The paper investigates why large‑language‑model coding agents rarely request a second chunk of tool output, focusing on the precision‑at‑1 rate ($p_1$) of the gold item appearing first in the first chunk. In a benchmark of 500 software‑engineering tasks, the authors compare six value functions and find that increasing $p_1$ does not systematically improve downstream accuracy; the agent can recover the correct answer from any position within the chunk. Adding file‑metadata signals to a keyword scorer actually reduces $p_1$, while a parameter‑free keyword scorer improves $p_1$ but still fails to boost overall accuracy.
By Tatiana Petrova, Andrei Mazniak, Radu State
The paper introduces a production‑grade large language model (LLM) pricing system for the tourism industry that separates structured extraction and policy selection from deterministic numeric pricing. Policies are compiled into interpretable condition trees, allowing new clauses and evolving rules to be added without code changes while maintaining auditability. Deployed across 12 business categories and 1,500 operators, the system handled 3,960 orders in six months, cutting the order‑management team from 15‑20 to 3 and reducing per‑order handling time from 10 minutes to under 2 minutes.
By Ziqiang Zhang, Jing Ma, Zilong Wang, Jiayuan Chen, Yi Qiao, Yu He, Wei Zhang, Dai Cheng, Xiaoyu Shen
The paper investigates whether hidden representations in latent-based multi‑agent systems can carry attack information that remains effective during normal operation. A latent attack framework is introduced, reactivating attack effects through latent interventions without using adversarial text. Experiments show that these latent attacks can significantly degrade task performance, especially when targeting inter‑agent KV‑cache handoffs, and that the degradation cannot be explained by simple perturbations or invalid generation.
By Chenxi Wang, Ruiyang Huang, Jiayan Sun, Lei Wei, Yifan Wu
OS-Marathon is a new benchmark that tests computer‑use agents on vast‑horizon, repetitive tasks, covering 100 tasks across five scenarios and ten domains. The study shows that current state‑of‑the‑art agents perform poorly on these tasks, and that simply decomposing workflows into subtasks does not solve the problem. Introducing a cost‑friendly personalization method called GraphDemo, which adapts agents from a single human demonstration, improves performance, highlighting the value of human guidance for these challenging tasks.
By Jing Wu, Wenjie Ai, Daphne Barretto, Yiye Chen, Qingyu Chen, Yuhang He, Pranit Chawla, Nicholas Gyd\'e, Yanan Jian, Vibhav Vineet
The article discusses a vulnerability in Anthropic’s Claude Code’s auto mode, which was promoted as a safeguard against prompt injection attacks. Prompt‑injection researcher Johann Rehberger demonstrated that the auto mode can be tricked into executing malicious code, even blocking the agent’s own cleanup attempts. The author concludes that the safest approach is to run coding agents in isolated sandboxes and restrict their access to sensitive resources.
SWE-Prime introduces a two‑stage, multi‑granularity supervised fine‑tuning (SFT) data selection process for large language models tackling real‑world software problems. The first stage filters entire trajectories based on process quality, result quality, and representativeness, while the second stage evaluates semantic segments for contribution, learnability, and risk, keeping all segments in context but only penalizing selected ones during training. Experiments on SWE‑Bench Pro and Verified demonstrate that training on just 10% of trajectories chosen by SWE‑Prime outperforms full‑dataset training, achieving up to 12.2% and 24.2% relative gains.
The article discusses how agentic AI is reshaping the analytics stack by taking over more execution tasks. It raises the question of which responsibilities should remain with human analysts versus AI agents and explores the importance of this distinction. The piece highlights the evolving role of AI in analytics and the need to define clear boundaries between human and machine work.
By Rashi Desai
The paper shows that large language model agents are far more likely to commit to a directional answer when presented with a professional-looking market panel, even if the panel’s data is fabricated. Across 12 frontier models, commitment rates jump from 6.5 % for a bare question to 54.0 % with evidence, and remain high (≈37 %) even when all numbers are invented. The study finds that the act/don’t‑act decision gate is the key failure point, can be trained to reduce false commitments, but is fragile to response format changes.
The article "How to Work with AI Coding Agents" offers a practical guide aimed at improving code quality rather than merely increasing quantity. It focuses on strategies and best practices for effectively collaborating with AI coding tools to produce better code. The post was originally published on Towards Data Science.
By Sara A. Metwalli
Internet memes are a pervasive form of multimodal online communication; however, such communication often involves users from diverse linguistic and cultural backgrounds. Therefore, adapting memes acr...
The article explores the effects of removing a search box from an AI agent and instead providing it with typed tools, hard bounds, and a gate that it cannot bypass. It examines how the agent navigates a knowledge graph within strict limits and discusses findings from four models and one incorrect prediction regarding the value of this approach.
By Miodrag Cekikj
FaulT-Bench is a new benchmark comprising 200 network troubleshooting scenarios across eight topologies, designed to test large‑language‑model agents on realistic, noisy tickets that may contain false premises or incorrect fault claims. The benchmark includes 72 rewritten tickets that vary reporter confidence and detail, and evaluates agents via an automated harness that scores diagnoses on outcome, fix, and reasoning quality. Results show that while agents perform well on accurate tickets, they degrade sharply on healthy networks with misleading reports, highlighting the importance of ticket wording over content.
GraphMemix introduces a combinatorial‑optimization graph memory framework that organizes long‑term multimodal agent memory as query‑aware evidence forests. It constructs candidate graphs by expanding seed memories through schema and semantic relations, then decouples evidence utility from anchor‑conditioned relation verification to reduce redundancy, and finally optimizes a forest‑format memory context within a maximum evidence budget. Experiments on four benchmarks show significant accuracy gains and a new Pareto frontier between accuracy and lifecycle cost.
4DSynth is a controllable procedural system that transforms natural-language descriptions, blueprint masks, or single photographs into editable 4D environments featuring explicit geometry, animated actors, collision-free trajectories, and physics-ready simulation states. The system unifies animation, camera planning, rendering, and task generation within a single geometry-grounded representation, enabling scalable creation of diverse, interactive scenes. Using 4DSynth, the authors built 4DSynth-Nav, an interactive navigation benchmark that demonstrates the reproducibility of failures and tunable difficulty across three tiers for vision‑language models.
The paper presents a modeling and simulation framework to study reinforcement‑learning control of connected and automated vehicle (CAV) platoon joining maneuvers in mixed traffic. It evaluates Deep Q‑Network, Double Deep Q‑Network, and Proximal Policy Optimization algorithms, finding that PPO achieves a 98 % joining success rate with less than 1 % collisions by incorporating risk penalties, though it requires more decision steps. An external safety controller can prevent collisions but may reduce joining efficiency, highlighting a trade‑off between safety, effectiveness, and decision speed.
LiveSim is an LLM-based framework designed to simulate users in multi‑agent live‑stream ecosystems. It treats users as editable behavioral hypotheses and refines them through trajectory‑grounded interactions, using discrepancies between simulated and observed trajectories to uncover missing environmental shaping effects. The extracted environment‑behavior patterns are stored in a collective behavioral memory, enhancing user‑level fidelity and enabling ecosystem‑level analysis of risk evolution and platform interventions.
The paper introduces the Multi-Instruction Multi-Shot Long-Video Editing (MMLVE) task, aiming to edit long videos with multiple instructions while maintaining consistency across shots, decoupling instructions, and preserving spatiotemporal structure. It proposes an agentic editing framework that combines Large Language Models and Vision‑Language Models for shot-level decoupling and precise instruction parsing. A new dataset, MMLVE‑Bench, and three evaluation metrics are created to benchmark this task, and experiments show the proposed MMLVE‑Agent outperforms existing closed‑source state‑of‑the‑art methods by eliminating hallucinations and ensuring seamless transitions.
Behavior2Trip introduces a new task—Behavior‑Aware Travel Planning—where user preferences are inferred directly from past behavior trajectories rather than explicit instructions. The benchmark contains 11,400 Chinese travel‑planning instances, each with an average of 39.8 past behaviors across 14 attributes and 5 preference dimensions. A reinforcement‑learning agent, B2T‑Agent, leveraging behavior trajectories, external retrieval tools, and internal memory, outperforms strong baselines such as GPT‑4.1 on this challenging dataset.
The paper introduces Instruct-to-Act, a system that decouples high‑level planning from low‑latency control by combining a vision‑language model (VLM) planner with a world‑model controller. The VLM generates sparse, high‑level text instructions, while the controller executes them autonomously at high frequency. Experiments across seven embodied environments, including multi‑agent settings, show that this approach outperforms both controller‑only and direct VLM action‑generation methods, maintains fast control, and allows swapping in different pretrained VLM planners without fine‑tuning.
The paper introduces AI Control Scientist (AICS), a large language model–driven agent that automatically generates optimized controllers from language design requirements. AICS consists of a Task Modeling Agent that translates user needs into engineering constraints, a Controller Design Agent that produces candidate controller structures and code, and a Parameter Tuning Agent that refines parameters to meet closed‑loop performance criteria. Experiments show that AICS outperforms existing automated baselines in design success rate and optimization efficiency, demonstrating its potential to shift control system design from human‑driven to agent‑driven approaches.