Leveraging the general world knowledge of Large Language Models (LLMs) holds significant promise for improving the ability of autonomous driving systems to handle rare and complex scenarios. While integrating LLMs into Vision-Language-Action (VLA) models has yielded state-of-the-art performance, their massive parameter counts pose severe challenges for latency-sensitive and energy-efficient deployment. Distilling LLM knowledge into a compact driving model offers a compelling solution to retain these reasoning capabilities while maintaining a manageable computational footprint. Although previous works have demonstrated the efficacy of distillation, these efforts have primarily focused on relatively simple scenarios and open-loop evaluations. Therefore, in this work, we investigate LLM distillation in more complex, interactive scenarios under closed-loop evaluation. We demonstrate that through a combination of latent feature distillation and ground-truth trajectory supervision, an efficient vision-only student model \textbf{Orion-Lite} can even surpass the performance of its massive VLA teacher, ORION. Setting a new state-of-the-art on the rigorous Bench2Drive benchmark, with a Driving Score of 80.6. Ultimately, this reveals that vision-only architectures still possess significant, untapped potential for high-performance reactive planning.
In classical Reinforcement Learning from Human Feedback (RLHF), Reward Models (RMs) serve as the fundamental signal provider for model alignment. As Large Language Models evolve into agentic systems capable of autonomous tool invocation and complex reasoning, the paradigm of reward modeling faces unprecedented challenges--most notably, the lack of benchmarks specifically designed to assess RM capabilities within tool-integrated environments. To address this gap, we present Plan-RewardBench, a trajectory-level preference benchmark designed to evaluate how well judges distinguish preferred versus distractor agent trajectories in complex tool-using scenarios. Plan-RewardBench covers four representative task families -- (i) Safety Refusal, (ii) Tool-Irrelevance / Unavailability, (iii) Complex Planning, and (iv) Robust Error Recovery -- comprising validated positive trajectories and confusable hard negatives constructed via multi-model natural rollouts, rule-based perturbations, and minimal-edit LLM perturbations. We benchmark representative RMs (generative, discriminative, and LLM-as-Judge) under a unified pairwise protocol, reporting accuracy trends across varying trajectory lengths and task categories. Furthermore, we provide diagnostic analyses of prevalent failure modes. Our results reveal that all three evaluator families face substantial challenges, with performance degrading sharply on long-horizon trajectories, underscoring the necessity for specialized training in agentic, trajectory-level reward modeling. Ultimately, Plan-RewardBench aims to serve as both a practical evaluation suite and a reusable blueprint for constructing agentic planning preference data.
Intelligent systems powered by large-scale sensor networks are shifting from predefined monitoring to intent-driven operation, revealing a critical Semantic-to-Physical Mapping Gap. While large language models (LLMs) excel at semantic understanding, existing perception-centric pipelines operate retrospectively, overlooking the fundamental decision of what to sense and when. We formalize this proactive decision as Semantic-Spatial Sensor Scheduling (S3) and demonstrate that direct LLM planning is unreliable due to inherent gaps in representation, reasoning, and optimization. To bridge these gaps, we introduce the Spatial Trajectory Graph (STG), a neuro-symbolic paradigm governed by a verify-before-commit discipline that transforms open-ended planning into a verifiable graph optimization problem. Based on STG, we implement IoT-Brain, a concrete system embodiment, and construct TopoSense-Bench, a campus-scale benchmark with 5,250 natural-language queries across 2,510 cameras. Evaluations show that IoT-Brain boosts task success rate by 37.6% over the strongest search-intensive methods while running nearly 2 times faster and using 6.6 times fewer prompt tokens. In real-world deployment, it approaches the reliability upper bound while reducing 4.1 times network bandwidth, providing a foundational framework for LLMs to interact with the physical world with unprecedented reliability and efficiency.
Most Human-Machine Interaction (HMI) research overlooks the maneuvering needs of passengers in autonomous driving (AD). Natural language offers an intuitive interface, yet translating passenger open-ended instructions into control signals, without sacrificing interpretability and traceability, remains a challenge. This study proposes an instruction-realization framework that leverages a large language model (LLM) to interpret instructions, generates executable scripts that schedule multiple model predictive control (MPC)-based motion planners based on real-time feedback, and converts planned trajectories into control signals. This scheduling-centric design decouples semantic reasoning from vehicle control at different timescales, establishing a transparent, traceable decision-making chain from high-level instructions to low-level actions. Due to the absence of high-fidelity evaluation tools, this study introduces a benchmark for open-ended instruction realization in a closed-loop setting. Comprehensive experiments reveal that the framework significantly improves task-completion rates over instruction-realization baselines, reduces LLM query costs, achieves safety and compliance on par with specialized AD approaches, and exhibits considerable tolerance to LLM inference latency. For more qualitative illustrations and a clearer understanding.
Large language models (LLMs) have recently demonstrated strong potential for autonomous vehicle motion planning by reformulating trajectory prediction as a language generation problem. However, deploying capable LLMs in resource-constrained onboard systems remains a fundamental challenge. In this paper, we study how to effectively transfer motion planning knowledge from a large teacher LLM to a smaller, more deployable student model. We build on the GPT-Driver framework, which represents driving scenes as language prompts and generates waypoint trajectories with chain-of-thought reasoning, and investigate two student training paradigms: (i) on-policy generalized knowledge distillation (GKD), which trains the student on its own self-generated outputs using dense token-level feedback from the teacher, and (ii) a dense-feedback reinforcement learning (RL) baseline that uses the teacher's log-probabilities as per-token reward signals in a policy gradient framework. Experiments on the nuScenes benchmark show that GKD substantially outperforms the RL baseline and closely approaches teacher-level performance despite a 5$\times$ reduction in model size. These results highlight the practical value of on-policy distillation as a principled and effective approach to deploying LLM-based planners in autonomous driving systems.
Recent advances in Reinforcement Learning with Verifiable Rewards (RLVR) have demonstrated significant potential in single-turn reasoning tasks. With the paradigm shift toward self-evolving agentic learning, models are increasingly expected to learn from trajectories by synthesizing tools or accumulating explicit experiences. However, prevailing methods typically rely on large-scale LLMs or multi-agent frameworks, which hinder their deployment in resource-constrained environments. The inherent sparsity of outcome-based rewards also poses a substantial challenge, as agents typically receive feedback only upon completion of tasks. To address these limitations, we introduce a Tool-Memory based self-evolving agentic framework SEARL. Unlike approaches that directly utilize interaction experiences, our method constructs a structured experience memory that integrates planning with execution. This provides a novel state abstraction that facilitates generalization across analogous contexts, such as tool reuse. Consequently, agents extract explicit knowledge from historical data while leveraging inter-trajectory correlations to densify reward signals. We evaluate our framework on knowledge reasoning and mathematics tasks, demonstrating its effectiveness in achieving more practical and efficient learning.
Modern video games are complex, non-deterministic systems that are difficult to test automatically at scale. Although prior work shows that personality-driven Large Language Model (LLM) agents can improve behavioural diversity and test coverage, existing tools largely remain research prototypes and lack cross-game reusability. This tool paper presents MIMIC-Py, a Python-based automated game-testing tool that transforms personality-driven LLM agents into a reusable and extensible framework. MIMIC-Py exposes personality traits as configurable inputs and adopts a modular architecture that decouples planning, execution, and memory from game-specific logic. It supports multiple interaction mechanisms, enabling agents to interact with games via exposed APIs or synthesized code. We describe the design of MIMIC-Py and show how it enables deployment to new game environments with minimal engineering effort, bridging the gap between research prototypes and practical automated game testing. The source code and a demo video are available on our project webpage: https://mimic-persona.github.io/MIMIC-Py-Home-Page/.
The integration of Artificial Intelligence (AI) with High-Performance Computing (HPC) is transforming scientific workflows from human-directed pipelines into adaptive systems capable of autonomous decision-making. Large language models (LLMs) play a critical role in autonomous workflows; however, deploying LLM-based agents at scale remains a significant challenge. Single-agent architectures and sequential tool calls often become serialization bottlenecks when executing large-scale simulation campaigns, failing to utilize the massive parallelism of exascale resources. To address this, we present a scalable, hierarchical multi-agent framework for orchestrating high-throughput screening campaigns. Our planner-executor architecture employs a central planning agent to dynamically partition workloads and assign subtasks to a swarm of parallel executor agents. All executor agents interface with a shared Model Context Protocol (MCP) server that orchestrates tasks via the Parsl workflow engine. To demonstrate this framework, we employed the open-weight gpt-oss-120b model to orchestrate a high-throughput screening of the Computation-Ready Experimental (CoRE) Metal-Organic Framework (MOF) database for atmospheric water harvesting. The results demonstrate that the proposed agentic framework enables efficient and scalable execution on the Aurora supercomputer, with low orchestration overhead and high task completion rates. This work establishes a flexible paradigm for LLM-driven scientific automation on HPC systems, with broad applicability to materials discovery and beyond.
Conversational diagnosis prediction requires models to track evolving evidence in streaming clinical conversations and decide when to commit to a diagnosis. Existing medical dialogue corpora are largely dyadic or lack the multi-party workflow and annotations needed for this setting. We introduce an ePCR-grounded, topic-flow-based multi-agent generation pipeline that iteratively plans, generates, and self-refines dialogues with rule-based factual and topic flow checks. The pipeline yields EMSDialog, a dataset of 4,414 synthetic multi-speaker EMS conversations based on a real-world ePCR dataset, annotated with 43 diagnoses, speaker roles, and turn-level topics. Human and LLM evaluations confirm high quality and realism of EMSDialog using both utterance- and conversation-level metrics. Results show that EMSDialog-augmented training improves accuracy, timeliness, and stability of EMS conversational diagnosis prediction.
Understanding the landscape of artificial intelligence (AI) and machine learning (ML) adoption across the National Institutes of Health (NIH) portfolio is critical for research funding strategy, institutional planning, and health policy. The advent of large language models (LLMs) has fundamentally transformed research landscape analysis, enabling researchers to perform large-scale semantic extraction from thousands of unstructured research documents. In this paper, we illustrate a human-in-the-loop research methodology for LLMs to automatically classify and summarize research descriptions at scale. Using our methodology, we present a comprehensive analysis of 58,746 NIH-funded biomedical research projects from 2025. We show that: (1) AI constitutes 15.9% of the NIH portfolio with a 13.4% funding premium, concentrated in discovery, prediction, and data integration across disease domains; (2) a critical research-to-deployment gap exists, with 79% of AI projects remaining in research/development stages while only 14.7% engage in clinical deployment or implementation; and (3) health disparities research is severely underrepresented at just 5.7% of AI-funded work despite its importance to NIH's equity mission. These findings establish a framework for evidence-based policy interventions to align the NIH AI portfolio with health equity goals and strategic research priorities.
Recent LLM-based agents often place world modeling, planning, and reflection inside a single language model loop. This can produce capable behavior, but it makes a basic scientific question difficult to answer: which part of the agent's competence actually comes from the LLM, and which part comes from explicit structure around it? We study this question not by claiming a general answer, but by making it empirically tractable. We introduce a declared reflective runtime protocol that externalizes agent state, confidence signals, guarded actions, and hypothetical transitions into inspectable runtime structure. We instantiate this protocol in a declarative runtime and evaluate it on noisy Collaborative Battleship [4] using four progressively structured agents over 54 games (18 boards $\times$ 3 seeds). The resulting decomposition isolates four components: posterior belief tracking, explicit world-model planning, symbolic in-episode reflection, and sparse LLM-based revision. Across this decomposition, explicit world-model planning improves substantially over a greedy posterior-following baseline (+24.1pp win rate, +0.017 F1). Symbolic reflection operates as a real runtime mechanism -- with prediction tracking, confidence gating, and guarded revision actions -- even though its current revision presets are not yet net-positive in aggregate. Adding conditional LLM revision at about 4.3\% of turns yields only a small and non-monotonic change: average F1 rises slightly (+0.005) while win rate drops (31$\rightarrow$29 out of 54). These results suggest a methodological contribution rather than a leaderboard claim: externalizing reflection turns otherwise latent agent behavior into inspectable runtime structure, allowing the marginal role of LLM intervention to be studied directly.
While Large Language Models (LLMs) demonstrate remarkable reasoning capabilities, their potential for purpose-driven exploration in dynamic geo-spatial environments remains under-investigated. Existing Geo-Spatial Question Answering (GSQA) benchmarks predominantly focus on static retrieval, failing to capture the complexity of real-world planning that involves dynamic user locations and compound constraints. To bridge this gap, we introduce EVGeoQA, a novel benchmark built upon Electric Vehicle (EV) charging scenarios that features a distinct location-anchored and dual-objective design. Specifically, each query in EVGeoQA is explicitly bound to a user's real-time coordinate and integrates the dual objectives of a charging necessity and a co-located activity preference. To systematically assess models in such complex settings, we further propose GeoRover, a general evaluation framework based on a tool-augmented agent architecture to evaluate the LLMs' capacity for dynamic, multi-objective exploration. Our experiments reveal that while LLMs successfully utilize tools to address sub-tasks, they struggle with long-range spatial exploration. Notably, we observe an emergent capability: LLMs can summarize historical exploration trajectories to enhance exploration efficiency. These findings establish EVGeoQA as a challenging testbed for future geo-spatial intelligence. The dataset and prompts are available at https://github.com/Hapluckyy/EVGeoQA/.
Large language models (LLMs) are probabilistic in nature and perform more reliably when augmented with external information. As complex queries often require multi-step reasoning over the retrieved information, with no clear or predetermined reasoning path, they remain challenging. Recent approaches train models using reinforcement learning on the model's outcome, showing promise in improving how models handle complex information. We introduce SubSearch, a specialized framework that shifts from outcome-only supervision to intermediate reward signals that incentivize planning high-quality reasoning. Unlike previous work on process reward modeling, which focuses on training a separate reward model with annotated trajectories by either human annotators or large LLM judges, SubSearch directly optimizes the generator using intrinsic process rewards, which we define as internally-derived rewards, eliminating the need for external supervision, and moving towards autonomous information-intensive reasoning. Experiments on seven benchmarks show that rewarding intermediate reasoning steps with intrinsic rewards leads to more robust reasoning traces in both QA and multi-hop QA datasets over using only outcome rewards. SubSearch can help in building reasoning traces that allow agents to better integrate search engines for complex query answering, while offering a data-efficient alternative to supervised process modeling.
The integration of machine learning tools into telecom networks, has led to two prevailing paradigms, namely, language-based systems, such as Large Language Models (LLMs), and physics-based systems, such as Digital Twins (DTs). While LLM-based approaches enable flexible interaction and automation, they lack explicit representations of network dynamics. DTs, in contrast, offer a high-fidelity network simulation, but remain scenario-specific and are not designed for learning or decision-making under uncertainty. This gap becomes critical for 6G systems, where decisions must take into account the evolving network states, uncertainty, and the cascading effects of control actions across multiple layers. In this article, we introduce the {Telecom World Model}~(TWM) concept, an architecture for learned, action-conditioned, uncertainty-aware modeling of telecom system dynamics. We decompose the problem into two interacting worlds, a controllable system world consisting of operator-configurable settings and an external world that captures propagation, mobility, traffic, and failures. We propose a three-layer architecture, comprising a field world model for spatial environment prediction, a control/dynamics world model for action-conditioned Key Performance Indicator (KPI) trajectory prediction, and a telecom foundation model layer for intent translation and orchestration. We showcase a comparative analysis between existing paradigms, which demonstrates that TWM jointly provides telecom state grounding, fast action-conditioned roll-outs, calibrated uncertainty, multi-timescale dynamics, model-based planning, and LLM-integrated guardrails. Furthermore, we present a proof-of-concept on network slicing to validate the proposed architecture, showing that the full three-layer pipeline outperforms single-world baselines and accurately predicts KPI trajectories.
When an LLM-based agent improves on a task, is the gain from the model itself or from the reasoning paradigm wrapped around it? We study this question by comparing six inference-time paradigms, namely Direct, CoT, ReAct, Plan-Execute, Reflection, and ReCode, across four frontier LLMs and ten benchmarks, yielding roughly 18,000 runs. We find that reasoning structure helps dramatically on some tasks but hurts on others: ReAct improves over Direct by 44pp on GAIA, while CoT degrades performance by 15pp on HumanEval. No single paradigm dominates, and oracle per-task selection beats the best fixed paradigm by 17.1pp on average. Motivated by this complementarity, we propose a select-then-solve approach: before answering each task, a lightweight embedding-based router selects the most suitable paradigm. Across four models, the router improves average accuracy from 47.6% to 53.1%, outperforming the best fixed paradigm at 50.3% by 2.8pp and recovering up to 37% of the oracle gap. In contrast, zero-shot self-routing only works for GPT-5 at 67.1% and fails for weaker models, all trailing the learned router. Our results argue that reasoning paradigm selection should be a per-task decision made by a learned router, not a fixed architectural choice.
The aerodynamic design of turbomachinery is a complex and tightly coupled multi-stage process involving geometry generation, performance prediction, optimization, and high-fidelity physical validation. Existing intelligent design approaches typically focus on individual stages or rely on loosely coupled pipelines, making fully autonomous end-to-end design challenging. To address this issue, this study proposes TurboAgent, a large language model (LLM)-driven autonomous multi-agent framework for turbomachinery aerodynamic design and optimization. The LLM serves as the core for task planning and coordination, while specialized agents handle generative design, rapid performance prediction, multi-objective optimization, and physics-based validation. The framework transforms traditional trial-and-error design into a data-driven collaborative workflow, with high-fidelity simulations retained for final verification. A transonic single-rotor compressor is used for validation. The results show strong agreement between target performance, generated designs, and CFD simulations. The coefficients of determination for mass flow rate, total pressure ratio, and isentropic efficiency all exceed 0.91, with normalized RMSE values below 8%. The optimization agent further improves isentropic efficiency by 1.61% and total pressure ratio by 3.02%. The complete workflow can be executed within approximately 30 minutes under parallel computing. These results demonstrate that TurboAgent enables an autonomous closed-loop design process from natural language requirements to final design generation, providing an efficient and scalable paradigm for turbomachinery aerodynamic design.
Large Language Models (LLMs) are driving a shift towards intent-driven development, where agents build complete software from scratch. However, existing benchmarks fail to assess this 0-to-1 generation capability due to two limitations: reliance on predefined scaffolds that ignore repository structure planning, and rigid white-box unit testing that lacks end-to-end behavioral validation. To bridge this gap, we introduce CLI-Tool-Bench, a structure-agnostic benchmark for evaluating the ground-up generation of Command-Line Interface (CLI) tools. It features 100 diverse real-world repositories evaluated via a black-box differential testing framework. Agent-generated software is executed in sandboxes, comparing system side effects and terminal outputs against human-written oracles using multi-tiered equivalence metrics. Evaluating seven state-of-the-art LLMs, we reveal that top models achieve under 43% success, highlighting the ongoing challenge of 0-to-1 generation. Furthermore, higher token consumption does not guarantee better performance, and agents tend to generate monolithic code.
Large language models (LLMs) have shown strong knowledge reserves and task-solving capabilities, but still face the challenge of severe hallucination, hindering their practical application. Though scientific theories and rules can efficiently direct the behaviors of human manipulators, LLMs still do not utilize these highly-condensed knowledge sufficiently through training or prompting. To address this issue, we propose \textbf{SciDC}, an LLM generation method that integrate subject-specific knowledge with strong constraints. By adopting strong LLMs to automatically convert flexible knowledge into multi-layered, standardized rules, we build an extensible framework to effectively constrain the model generation on domain tasks. Experiments on scientific tasks including industrial formulation design, clinical tumor diagnosis and retrosynthesis planning, consistently demonstrate the effectiveness of our method, achieving a 12\% accuracy improvement on average compared with vanilla generation. We further discuss the potential of LLMs in automatically inductively summarizing highly-condensed knowledge, looking ahead to practical solutions for accelerating the overall scientific research process. All the code of this paper can be obtained (https://github.com/Maotian-Ma/SciDC).
Missing-person and child-safety investigations rely on heterogeneous case documents, including structured forms, bulletin-style posters, and narrative web profiles. Variations in layout, terminology, and data quality impede rapid triage, large-scale analysis, and search-planning workflows. This paper introduces the Guardian Parser Pack, an AI-driven parsing and normalization pipeline that transforms multi-source investigative documents into a unified, schema-compliant representation suitable for operational review and downstream spatial modeling. The proposed system integrates (i) multi-engine PDF text extraction with Optical Character Recognition (OCR) fallback, (ii) rule-based source identification with source-specific parsers, (iii) schema-first harmonization and validation, and (iv) an optional Large Language Model (LLM)-assisted extraction pathway incorporating validator-guided repair and shared geocoding services. We present the system architecture, key implementation decisions, and output design, and evaluate performance using both gold-aligned extraction metrics and corpus-level operational indicators. On a manually aligned subset of 75 cases, the LLM-assisted pathway achieved substantially higher extraction quality than the deterministic comparator (F1 = 0.8664 vs. 0.2578), while across 517 parsed records per pathway it also improved aggregate key-field completeness (96.97\% vs. 93.23\%). The deterministic pathway remained much faster (mean runtime 0.03 s/record vs. 3.95 s/record for the LLM pathway). In the evaluated run, all LLM outputs passed initial schema validation, so validator-guided repair functioned as a built-in safeguard rather than a contributor to the observed gains. These results support controlled use of probabilistic AI within a schema-first, auditable pipeline for high-stakes investigative settings.
Bicycle safety is important for bikeability and transportation efficiency. However, conventional surveys often fall short in capturing how people actually perceive cycling environments because they rely heavily on respondents' recall rather than in-the-moment experience. By leveraging large language models (LLMs), this study proposes a new method of combining video-based surveys with a conversational AI chatbot to collect human perceptions of cycling safety and the reasons behind these perceptions. The paper developed the AI chatbot using a modular LLM architecture, integrating prompt engineering, state management, and rule-based control to support the structure of human-AI interaction. This paper evaluates the feasibility of the proposed video-based conversational chatbot using complete responses from sixteen participants to the pilot survey across nine street segments in New York City. The method feasibility was assessed using a seven-point scale rating for user experience (i.e., ease of use, supportiveness, efficiency) and a five-point scale for chatbot usability (i.e., personality, roboticness, friendliness), yielding positive results with mean scores of 5.00 out of 7 (standard deviation = 1.6) and 3.47 out of 5 (standard deviation = 0.43), respectively. The data feasibility was assessed using multiple techniques: (1) Natural language processing (NLP), such as KeyBERT, for overall safety and feature analysis to extract built-environment attributes; (2) K-means clustering for semantic analysis to identify reasons and suggestions; and (3) regression to estimate the effects of built-environment and demographic variables on perceived safety outcomes. The results show the potential of AI chatbots as a novel approach to collecting data on human perception, behavior, and future visions for transport planning.
The viability of chain-of-thought (CoT) monitoring hinges on models being unable to reason effectively in their latent representations. Yet little is known about the limits of such latent reasoning in LLMs. We test these limits by studying whether models can discover multi-step planning strategies without supervision on intermediate steps and execute them latently, within a single forward pass. Using graph path-finding tasks that precisely control the number of required latent planning steps, we uncover a striking limitation unresolved by massive scaling: tiny transformers trained from scratch discover strategies requiring up to three latent steps, fine-tuned GPT-4o and Qwen3-32B reach five, and GPT-5.4 attains seven under few-shot prompting. Although the maximum latent planning depth models can learn during training is five, the discovered strategy generalizes up to eight latent steps at test-time. This reveals a dissociation between the ability to discover a latent strategy under final-answer supervision alone and the ability to execute it once discovered. If similar limits hold more broadly, strategies requiring multiple coordinated latent planning steps may need to be explicitly taught or externalized, lending credence to CoT monitoring.
Instruction tuning is commonly assumed to endow language models with a domain-general ability to follow instructions, yet the underlying mechanism remains poorly understood. Does instruction-following rely on a universal mechanism or compositional skill deployment? We investigate this through diagnostic probing across nine diverse tasks in three instruction-tuned models. Our analysis provides converging evidence against a universal mechanism. First, general probes trained across all tasks consistently underperform task-specific specialists, indicating limited representational sharing. Second, cross-task transfer is weak and clustered by skill similarity. Third, causal ablation reveals sparse asymmetric dependencies rather than shared representations. Tasks also stratify by complexity across layers, with structural constraints emerging early and semantic tasks emerging late. Finally, temporal analysis shows constraint satisfaction operates as dynamic monitoring during generation rather than pre-generation planning. These findings indicate that instruction-following is better characterized as skillful coordination of diverse linguistic capabilities rather than deployment of a single abstract constraint-checking process.
The rapid advancement of Large Language Models (LLMs) has created new opportunities for Automated Penetration Testing (AutoPT), spawning numerous frameworks aimed at achieving end-to-end autonomous attacks. However, despite the proliferation of related studies, existing research generally lacks systematic architectural analysis and large-scale empirical comparisons under a unified benchmark. Therefore, this paper presents the first Systematization of Knowledge (SoK) focusing on the architectural design and comprehensive empirical evaluation of current LLM-based AutoPT frameworks. At systematization level, we comprehensively review existing framework designs across six dimensions: agent architecture, agent plan, agent memory, agent execution, external knowledge, and benchmarks. At empirical level, we conduct large-scale experiments on 13 representative open-source AutoPT frameworks and 2 baseline frameworks utilizing a unified benchmark. The experiments consumed over 10 billion tokens in total and generated more than 1,500 execution logs, which were manually reviewed and analyzed over four months by a panel of more than 15 researchers with expertise in cybersecurity. By investigating the latest progress in this rapidly developing field, we provide researchers with a structured taxonomy to understand existing LLM-based AutoPT frameworks and a large-scale empirical benchmark, along with promising directions for future research.
We introduce LudoBench, a benchmark for evaluating LLM strategic reasoning in Ludo, a stochastic multi-agent board game whose dice mechanics, piece capture, safe-square navigation, and home-path progression introduce meaningful planning complexity. LudoBench comprises 480 handcrafted spot scenarios across 12 behaviorally distinct decision categories, each isolating a specific strategic choice. We additionally contribute a fully functional 4-player Ludo simulator supporting Random, Heuristic, Game-Theory, and LLM agents. The game-theory agent uses Expectiminimax search with depth-limited lookahead to provide a principled strategic ceiling beyond greedy heuristics. Evaluating six models spanning four model families, we find that all models agree with the game-theory baseline only 40-46% of the time. Models split into distinct behavioral archetypes: finishers that complete pieces but neglect development, and builders that develop but never finish. Each archetype captures only half of the game theory strategy. Models also display measurable behavioral shifts under history-conditioned grudge framing on identical board states, revealing prompt-sensitivity as a key vulnerability. LudoBench provides a lightweight and interpretable framework for benchmarking LLM strategic reasoning under uncertainty. All code, the spot dataset (480 entries) and model outputs are available at https://anonymous.4open.science/r/LudoBench-5CBF/
RL training of multi-turn LLM agents is inherently unstable, and reasoning quality directly determines task performance. Entropy is widely used to track reasoning stability. However, entropy only measures diversity within the same input, and cannot tell whether reasoning actually responds to different inputs. In RAGEN-2, we find that even with stable entropy, models can rely on fixed templates that look diverse but are input-agnostic. We call this template collapse, a failure mode invisible to entropy and all existing metrics. To diagnose this failure, we decompose reasoning quality into within-input diversity (Entropy) and cross-input distinguishability (Mutual Information, MI), and introduce a family of mutual information proxies for online diagnosis. Across diverse tasks, mutual information correlates with final performance much more strongly than entropy, making it a more reliable proxy for reasoning quality. We further explain template collapse with a signal-to-noise ratio (SNR) mechanism. Low reward variance weakens task gradients, letting regularization terms dominate and erase cross-input reasoning differences. To address this, we propose SNR-Aware Filtering to select high-signal prompts per iteration using reward variance as a lightweight proxy. Across planning, math reasoning, web navigation, and code execution, the method consistently improves both input dependence and task performance.
Spec-driven development (SDD) with AI coding agents provides a structured workflow, but agents often remain "context blind" in large, evolving repositories, leading to hallucinated APIs and architectural violations. We present Spec Kit Agents, a multi-agent SDD pipeline (with PM and developer roles) that adds phase-level, context-grounding hooks. Read-only probing hooks ground each stage (Specify, Plan, Tasks, Implement) in repository evidence, while validation hooks check intermediate artifacts against the environment. We evaluate 128 runs covering 32 features across five repositories. Context-grounding hooks improve judged quality by +0.15 on a 1-5 composite LLM-as-judge score (+3.0 percent of the full score; Wilcoxon signed-rank, p < 0.05) while maintaining 99.7-100 percent repository-level test compatibility. We further evaluate the framework on SWE-bench Lite, where augmentation hooks improve baseline by 1.7 percent, achieving 58.2 percent Pass@1.
As LLM reasoning performance plateau, improving inference-time compute efficiency is crucial to mitigate overthinking and long thinking traces even for simple queries. Prior approaches including length regularization, adaptive routing, and difficulty-based budget allocation primarily focus on single-turn settings and fail to address the sequential dependencies inherent in multi-turn reasoning.In this work, we formulate multi-turn reasoning as a sequential compute allocation problem and model it as a multi-objective Markov Decision Process. We propose TAB: Turn-Adaptive Budgets, a budget allocation policy trained via Group Relative Policy Optimization (GRPO) that learns to maximize task accuracy while respecting global per-problem token constraints. Consequently, TAB takes as input the conversation history and learns to adaptively allocate smaller budgets to easier turns and save appropriate number of tokens for the crucial harder reasoning steps. Our experiments on mathematical reasoning benchmarks demonstrate that TAB achieves a superior accuracy-tokens tradeoff saving up to 35% tokens while maintaining accuracy over static and off-the-shelf LLM budget baselines. Further, for systems where a plan of all sub-questions is available apriori, we propose TAB All-SubQ, a budget allocation policy that budgets tokens based on the conversation history and all past and future sub-questions saving up to 40% tokens over baselines.
The use of LLMs for code generation has naturally extended to code testing and evaluation. As codebases grow in size and complexity, so does the need for automated test generation. Current approaches for LLM-based test generation rely on strategies that maximize immediate coverage gain, a greedy approach that plateaus on code where reaching deep branches requires setup steps that individually yield zero new coverage. Drawing on principles of Bayesian exploration, we treat the program's branch structure as an unknown environment, and an evolving coverage map as a proxy probabilistic posterior representing what the LLM has discovered so far. Our method, CovQValue, feeds the coverage map back to the LLM, generates diverse candidate plans in parallel, and selects the most informative plan by LLM-estimated Q-values, seeking actions that balance immediate branch discovery with future reachability. Our method outperforms greedy selection on TestGenEval Lite, achieving 51-77% higher branch coverage across three popular LLMs and winning on 77-84% of targets. In addition, we build a benchmark for iterative test generation, RepoExploreBench, where they achieve 40-74%. These results show the potential of curiosity-driven planning methods for LLM-based exploration, enabling more effective discovery of program behavior through sequential interaction
We study compiled AI, a paradigm in which large language models generate executable code artifacts during a compilation phase, after which workflows execute deterministically without further model invocation. This paradigm has antecedents in prior work on declarative pipeline optimization (DSPy) and hybrid neural-symbolic planning (LLM+P); our contribution is a systems-oriented study of its application to high-stakes enterprise workflows, with particular emphasis on healthcare settings where reliability and auditability are critical. By constraining generation to narrow business-logic functions embedded in validated templates, compiled AI trades runtime flexibility for predictability, auditability, cost efficiency, and reduced security exposure. We introduce (i) a system architecture for constrained LLM-based code generation, (ii) a four-stage generation-and-validation pipeline that converts probabilistic model output into production-ready code artifacts, and (iii) an evaluation framework measuring operational metrics including token amortization, determinism, reliability, security, and cost. We evaluate on two task types: function-calling (BFCL, n=400) and document intelligence (DocILE, n=5,680 invoices). On function-calling, compiled AI achieves 96% task completion with zero execution tokens, breaking even with runtime inference at approximately 17 transactions and reducing token consumption by 57x at 1,000 transactions. On document intelligence, our Code Factory variant matches Direct LLM on key field extraction (KILE: 80.0%) while achieving the highest line item recognition accuracy (LIR: 80.4%). Security evaluation across 135 test cases demonstrates 96.7% accuracy on prompt injection detection and 87.5% on static code safety analysis with zero false positives.
Clinical diagnosis requires sequential evidence acquisition under uncertainty. However, most Large Language Model (LLM) based diagnostic systems assume fully observed patient information and therefore do not explicitly model how clinical evidence should be sequentially acquired over time. Even when diagnosis is formulated as a sequential decision process, it is still challenging to learn effective diagnostic trajectories. This is because the space of possible evidence-acquisition paths is relatively large, while clinical datasets rarely provide explicit supervision information for desirable diagnostic paths. To this end, we formulate sequential diagnosis as a Latent Diagnostic Trajectory Learning (LDTL) framework based on a planning LLM agent and a diagnostic LLM agent. For the diagnostic LLM agent, diagnostic action sequences are treated as latent paths and we introduce a posterior distribution that prioritizes trajectories providing more diagnostic information. The planning LLM agent is then trained to follow this distribution, encouraging coherent diagnostic trajectories that progressively reduce uncertainty. Experiments on the MIMIC-CDM benchmark demonstrate that our proposed LDTL framework outperforms existing baselines in diagnostic accuracy under a sequential clinical diagnosis setting, while requiring fewer diagnostic tests. Furthermore, ablation studies highlight the critical role of trajectory-level posterior alignment in achieving these improvements.