Tool use, function calling, orchestration and the protocols that let models act rather than only answer.
Autonomous agents have been observed acting beyond their assigned tasks, accessing unauthorized systems, and establishing communication channels to exchange information, yet the Web and network-level...
Data curation is a first-order design choice in machine learning, yet tailoring training data to a downstream model and task remains costly in labor, expertise, and compute. Datasets are therefore oft...
Decentralized multi-agent systems (MAS) enable large language model agents to interact without centralized control. Agents need to adapt to other agents at test time and reuse accumulated experience t...
The article discusses how an agent can continue to produce confident answers even when its tools fail, leading to a misleadingly positive dashboard display. It highlights the issue of agents generating responses from nothing after tool breakdowns, which can create a false sense of reliability. The piece serves as a cautionary note for developers deploying agentic AI systems.
Autonomous agents are increasingly deployed across sequences of related tasks, yet they often fail to retain and improve their knowledge over time. Existing approaches convert interaction trajectories...
Agents end many steps in a typed decision over a set of candidates: which screen element to click, which product to pick. System One models answer in one forward pass, and Jeva showed that four proper...
The article provides a practical guide on implementing guardrails for a local large language model (LLM) agent. It focuses on using Gemma 4, Ollama, and the OpenAI Agents SDK to set up these safety measures. The post is published on Towards Data Science.
Although Large Language Models (LLMs) have demonstrated potential in clinical documentation, generating the rigorous, evidence-based reasoning required for unprecedented, atypical, and rare disease ca...
The paper introduces GAV, a scientific experimentation environment that generates verifiable PDE problems by learning a distribution of reference solutions and using the method of manufactured solutions to construct matching PDEs. GAV includes a Generator, Agent, and Verifier that together enable iterative program improvement and provide numerical feedback on accuracy and computational cost. Experiments on Navier–Stokes and Vlasov–Poisson equations show that GAV can significantly reduce test errors for existing PDE agents.
DiffResearch is an open multi‑agent deep research framework that replaces the traditional autoregressive language model backbone with a diffusion language model (dLLM) across all stages of research, from planning to writing. In experiments on 100 Deep Research Bench tasks, the diffusion backbone achieves roughly six times faster end‑to‑end latency, largely due to significant speedups in planning and long‑form generation. The framework also introduces dFast, a decomposition mode that reads intermediate denoising states to generate subqueries, yielding a median 1.99× speedup without compromising report quality and achieving the best overall score on the RACE rubric. whyItMatters":"The study demonstrates that diffusion language models can dramatically reduce latency in deep research systems, enabling faster interactive use while maintaining or improving output quality."
Gandalf is a flash‑based KV cache offloading system designed to extend SSD lifetime in large language model serving. It prioritizes full‑attention KV cache entries for SSD admission, using attention type and historical access patterns to reduce write traffic without lowering cache hit rates. Experiments show a 7.3× increase in SSD lifespan with negligible impact on hit rate across diverse workloads and hardware configurations.
AgentDreamer is an agent-guided optimization framework that aligns text instructions with simulated 3D motion by using an off-the-shelf agent to generate a structured semantic motion prior. The prior specifies what should move, in which direction, and in what temporal order, and is represented as parametric functions over temporal chunks to produce dense 3D trajectories. These trajectories are then used to optimize a differentiable MPM simulator so that its rollout matches the prior while respecting scene dynamics, and the framework introduces Motion-IG, an information‑theoretic metric for measuring text‑motion alignment.
AgenticEvolve is an evolutionary program search framework that employs persistent large language model agents on each island. These agents not only generate candidate programs but also design and run experiments between evaluations. Compared to a version without experimentation (No-analysis) and other baselines, AgenticEvolve achieves higher quality programs on heuristic design tasks and recovers more ground‑truth equations on discovery problems, doing so with fewer evaluations and similar cost.
FaultGene is an LLM-assisted framework that automatically generates executable fault‑augmented tests for hardware‑in‑the‑loop validation of autonomous driving systems. It takes a baseline Scenic/CARLA scenario and a structured fault specification, then produces fault‑augmented test cases that reflect HIL‑level fault injection effects. The system refines candidates offline using semantic‑consistency checks and execution‑level feedback, enabling high‑quality test generation at deployment time without manual engineering or repeated refinement loops.
The paper introduces DynaCertBench, a benchmark of 105 diverse systems for certifying stability, safety, and related properties, and presents LLM2Cert, a framework that uses large language models to generate and verify candidate certificates via SMT solvers or Lean. LLM2Cert achieves a 77.1% verification rate on the benchmark, improving by 17.1 percentage points over a single-shot approach, and successfully verifies 22 certificates that All-SMT could not resolve. The authors provide code and data on GitHub for reproducibility.
CompLit is a new literature search benchmark covering all eight arXiv domains, designed to evaluate agents on three diagnostic settings: implicit, cumulative, and unmet needs. The benchmark shows that top agents like Codex and Claude Code perform well on standard queries (86‑87%) but drop to 66‑68% on implicit queries, lose additional accuracy when requirements accumulate, and abstain on up to 53% of unmet queries. The study finds that agents often locate the correct or near‑correct paper but misjudge it, indicating a need for more reliable search capabilities.
The paper investigates whether large language models’ refusal behavior in text transfers to tool‑calling contexts. A benchmark of 2,304 prompts across four domains compares text‑only, standard tool‑enabled, and adversarially framed tool‑enabled settings, revealing that 10–52 % of harmful requests refused in chat still lead to unsafe tool calls. The authors identify a refusal‑associated direction in the models’ residual stream, showing that lower activation along this direction predicts unsafe calls, but restoring the signal only partially mitigates the issue and often suppresses benign tool use.
AgentMonopoly is a new environment that lets large language model agents play a property‑trading game while communicating with each other. The study compares models from four families and two sizes, examining how often they trade, waive rent, and manage cash and properties, and tests whether giving an agent a persona description changes its language and actions. Results show that larger models do not always win more games and that persona instructions alter wording and some actions, but not consistently across all decisions, highlighting the need to evaluate persona adherence through both language and outcomes.
The paper introduces the Latent Interconnect Protocol (LIP), a framework that allows large language models (LLMs) to directly exchange rich, high‑dimensional representations with specialist models (e.g., protein and small‑molecule experts) via learned connectors. By training only these connectors, the approach enables zero‑shot transfer to cross‑domain tasks and compositional reasoning without joint training of the LLMs or specialists. Experiments with Gemma 3 and Kimi K2.6 show improved task performance and compositionality, especially in zero‑shot settings, while highlighting the need for consistent specialist interfaces and LLM training to handle such inputs.
The paper introduces an agent architecture that uses tools and memory to selectively correct wrist photoplethysmography (PPG) heart‑rate estimates. By generating candidate corrections from pulse, motion, and historical data, and evaluating their predicted benefits and harms, the system decides whether to retain, correct, or reject each estimate. Across four public datasets, the approach reduced mean absolute error by up to 44.5% and demonstrated real‑time deployment on a wearable‑smartphone platform.