Search arXiv⌕ Search
Home/Archive/arXiv

arXiv · 2610.02405

When Terminal-Agent Training Stalls: Demystifying Data Generation and Verification Challenge

Xi Qin·Isabel Kurth·Xin Cui·Elin Park·Alexander Schaefer·Yaad Oren
View original source ↗Read source document ↗Export metadata ↓

Abstract

Using a frontier model like Claude Opus as a meta-agent to generate terminal tasks and verifiers for RL training is increasingly common. Yet a runnable Docker image and executable test suite do not guarantee a faithful end-to-end pipeline for terminal agent training. We present a meta-agent pipeline motivated by this gap, diagnosing three classes of failure: benchmark invalidity, harness brittleness, and reward misalignment. Prompt redesign and context extension raise baseline solvability 5.6 times, but a 9B model saturates at 81.3% mean pass@2 within 20 steps on Claude Opus-generated tasks. Adding hard tasks reduces mean pass@2 to 20.6% without changing the training configuration, a strong evidence that the solvability band is model-specific. These findings demonstrate that meta-agent reliability requires solvability-band calibration, verifier audits, and infrastructure error accounting as first-class evaluation criteria, not post-hoc diagnost.

Explore related subjects

cs.AI

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Xi Qin, Isabel Kurth, Xin Cui, Elin Park, Alexander Schaefer, Yaad Oren. 2026-10-01. When Terminal-Agent Training Stalls: Demystifying Data Generation and Verification Challenge. https://arxiv.org/abs/2610.02405

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

AstroAgentBench: Evaluating Agentic Planning on Space Mission Planning Tasks

Recent LLM-for-Space systems address mission planning, scheduling, operations support, simulator control, and autonomy, but their evaluations use different task contracts, control settings, simulators, and success criteria. We introduce AstroAgentBench, a seven-family benchmark for executable space mission planning in the domains of scheduling, observation planning, constellation design, and relay support. For each case, an agent submits a planning artifact that is checked by an external verifier for schema, timing, geometry, resources, and mission value. Results report validity and normalized scores, with comparisons to task-specific solver references. Across five LLM agent systems and 35 held-out cases, the strongest systems approach or exceed solver-reference scores on several families, while weaker systems often fail to produce high-value valid plans and even strong systems lose quality on geometric, product-level, or design-heavy tasks. Trace analyses separate two failure points: task-contract misformulation and weak solution construction. Successful runs instead calibrate agent-written implementations against verifier feedback and adapt search to case-specific structure. Ablations show that procedure injection and memory accumulation help selectively, when they supply the missing formulation, calibration, or search support.

cs.AI↗
arXiv2026

LPS-Bench: Benchmarking Safety Awareness of Computer-Use Agents in Long-Horizon Planning under Benign and Adversarial Scenarios

Computer-use agents (CUAs) execute multi-stage tasks through tools, where an early unsafe decision can propagate to consequential actions. Evaluating only final outcomes can miss such decisions, while constructing executable environments for new tasks can make benchmark expansion costly. We present LPS-Bench, a benchmark of long-horizon planning safety in MCP-style tool workflows under benign requests and adversarial steering. A template-guided multi-agent pipeline generates user instructions, simulated toolkits, and case-specific safety criteria, followed by human review. This design supports scalable case expansion without building a separate application environment for every test case. LPS-Bench comprises 570 cases derived from 65 scenarios across 7 task domains and 9 planning-risk types, with representative cases additionally adapted to reusable skills. An LLM-based evaluator applies case-specific criteria to complete interaction records, examining tool choices, arguments, and responses to environmental feedback throughout execution. Evaluations of 13 LLM agents reveal persistent failures in both benign and adversarial settings. Prompt-based interventions yield model-dependent gains, but substantial safety failures remain.

cs.AI↗
arXiv2026

On the Tip of the Tongue: Why LLMs Hallucinate Answers They Can Decode

A language model can give the wrong answer even when the correct answer is decodable from its intermediate states. To study this gap between decodability and selection, we distinguish \textit{read} from \textit{write} at the first answer token. Read asks whether the gold token can be decoded from intermediate residual states under same-relation decoy controls. Write asks whether the final readout ranks that token first among content tokens. Under three different readers, with a randomized-label control, a substantial fraction of failures remain readable while another content token is selected. We explain this through the selection margin at the final readout, the difference between the answer logit and the logit of its strongest alternative, which is answer support minus alternative support, and can also be split into a context-averaged baseline linked to token frequency and an item-specific term. Setting the answer support to the level typical of successful generations is sufficient to recover first-token selection for the majority of failures in most of the models we study; the original alternative remains ahead in most remaining failures under this edit, and this outcome follows directly from the readout geometry. Removing the frequency direction alone shifts selection but rarely recovers the answer. Prompt variants of the same fact that succeed supply support that transfers to failing variants through the residual stream and through late MLP outputs, with less consistent effects through late attention. First-token recovery leaves most full answers wrong, which limits the recovery achieved by these edits and separates three things that are easily conflated, decodability, recoverability, and generation.

cs.AI↗