Proactive coding assistants aim to infer latent developer intent from integrated development environment (IDE) interactions and repository context, thereby reducing interaction overhead and supporting more seamless assistance. However, research in this direction is limited by the scarcity of large-scale real-world developer behavior data. Existing studies therefore often rely on LLM-simulated IDE traces, whose fidelity to real development behavior remains unclear. In this paper, we investigate this simulation-to-reality gap through a large-scale empirical study. We collect real IDE interaction traces from 1{,}246 experienced industry developers over three consecutive days using a custom Visual Studio Code extension, and construct paired LLM-simulated traces for controlled comparison. Our analysis shows that simulated traces differ substantially from real traces in behavioral diversity, temporal structure, and exploratory patterns. Based on the collected data, we introduce \textbf{ProCodeBench}, a real-world benchmark for proactive intent prediction. Experiments with representative LLMs, retrieval-augmented methods, and agentic baselines show that current approaches remain far from reliable under real IDE traces, suggesting that simulation-based evaluation can overestimate real-world performance. Finally, our training study shows that simulated data cannot replace real data, but can complement it when used before real-world fine-tuning. These findings highlight the importance of real developer behavior data for evaluating and training proactive coding assistants.

Deep reinforcement learning (DRL) has delivered strong results in domains such as Atari and Go, but it still suffers from high sample cost and weak transfer beyond the training setting. A common response is to reuse information from previously trained models through transfer, distillation, ensemble methods, or federated training instead of learning each target task from random initialization. The literature on these mechanisms is fragmented, and published comparisons are hard to interpret because tasks, baselines, and compute budgets differ. This paper presents a PRISMA-guided systematic review of empirical studies on pretrained knowledge reuse in DRL. Starting from 589 records retrieved from IEEE Xplore, the ACM Digital Library, and citation tracing, we screened 570 unique records and assessed 89 full texts. After applying the final eligibility criteria, 15 empirical studies remained in the main synthesis. We analyzed them qualitatively across three factors: source-target similarity, diversity among reused models, and the fairness of comparisons against from-scratch baselines. Three patterns recur across the surviving corpus. First, positive results are concentrated in settings where source and target tasks share substantial structure or where the method includes an explicit gating or alignment mechanism. Second, evidence for ensembles and federated aggregation is promising but sparse and mostly limited to narrow settings. Third, compute-matched comparisons are rare, which weakens claims about efficiency gains over stronger single-agent baselines. The paper contributes a narrower and internally consistent review scope, a study-level synthesis of empirical evidence, and a provisional independence spectrum that should be treated as a hypothesis for future benchmarking rather than a validated metric.

We present StratFormer, a transformer-based meta-agent that learns to simultaneously model and exploit opponents in imperfect-information games through a two-phase curriculum. The first phase trains an opponent modeling head to identify behavioral patterns from action histories while the agent plays a game-theoretic optimal (GTO) policy. The second phase progressively shifts the policy toward best-response (BR) exploitation, guided by a per-opponent regularization schedule tied to exploitability. Our architecture introduces dual-turn tokens -- feature vectors constructed at both agent and opponent decision points -- coupled with bucket-rate features that encode opponent tendencies across five strategic contexts. On Leduc Hold'em, a small poker variant with six cards and two betting rounds, we test against six opponent archetypes at two strength levels each, with exploitability ranging from 0.15 to 1.26 Big Blinds (BB) per hand. StratFormer achieves an average exploitation gain of +0.106 BB per hand over GTO, with peak gains of +0.821 against highly exploitable opponents, while maintaining near-equilibrium safety.

Instructed code editing is a significant challenge for large language models (LLMs). On the EditBench benchmark, 39 of 40 evaluated models obtain a task success rate (TSR) below 60 percent, highlighting a gap between general code generation and the ability to perform instruction-driven editing under executable test constraints. To address this, we propose SAFEdit, a multi-agent framework for instructed code editing that decomposes the editing process into specialized roles to improve reliability and reduce unintended code changes. A Planner Agent produces an explicit, visibility-aware edit plan, an Editor Agent applies minimal, literal code modifications, and a Verifier Agent executes real test runs. When tests fail, SAFEdit uses a Failure Abstraction Layer (FAL) to transform raw test logs into structured diagnostic feedback, which is fed back to the Editor to support iterative refinement. We compare SAFEdit against both prior single-model results reported for EditBench and an implemented ReAct single-agent baseline under the same evaluation conditions. We used EditBench to evaluate SAFEdit on 445 code editing instances in five languages (English, Polish, Spanish, Chinese, and Russian) under varying spatial context variants. SAFEdit achieved 68.6 percent TSR, outperforming the single-model baseline by 3.8 percentage points and the ReAct single-agent baseline by 8.6 percentage points. The iterative refinement loop was found to contribute 17.4 percentage points to SAFEdit's overall success rate. SAFEdit's automated error analysis further indicates a reduction in instruction-level hallucinations compared to single-agent approaches, providing an additional framework component for interpreting failures beyond pass or fail outcomes.

At present, executable visual workflows have emerged as a mainstream paradigm in real-world industrial deployments, offering strong reliability and controllability. However, in current practice, such workflows are almost entirely constructed through manual engineering: developers must carefully design workflows, write prompts for each step, and repeatedly revise the logic as requirements evolve -- making development costly, time-consuming, and error-prone. To study whether large language models can automate this multi-round interaction process, we introduce Chat2Workflow, a benchmark for generating executable visual workflows directly from natural language, and propose a robust agentic baseline to improve performance. The benchmark is built from a large collection of real-world business workflows, with each instance designed so that the generated workflow can be transformed and directly deployed to practical workflow platforms such as Dify and Coze. Experimental results show that while state-of-the-art language models can often capture high-level intent, they struggle to generate correct, stable, and executable workflows, especially given complex and evolving requirements. Although our agentic baseline yields up to 6.05% resolve rate gains, the remaining real-world gap positions Chat2Workflow as a foundation for advancing industrial-grade automation. Code is available at https://github.com/zjunlp/Chat2Workflow.

Test-Time Learning (TTL) enables language agents to iteratively refine their performance through repeated interactions with the environment at inference time. At the core of TTL is an adaptation policy that updates the actor policy based on experience from previous episodes, thereby improving future behavior. Existing methods rely on fixed, hand-crafted adaptation policies rather than optimizing them for downstream improvement. We argue that optimal adaptation policies should be learned from task environments, not hand-engineered based on human intuition. To achieve this, we introduce Meta-TTL, a framework that formulates the discovery of effective adaptation policies as a bi-level optimization problem. Within this framework, the inner loop executes the standard TTL process, measuring how effectively a candidate adaptation policy helps an agent correct errors across sequential episodes. Guided by the agent's performance, the outer loop employs evolutionary search over a diverse distribution of training tasks to continually optimize the adaptation policy. We evaluate Meta-TTL on Jericho, WebArena-Lite, and $τ^2$-bench across both in-distribution (ID) and out-of-distribution (OOD) settings. Results on all three show that Meta-TTL consistently outperforms single-agent, prompt-optimization, and unoptimized meta-agent baselines, suggesting that the optimized adaptation policy encodes transferable strategies that generalize beyond the training task distribution.