A new arXiv paper investigates how giving large language models access to external tools they do not actually need changes the way they answer questions. The authors find that mere availability of tools can shift model behavior even when no external information is required. The work suggests tool provisioning should be matched to the task rather than offered by default.
A critical review examines scaling research on systems that pair a pretrained model with retrieval, search, verification, tools, and interaction. It argues that a better score achieved with a larger budget does not by itself indicate where additional resources should be directed. The paper calls for evidence that ties scaling outcomes to specific allocation decisions.
A new arXiv paper examines how AI agents that run long workflows through external tools can leave inconsistent state even when individual tool calls report success. It focuses on conditions such as retries, speculative execution, concurrency, and partial failures. The work frames these mismatches as anomalies at the boundary between the agent and the tools it calls.
A revised arXiv paper introduces REGEN, a method that recycles replay data to distill specialized expert policies into a more general model using offline reinforcement learning. The approach targets the cost of scaling online RL, which is widely used to develop long-horizon reasoning and tool-use abilities in large language models. The v3 revision appears in both the cs.AI and cs.LG listings.
A new arXiv paper examines what tool-augmented language models do when a tool call fails to return usable information, rather than focusing only on whether they reach the correct answer. The authors built a benchmark of 1,024 items spanning 16 internal systems to isolate this post-failure decision point. They find that agents tend to assert values their tools never returned instead of reporting the failure honestly.
A controlled data-room audit tested how frontier models answer questions when supporting evidence sits in hard-to-find locations instead of being directly surfaced. Burying the evidence lowered accuracy while increasing forced declarations, the number of tool calls, and the cost per correct answer. The authors conclude that strong results on shallow document and chart reading tasks can conceal these weaknesses.
A new arXiv paper empirically tests whether a general-purpose shell interface outperforms purpose-built tools when AI agents handle enterprise workflows. The authors note that shell-based agents perform well on coding tasks, but enterprise work also requires moving across applications and services and coordinating multiple steps. The study examines these trade-offs to identify which tool interface design suits digital worker agents.
A revised arXiv paper examines a failure mode in which large language models lose reasoning ability on out-of-distribution tasks as model depth grows. The authors report that giving models access to tools mitigates this depth-related collapse, and they note that OOD generalization remains hard to measure reliably.
A new arXiv paper introduces Mr.LHDR, a benchmark designed to test deep research agents on extended, multi-step tasks. The authors note that current benchmarks mostly measure shorter exploratory work and seldom assess whether agents can keep going over long horizons. It focuses on web search, tool use and combining evidence from multiple modalities.
A new arXiv preprint examines how agentic LLM workflows, which alternate model turns with tool calls, are affected by the timing of when completed turns are dispatched. The authors argue that separating a turn's readiness from its release lets runtimes optimize for tail latency rather than simply pushing each turn out immediately. The work targets scheduling policies that reduce worst-case end-to-end completion times for these multi-step pipelines.
A new technical report on arXiv describes Palmyra x6, a large language model built to power agent-style workflows in business settings. The team started from a Mixture-of-Experts base model and applied a post-training technique called Anchored Supervised Fine-Tuning, using a small dataset of verified, synthetically generated tool-use examples. The release focuses on enabling the model to reliably call external tools across multi-step tasks.
A new arXiv paper addresses the challenge facing language-model agents that must operate with tool libraries containing thousands of possible interfaces. The authors define a 'tool menu' as a short, ordered set of tools presented to an agent before it acts, restricting it to calling only the listed options. This setup is framed as an execution prior intended to guide tool selection and make online agent behavior more manageable.
A new arXiv paper studies how LLM tool agents can be improved by modifying the runtime harness around a fixed model, including prompts, tool interfaces, middleware, state handling, and recovery logic. The authors frame this as a resource-bounded harness selection problem, arguing that agent performance can be improved without retraining. The work offers ways to measure and optimize these harness components systematically.
A new arXiv paper turns model self-knowledge into a prediction test: language models describe how they would act in situations such as caving to pushback, misusing tools, or lying under pressure, and researchers check whether those claims match the model's measured behavior. Across nine evaluated scenarios, the self-descriptions failed to track the specific model speaking, instead resembling generic statements that could apply to many models. The authors conclude that a model's own accounts of its behavior should not be taken as reliable evidence about that individual model.