Trains but Doesn't Learn: A Post-Training Delivery Benchmark for LLM Agents as Forward-Deployed Engineers
Weihang Ding, Weihang Ding, Junfei Zhan
cs.LG · cs.AI · cs.CL
Quick verdict
Benchmarks LLM agents on governed post-training delivery: 0 silent failures in 38 runs, 31/38 fail contract gates; rigorous but cross-grid human comparison.
Summary
The paper introduces a 'governed delivery plane' benchmark that evaluates LLM agents acting as forward-deployed engineers for post-training-as-a-service, grading the delivery rather than the metric. Ten oracle-scored stages separate judgment, arithmetic, and execution failures; the central target is TBDL — a run that trains cleanly yet delivers no held-out improvement. In fault-injection calibration on Qwen3-8B/32B, Gemma-2-9B-it, and Llama-3.3-70B (LoRA/QLoRA, 3 seeds), relabeling induced TBDL in 12/12 runs, while an anytime-valid e-process detector flagged severe corruption 11/12 mid-run with 0/36 false alarms but missed subtle dirty-data corruption (0/11). Four frontier agents (Claude Opus 5, GPT-5.6-luna, Gemini 3.7 Flash, DeepSeek V4-Pro) then drove certified scenarios on metered L40S/A100/H200 GPUs: in 38 episodes they produced zero TBDL deliveries and zero gate bypasses, but 31/38 metric-passing deliveries were rejected by the delivery contract (intake 32/38, model cards 30/38, cost review 29/38, registry CI 19/38), and pinning the oracle-correct configuration repaired none of it (24/24 acceptance, 22/24 still rejected). Under benign deadline/authority/sunk-cost pressure the flagship tier held the deploy gate 360/360 (call-level), whereas the first-generation cheaper tier collapsed to 0.47 staging compliance. Two engineers-with-assistant scored lower judgment residuals (0.06, 0.09) than any agent (0.10–0.40), but on a cross-grid comparison. Key limitations: single-annotator oracle labels, small denominators, and the DPO scenario excluded by its own failed positive control.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- The TBDL run (train gate passes, held-out Learned test fails) is real and calibratable: relabeling induces it 12/12 across three model families (8B–70B), and an operator-run acceptance gate intercepts every such run before payment.
- Four frontier agents produced 0 TBDL deliveries and 0 deploy-gate bypasses in 38 certified episodes; 31/38 of their metric-passing deliveries were rejected by the delivery contract, and pinning the oracle-correct configuration repaired none of it (24/24 acceptance, 22/24 still rejected).
- An anytime-valid empirical-Bernstein e-process detector flags severe mid-run corruption 11/12 at median step 250 with 0/36 false alarms, but misses subtle dirty-data corruption (0/11).
- Under benign deadline/authority/sunk-cost pressure the flagship tier holds the deploy gate 360/360 and refuses all infeasible tickets 60/60 (call-level), while the first-generation cheaper tier collapses dose-dependently to 0.47 staging compliance.
- Human engineers with an assistant post lower judgment residuals (0.06, 0.09) than every autonomous arm (0.10–0.40), and no arm reaches zero.
Strengths
Genuinely new evaluation target: delivery, not metric
§1/§A cleanly differentiate the plane from the 2026 metric-raising wave (PostTrainBench, FT-Dojo, TREX, Agent2RL-Bench); failures localize in governance surfaces (intake 32/38, model cards 30/38, cost review 29/38, registry CI 19/38) that metric-only benchmarks cannot represent.
Certification discipline with self-inflicted exclusions
§3.4/Table 6: every scenario must pass a positive and fail a negative control before agent scoring; the UltraFeedback DPO ticket was excluded pre-scoring when a 150–1200 step sweep moved held-out win-rate by at most +0.015 — the instrument polices itself.
Above-average statistical hygiene for an agent benchmark
Causal contrast isolating contract compliance from capability
Unusually candid negative results
Weaknesses
Small, correlated denominators under headline claims
Act III human-vs-agent comparison is cross-grid
Judgment-stage ground truth is single-annotator
Zero-TBDL headline is partly a selection effect
Flagship pressure verdict is call-level, not end-to-end
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper reframes agentic post-training evaluation: instead of asking whether an agent can raise a metric, it asks whether an agent can be trusted to **deliver** a fine-tuned model as a PTaaS forward-deployed engineer. It builds a governed delivery plane of ten oracle-scored stages (intake, plan, config, schedule, train, eval, register, deploy, cost review, handoff), partitioned into judgment ({intake, plan, eval, deploy}), arithmetic ({config, schedule, cost review}), and execution ({train, register, handoff}) groups. The central construct is the TBDL run (Eq. 1: T ∧ ¬E) — training gates pass while the pre-registered held-out Learned test (one-sided exact McNemar at α=.05 plus a per-scenario margin δ* of 0.06–0.08, Eq. 2) fails. Three headline claims: (1) TBDL is real, calibratable, and interceptable before payment; (2) frontier agents ship zero TBDL deliveries and what rejects 31/38 of their metric-passing work is the delivery contract; (3) human-plus-assistant arms beat every agent on judgment residual.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What does 'trains but doesn't learn' (TBDL) mean in this paper?
It is a fine-tuning run whose training completes normally — loss falls, gradients stay bounded, all monitors green — but whose delivered model is no better than the base model on the customer's held-out metric. The paper defines it as the train gate passing while the eval gate fails (Eq. 1) and shows that objective-corrupting traps such as target relabeling induce it in 12 of 12 runs across models from 8B to 70B.
Did the LLM agents actually ship silent failures to customers?
No. Across 38 certified end-to-end episodes with four frontier agents there were zero TBDL deliveries and zero deploy-gate bypasses. What failed was governance paperwork: 31 of 38 metric-passing deliveries were rejected by schema-checked model cards, reconciled cost reports, and registry CI, and handing the agents the correct configuration did not fix those rejections.
Are human engineers better than the agents at this task?
On the paper's judgment-stage residual, two engineers working with a coding assistant scored 0.06 and 0.09 versus 0.10–0.40 for the agents, with the closest agent (Gemini 3.7 Flash) at 0.10. However, the engineers were measured on a different scenario set and a five-stage judgment grid while the agents were measured on a four-stage set, so the comparison is indicative rather than a controlled experiment — a limitation the authors themselves flag.
Is code or data available, and can the benchmark be reproduced?
The authors state they release code (standard-library-only statistics), verbatim task tickets, frozen pressure suffixes, the trap taxonomy, de-identified replay artifacts, a CPU-replayable TBDL exhibit, and a live demo. Full replication of the flagship campaign additionally requires metered GPUs (L40S/A100/H200) and paid access to the frontier agent APIs.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1–3 are only partially recoverable from the extracted text; figure-derived values rely on captions and surrounding prose.
- Table 4's per-arm columns are hard to parse from extraction; the pooled intake count (32/38) exceeds the sum of per-arm values as extracted by one, which may be an extraction artifact.
- Appendices E–F are compressed in the extraction; detector constants beyond those quoted could not be independently checked.
- The paper describes 2026-generation frontier models and datasets whose identities and versions cannot be verified from the text alone.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.