The paper presents a reinforcement learning framework for designing solar PV adoption policies under uncertainty, integrating RL with a stochastic agent‑based model to simulate yearly adoption over a 16‑year horizon. Policymakers can choose annual incentives such as grants, subsidised loans, and feed‑in tariffs, and the study evaluates three RL algorithms—PPO, SAC, and TD3—within a scalarised reward framework that balances adoption gains against costs. Results show clear trade‑off patterns, with TD3 yielding the highest adoption at higher cost, PPO achieving the lowest cost with fewer adopters, and a balanced PPO policy offering a middle ground, all outperforming static baseline policies.
By Iias Faiud, Jonaid Shianifar, Michael Schukat, Karl Mason
TruthInsightBench is a new benchmark designed to evaluate automated scientific discovery agents by presenting them with 40 blind tasks drawn from peer‑reviewed studies across ten domains. Each task provides only a neutral objective and frozen data, withholding source conclusions, expected values, and analysis paths, forcing agents to determine which claim the data support. A fixed LLM‑based judge scores agents on evidentiary maturity across six dimensions, using 29 artifact‑grounded items, enabling fully automated, repeatable evaluation without human grading.
By Zhibo Yang, Chen Zhang, Yuewei Zhang, Hao Wang
Compact Bellman-Grounded Cognitive Maps (BCM) are introduced as a new method for cost-aware navigation that reuses a single learned map for different goals without per-goal retraining. BCM grounds the map in local edge costs using a self-supervised Bellman objective and a compact coordinate encoding, achieving near-optimal performance on weighted grids up to 1600 nodes with only a 5% mean gap to exact Dijkstra search. Its memory footprint grows sublinearly with graph size, outperforming connectivity-based spectral baselines and demonstrating scalability to complex environments.
By Yuzhe Han, Mingkun Xu, Yujie Wu
The paper introduces the Mirror Agent Model, a Bayesian architecture designed to produce interpretable agent behavior and explanations. It defines an observer model that mirrors the agent’s internal state, enabling explicit and implicit communication of intentions. The architecture incorporates off‑the‑shelf saliency methods to generate explanations, and the authors present preliminary qualitative results demonstrating its potential.
By Michele Persiani, Thomas Hellstr\"om
The paper proposes a distributional theory explaining how large language models (LLMs) incorporate external evidence into their decision-making process. It identifies three key predictions: (1) evidence is more persuasive when it aligns with the model’s prior beliefs, (2) models more readily accept errors from their own internal processes than from external sources, and (3) the same evidence can improve weaker models while harming stronger ones. Extensive experiments across ten million trials, twelve LLMs from four families, and eight domains—including quantum mechanics, physics, genetics, and molecular biology—confirm these predictions and reveal that evidence integration occurs late in the network as a structured sequence of steps rather than through a simple trust metric.
By Sebastien Kawada, Manolis Kellis
Repeat-After-Me is a black-box adaptive visual prompt injection technique that can reveal personally identifiable information or trigger malicious tool calls in both open-weight and commercial vision‑language models, achieving attack success rates above 80% on Qwen3.6‑27B and 47% on GPT‑5.5. The method works even when the benign user prompt is unrelated to the injected task and does not explicitly authorize it, and it retains significant effectiveness when transferred across models or optimized on surrogate systems. In a real‑world OpenClaw Discord deployment, a minimally injected image can overwrite TOOLS.md, enabling remote code execution and secret exfiltration.
By Sizhe Chen, Yu-Lin Tsai, Ivan Evtimov, Kamalika Chaudhuri, Raluca Ada Popa, David Wagner, Arman Zharmagambetov
The paper introduces GRACE, a framework that breaks down large language model (LLM) responses into atomic claims and grounds them against trusted knowledge priors using a weighted bipartite graph. Edge weights enable weighted centrality analysis to classify claims as Grounded, Refuted, or Boundary, identifying hallucinations and frontier knowledge. An objective called Return on Attention (RoA) prioritizes expert review only for high‑uncertainty claims, and verified claims become new evidence anchors, creating a loop that expands the knowledge base across iterations.
By John Seon Keun Yi, Joshua R. Minot, Dokyun Lee
AccidentSim is a framework that generates physically realistic vehicle collision videos by extracting physical clues from real-world accident reports. It uses a reliable physical simulator to replicate post-collision trajectories, builds a trajectory dataset, fine‑tunes a language model to predict consistent trajectories from user prompts, and finally renders high‑quality videos with Neural Radiance Fields. The resulting videos show strong visual and physical authenticity compared to existing methods.
By Xiangwen Zhang, Qian Zhang, Longfei Han, Qiang Qu, Xiaoming Chen, Weidong Cai
SCRIPT is a scalable diffusion policy that uses a Joint Action-State-Text Diffusion Transformer (JAST‑DiT) to jointly encode actions, physical states, and natural‑language instructions, enabling direct interaction between language semantics and control dynamics. The method employs a multi‑stage training framework, including supervised imitation pre‑training, a nonlinear history conditioning mechanism for stable autoregressive control, and a post‑training stage with Reinforcement Learning with Hybrid Rewards (RLHR) that injects learnable noise to improve motion quality and instruction following. Experiments on the 1200‑hour MotionMillion dataset show that SCRIPT outperforms prior state‑of‑the‑art methods across text alignment, motion quality, and physical realism, and its performance scales consistently with model size.
By Jingyan Zhang, Han Liang, Ruichi Zhang, Bin Li, Juze Zhang, Xin Chen, Jingya Wang, Lan Xu, Jingyi Yu
The paper examines how AI coding agents are evolving beyond simple autocomplete to perform complex tasks such as repository inspection, multi-file editing, tool execution, test writing, pull request creation, and long-duration work with minimal supervision. It highlights that while these agents boost coding activity, significant bottlenecks remain in review, integration, testing, security, deployment, and production operations, and that the economics of software development are shifting toward variable token, tool, sandbox, CI, and rework costs. The authors synthesize recent research and industry data to propose four engineering concepts—Agentic SDLC Throughput Paradox, Production-Qualified Change, Verification Tax, and an Agentic SDLC Control Plane—to guide the allocation of autonomy within cost, reliability, and human-attention constraints, ultimately reframing the research focus to production-qualified value per dollar, reviewer-hour, and operational risk.
By Happy Bhati
The paper presents a single pretrained diffusion traffic model that serves both as an ego motion planner and as a controllable generator of safety‑critical scenarios for autonomous driving. It introduces a Single‑Stream Dual‑Stream diffusion‑transformer decoder (SSDS) that fuses scene context via joint attention, improving closed‑loop performance on the nuPlan benchmark, and a training‑free guidance scheme called Decoupled Annealing Posterior Sampling with Energy (DAPSE) that injects arbitrary energy functions at inference time. Using the same model, the authors generate realistic long‑tail driving interactions—such as aggressive cut‑ins and lead‑vehicle braking—through inference‑time guidance, exposing failure modes in black‑box planners that standard benchmarks miss.
By Arka Pal, Rajesh Kumar, Hannes Eriksson, R\'emi Lacombe, Arvid Laveno Ling, Ankit Gupta, Maciej Wozniak
The paper introduces the Neutral Prompting Attack (NPA), a stealthy method that uses semantically benign instructions to increase the likelihood of large language models hallucinating non‑existent package names in coding agents. Unlike traditional dependency steering, NPA does not target a specific package but shifts the model’s output toward more speculative names. Experiments across multiple LLMs show that NPA raises hallucination rates, affects pip install success, alters the distribution of hallucinated packages, and bypasses existing static‑analysis, LLM‑based, and agent‑based defenses.
By Chia-Yi Hsu, Chia-Mu Yu, Chun-Ying Huang, Jun Sakuma
The paper introduces an online skill‑evolution framework that transforms interaction traces and evaluator feedback into a persistent, versioned library of reusable procedures for computer‑use agents. By executing each iteration against a frozen library snapshot, the system updates skills without altering the underlying model parameters. Experiments across four OSWorld domains show that the evolving library consistently outperforms an empty‑library baseline, with gains ranging from 5.7 to 18.6 percentage points, while also revealing domain‑specific temporal stability and challenges in skill retrieval and revision.
By Longtao Hu, Xiao Liang, Linchao Zhu
Trace2Tower is a transition‑aware EigenTrace framework that transforms raw execution traces of large language model agents into a robust skill hierarchy. By abstracting step‑level interactions into canonical events and constructing a unified graph based on semantic compatibility, transition dynamics, and outcome evidence, it isolates stable, success‑aligned behavioral modes through contrastive spectral decomposition. These modes populate a dynamic skill tower of action templates, procedural routines, and overarching task strategies, which are continuously refined via verifier‑guided feedback, achieving superior performance on ALFWorld and WebShop benchmarks.
By Jiazheng Sun, Boyu Yang, Binhao Yuan, Mingxuan Li, Xin Peng
The paper presents a decentralized navigation framework for composite heterogeneous robots that integrates a large language model (LLM) policy agent, an Upper Confidence Bound (UCB) bandit, and a Double Deep Q-Network (Double DQN) controller. Each robot independently generates and refines policies at the round level using LLM inference, while the Double DQN handles tick-level action selection based on navigation variables and LLM priors. Across 30 rounds, the full configuration achieved all goals with the lowest median completion time (42 ticks) and a 25–39% improvement over other setups.
By Chongwen Dong, Mithun Paul Saint-Germain, Pinjari Asif, Carlo R. daCunha
The paper introduces CONTINUITY, a framework that ensures secure composition of large language model (LLM) agent controls by using assume‑guarantee contracts and authenticated security contexts. It models each component with signed root grants, provenance commitments, and other mechanisms to carry security context across transitions, formalizing end‑to‑end consequence integrity. A reference verifier and fault‑injection suite demonstrate that the full configuration prevents harmful effects while completing all benign tasks and correctly escalating ambiguous cases.
By Chris Zheng, Geng Yang
The paper introduces SMART, a symbolic performance‑modeling library for machine‑learning systems that relies almost entirely on natural‑language design documents rather than code. By using AI coding agents to regenerate the implementation from these documents, the framework eliminates the need for continuous refactoring as models and systems evolve. The authors demonstrate that regenerated implementations match hand‑audited reference models to round‑off precision, suggesting that design documents can serve as the durable artifact for ML‑systems co‑design tools.
By Samuel Kushnir, Kimia Noorbakhsh, Kavya Sreedhar, Liqun Cheng, Ming Liu, Parthasarathy Ranganathan, Mohammad Alizadeh, Fred Kjolstad, Suvinay Subramanian
The paper presents an active inference route‑planning method for autonomous agents tasked with reconnaissance missions. It builds an evidence map that integrates both positive and negative sensor observations over time, using Dempster‑Shafer theory and a Gaussian sensor model to update a posterior probability distribution. By computing variational free energy across positions, the agent moves incrementally toward locations that minimize free energy, balancing exploration of large areas with exploitation of identified targets.
By Johan Schubert, Farzad Kamrani, Tove Gustavi
MemMA is a plug‑and‑play multi‑agent framework that coordinates the memory cycle of memory‑augmented LLM agents on both forward and backward paths. On the forward path, a Meta‑Thinker guides a Memory Manager for construction and a Query Reasoner for iterative retrieval. On the backward path, MemMA performs in‑situ self‑evolving memory construction, generating probe QA pairs, verifying the memory, and converting failures into repair actions before finalization. Experiments on LoCoMo show that MemMA consistently outperforms existing baselines across multiple LLM backbones and improves three different storage backends.
By Minhua Lin, Zhiwei Zhang, Hanqing Lu, Hui Liu, Xianfeng Tang, Qi He, Xiang Zhang, Suhang Wang
The paper surveys 87 influential studies on machine‑learning‑based automated vulnerability detection (ML4AVD), categorizing them by problem formulation, input and detection granularity, target languages, evaluation metrics, datasets, and detection approaches. It identifies twelve self‑reinforcing pain points—such as overreliance on binary classification of C/C++ function‑level vulnerabilities, limited language coverage, and intertwined datasets, baselines, and metrics—that trap the field in a narrow, artificial problem space. The authors propose concrete recommendations to break these feedback loops and evaluate a recent high‑profile effort, AIxCC, against these guidelines, reflecting on ML4AVD’s relevance amid the rise of agentic AI.
By Dan Ristea, Shae McFadden, Ezzeldin Shereen, Madeleine Dwyer, Sanyam Vyas, Chris Hicks, Vasilios Mavroudis