Llm agents benchmarked for trusted post-training model delivery
Trains but Doesn't Learn: A Post-Training Delivery Benchmark for LLM Agents as Forward-Deployed Engineers
Machine LearningArtificial IntelligenceComputation and Language
Summary
Training large language models (LLMs) after deployment is now offered as a service, where an engineer tunes and delivers a model based on user data and goals. The authors created a new benchmark that tests whether an LLM acting as this engineer can be trusted to truly improve the model, not just reduce training loss without learning. They ran four advanced AI agents and compared them to human engineers using a detailed scoring process. The study found that some runs look successful but actually produce no improvement, and they developed techniques to detect and block such failures.
What this means in practice
- •For machine learning operations teams: Validate and monitor AI agents that autonomously fine-tune and deploy models to ensure reliable improvements before release.
- •For ai deployment platform providers: Integrate trustworthiness tests and corruption detectors into post-training services to prevent costly failures when delivering client models.$Commercial implications: Enables platforms to offer certified safe model fine-tuning services using automatic detection to gain customer trust.
Authors
Weihang Ding, Junfei Zhan
Abstract
Post-training is becoming a service (PTaaS): a customer hands an operator data and a goal, and a forward-deployed engineer (FDE) returns a fine-tuned, evaluated, and deployed model under a budget, a human-approval gate, and reproducibility requirements. Seating an LLM agent in the FDE seat raises a question existing benchmarks cannot answer: not whether an agent can raise a metric, but whether it can be trusted to deliver. We answer it on a governed delivery plane, where an agent drives ten stages and an oracle scores each stage from platform-recorded facts. The central silent failure is the run that trains but does not learn (TBDL): loss falls, every signal stays green, and the delivered model is no better than the base. An operator-run acceptance gate catches every such run before payment, and a detector calibrated on known-corrupted runs flags severe corruption mid-run. We ran four frontier agents (Claude Opus 5, GPT-5.6-luna, Gemini 3.7 Flash, DeepSeek V4-Pro) end to end on metered L40S, A100, and H200 GPUs across 8B to 70B open bases, certifying every scenario before scoring. We also ran a human FDE arm under the same oracle and compare every agent against it.