Agents Learn Their Runtime: Interpreter Persistence as Training-Time Semantics

• URL: http://arxiv.org/abs/2603.01209v2
• Date: Thu, 05 Mar 2026 05:07:31 GMT
• Title: Agents Learn Their Runtime: Interpreter Persistence as Training-Time Semantics
• Abstract summary: Tool-augmented LLMs are increasingly deployed as agents that interleave natural-language reasoning with executable Python actions.<n>We ask whether interpreter persistence is merely a runtime scaffold, or a property of the training data that shapes how agents learn to use the interpreter.<n>We introduce Opaque Knapsack, a procedurally generated family of partially observable optimization tasks designed to prevent one-shot solutions.
• Score: 4.774525456207306
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Tool-augmented LLMs are increasingly deployed as agents that interleave natural-language reasoning with executable Python actions, as in CodeAct-style frameworks. In deployment, these agents rely on runtime state that persists across steps. By contrast, the traces used to post-train these models rarely encode how interpreter state is managed. We ask whether interpreter persistence is merely a runtime scaffold, or a property of the training data that shapes how agents learn to use the interpreter. We isolate state persistence as a training-time variable. We introduce Opaque Knapsack, a procedurally generated family of partially observable optimization tasks designed to prevent one-shot solutions. Item attributes and constraints are hidden behind budgeted tool calls, forcing multi-turn control flow and iterative state revision. Holding task instances, prompts, tools, model, and supervision fixed, we generate matched trajectories differing only in whether interpreter state persists across steps or resets after each action. We then fine-tune identical base models (Qwen3-8B) on each trace variant and evaluate all four train-runtime combinations. Our 2x2 cross-evaluation shows that interpreter persistence shapes how agents reach solutions, not whether they do: solution quality is statistically indistinguishable across conditions, but token cost and stability differ substantially. A persistent-trained model in a stateless runtime triggers missing-variable errors in roughly 80% of episodes; a stateless-trained model in a persistent runtime redundantly re-derives retained state, using roughly 3.5x more tokens. Interpreter persistence should be treated as a first-class semantic of agent traces. Aligning fine-tuning data with deployment runtimes improves efficiency and reduces brittle train-runtime mismatches.

Related papers

• AgentPatch: Coarse-to-Fine Weak-Task Repair for Merging Agentic Multimodal Large Language Models [22.004098225093177]
  AgentPatch is a training-free coarse-to-fine repair framework for agentic large language models.<n>It selects a stable merged backbone, restores diluted weak-task-specific signals through Weak-Task Unique Residual Recovery, and applies an Agent-Guided Behavior-Critical Patch.<n>AgentPatch produces a single static checkpoint without routing or ensembles.
  arXiv  Detail & Related papers  (2026-08-07T01:52:45Z)
• Verified Tool Calls Improve LLM Agent Reliability Under Non-Atomic Failures [0.0]
  Real-world systems exhibit non-atomic behaviors such as timeouts after dispatch, delayed visibility, and partial state updates.<n>These mismatches lead to reliability issues including duplicate actions, task success, and unnecessary tool executions.<n>A lightweight, verification-aware tool wrapper is introduced that augments tool calls with postcondition verification, verify-before-retry logic, and idempotency keys.
  arXiv  Detail & Related papers  (2026-07-31T16:16:14Z)
• PhyAgentOS: A Self-Evolving Operating System for Embodied Agents with Decoupled Cognitive Planning and Physical Execution [49.776611937968]
  We present PhyAgentOS, a foundation delivering scheduling, verification, memory, benchmarking, and safety as system-level services.<n>Its Session-Centered treats a session, not an action, as the minimum unit of scheduling, compatibility, supervised execution, evidence collection, and acceptance.<n>SessionVerifier distinguishes execution termination from semantic task completion via evidence-grounded verdicts of success, failure, or replan.<n> Benchmarking reuses the deployment session and verification path, so results trace to real execution.
  arXiv  Detail & Related papers  (2026-07-18T04:46:53Z)
• AutoTrainess: Teaching Language Models to Improve Language Models Autonomously [18.706079441467992]
  Training language models (LMs) remains a highly human-intensive process.<n>We present AutoTrainess, a LM agent that exposes these operations as a repository of agent-computer interfaces for planning, data preparation, training, evaluation, and logging.<n>On PostBench, AutoTrainess consistently outperforms CLI-only baselines, achieving 26.94 average score with GPT-5.4 (Codex) versus 23.21 for CLI-only.
  arXiv  Detail & Related papers  (2026-06-30T12:09:51Z)
• ClawForge: Generating Executable Interactive Benchmarks for Command-Line Agents [59.626170560327274]
  We present textbfClawForge, a generator-backed benchmark framework for executable command-line categories under state conflict.<n>We instantiate this framework as the ClawForge-Bench (17 scenarios, 6 ability categories)
  arXiv  Detail & Related papers  (2026-05-13T21:34:08Z)
• CodeTracer: Towards Traceable Agent States [40.51936201889185]
  We present CodeTracer, a tracing architecture that parses heterogeneous run artifacts through evolving extractors.<n>CodeTracer reconstructs the full state transition history as a hierarchical trace tree with persistent memory.<n>It performs failure onset localization to pinpoint the failure origin and its downstream chain.
  arXiv  Detail & Related papers  (2026-04-13T15:52:03Z)
• Training LLMs for Multi-Step Tool Orchestration with Constrained Data Synthesis and Graduated Rewards [76.49428173793386]
  LLMs must invoke multiple dependent APIs in the correct order while propagating intermediate outputs.<n>Training models to handle such faces two obstacles: existing environments focus on simple per-turn function calls with simulated data, and binary rewards.<n>We present a framework addressing both challenges. First, we construct a reinforcement learning environment backed by a large-scale cache of real API responses, enabling a data synthesis pipeline that samples valid multi-step orchestration traces.<n>Second, we propose a graduated reward design that decomposes correctness into atomic validity and orchestration.
  arXiv  Detail & Related papers  (2026-03-25T18:31:39Z)
• Efficient Continual Learning in Language Models via Thalamically Routed Cortical Columns [0.16921396880325779]
  We introduce TRC$2$ (Thalamically Routed Cortical Columns), a decoder-only backbone that addresses continual learning at the architectural level.<n>The resulting block is sparse and chunk-parallel, enabling efficient training and inference while preserving clean ablations of each subsystem.
  arXiv  Detail & Related papers  (2026-02-25T23:38:16Z)
• Transporting Task Vectors across Different Architectures without Training [18.853461748616333]
  We introduce Theseus, a training-free method for transporting task-specific updates across heterogeneous models.<n>Rather than matching parameters directly, we characterize a task update by the functional effect it induces on intermediate representations.
  arXiv  Detail & Related papers  (2026-02-13T14:16:34Z)
• Decomposing Reasoning Efficiency in Large Language Models [2.4149105714758545]
  We decompose token efficiency into interpretable factors: completion under a fixed token budget, conditional correctness given completion, and verbosity.<n>When reasoning traces are available, we add deterministic trace-quality measures to separate looping from verbose-but-engaged reasoning.<n>Our decomposition reveals distinct bottleneck profiles that suggest different efficiency interventions.
  arXiv  Detail & Related papers  (2026-02-10T14:09:18Z)
• ET-Agent: Incentivizing Effective Tool-Integrated Reasoning Agent via Behavior Calibration [68.89572566071575]
  ETAgent is a training framework for calibrating agent's tool-use behavior.<n>It is designed to progressively calibrate erroneous behavioral patterns to optimal behaviors.
  arXiv  Detail & Related papers  (2026-01-11T11:05:26Z)
• Tree Training: Accelerating Agentic LLMs Training via Shared Prefix Reuse [21.642997639835396]
  We propose Tree Training, a paradigm that computes each shared prefix only once and reuses its intermediate results across related branches during both forward and backward passes.<n>Experiments on multiple open-source models demonstrate up to 3.9x reduction in total training time, enabling more efficient agentic LLM SFT and RL training.
  arXiv  Detail & Related papers  (2025-11-01T05:56:49Z)
• Training-Free Time Series Classification via In-Context Reasoning with LLM Agents [29.14242392533328]
  Time series classification (TSC) spans diverse application scenarios, yet labeled data are often scarce.<n>We propose FETA, a multi-agent framework for training-free TSC via exemplar-based in-context reasoning.
  arXiv  Detail & Related papers  (2025-10-07T14:07:43Z)
• Impatient Users Confuse AI Agents: High-fidelity Simulations of Human Traits for Testing Agents [58.00130492861884]
  TraitBasis is a lightweight, model-agnostic method for systematically stress testing AI agents.<n>TraitBasis learns directions in activation space corresponding to steerable user traits.<n>We observe on average a 2%-30% performance degradation on $tau$-Trait across frontier models.
  arXiv  Detail & Related papers  (2025-10-06T05:03:57Z)
• Reinforcement Learning for Machine Learning Engineering Agents [52.03168614623642]
  We show that agents backed by weaker models that improve via reinforcement learning can outperform agents backed by much larger, but static models.<n>We propose duration- aware gradient updates in a distributed asynchronous RL framework to amplify high-cost but high-reward actions.<n>We also propose environment instrumentation to offer partial credit, distinguishing almost-correct programs from those that fail early.
  arXiv  Detail & Related papers  (2025-09-01T18:04:10Z)
• Agentic-R1: Distilled Dual-Strategy Reasoning [58.73951532294446]
  Current long chain-of-thought (long-CoT) models excel at mathematical reasoning but rely on slow and error-prone natural language traces.<n>We introduce a fine-tuning framework, DualDistill, that distills complementary reasoning strategies from multiple teachers into a unified student model.<n>Our method improves accuracy across a range of tasks, including both computation-intensive and standard benchmarks.
  arXiv  Detail & Related papers  (2025-07-08T06:35:16Z)
• Runaway is Ashamed, But Helpful: On the Early-Exit Behavior of Large Language Model-based Agents in Embodied Environments [54.67512489842682]
  Large language models (LLMs) have demonstrated strong planning and decision-making capabilities in complex embodied environments.<n>We take a first step toward exploring the early-exit behavior for LLM-based agents.
  arXiv  Detail & Related papers  (2025-05-23T08:23:36Z)
• Sports-Traj: A Unified Trajectory Generation Model for Multi-Agent Movement in Sports [53.637837706712794]
  We propose a Unified Trajectory Generation model, UniTraj, that processes arbitrary trajectories as masked inputs.<n>Specifically, we introduce a Ghost Spatial Masking (GSM) module, embedded within a Transformer encoder, for spatial feature extraction.<n>We benchmark three practical sports datasets, Basketball-U, Football-U, and Soccer-U, for evaluation.
  arXiv  Detail & Related papers  (2024-05-27T22:15:23Z)
• Sequential Reptile: Inter-Task Gradient Alignment for Multilingual Learning [61.29879000628815]
  We show that it is crucial for tasks to align gradients between them in order to maximize knowledge transfer. We propose a simple yet effective method that can efficiently align gradients between tasks. We extensively validate our method on various multi-task learning and zero-shot cross-lingual transfer tasks.
  arXiv  Detail & Related papers  (2021-10-06T09:10:10Z)
• Stabilizing Label Assignment for Speech Separation by Self-supervised Pre-training [58.30339239234169]
  We propose to perform self-supervised pre-training to stabilize the label assignment in training the speech separation model. Experiments over several types of self-supervised approaches, several typical speech separation models and two different datasets showed that very good improvements are achievable if a proper self-supervised approach is chosen.
  arXiv  Detail & Related papers  (2020-10-29T06:07:01Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.