Long-Horizon AI Is Not the Same as Reliable AI — Stanford CS329A Part 8

작성자:

Series: Self-Improving AI Agents — Part 8 of 9

Previous: Part 7 — Search and Deep Research Agents

Lecture: Stanford CS329A | Agentic Evaluations and Long-Horizon Tasks

Source video: https://www.youtube.com/watch?v=8JAqLnTaZu4

한국어판

1. Why Part 8 matters after Part 5

Part 5 already established a basic distinction: capability is not the same as reliability. Part 8 revisits evaluation from a deployment perspective.

The question is no longer only whether a model can do a task. It becomes: how long a task can the agent handle, how reliably can it handle it, and is the resulting work actually valuable?

The lecture organizes this around three evaluation families:

  1. METR: task-completion time horizon and reliability
  2. GDPval: professional deliverable quality and economic usefulness
  3. DeepScholar-Bench: research synthesis, retrieval, and citation verifiability

These are not interchangeable scores.

Longer autonomy, higher reliability, better economic output, and stronger research synthesis are different variables.

2. METR changes the unit of evaluation from questions to tasks

Traditional benchmarks often ask whether a model answered a question correctly. METR asks what length of task an AI agent can complete with a specified probability of success.

The key metric is the time horizon.

A 60-minute time horizon does not mean the AI literally works autonomously for 60 minutes. It means that, on METR’s task distribution, the model is predicted to succeed at a specified probability on tasks that take human experts about 60 minutes to complete.

The duration is therefore a human-equivalent measure of task difficulty.

3. The historical doubling result is real, but it is a trend, not a law

METR’s original long-task study found that the 50% task-completion time horizon of frontier agents had doubled roughly every seven months over the preceding six years.

Claude 3.7 Sonnet was measured at roughly 55–59 minutes depending on evaluation setup.

That result is important because it captures something short benchmarks miss: frontier agents have been getting better at completing increasingly long, multi-step software and research tasks.

But it should not be treated as a physical law. METR has since changed task suites and tooling, and its current methodology warns that measurements above 16 hours are unreliable with the present task suite.

Agent time horizons have shown a strong historical exponential trend, but the exact slope and future extrapolation are method-dependent.

4. The most important METR result is not 50% reliability

A 50% success probability is useful for tracking capability growth. It is much less useful as a deployment standard.

If an autonomous agent succeeds only half the time on a one-hour task, the user still needs supervision, checking, and recovery.

Higher-reliability time horizons are substantially shorter than 50% horizons. This is the practical reliability gap.

The amount of work an agent can attempt is growing faster than the amount of work we can safely delegate without supervision.

5. High-reliability estimates are themselves sensitive to methodology

METR’s 2026 methodological analysis shows that high-reliability time horizons can be sensitive to the assumed shape of the fitted success curve.

Language models sometimes fail even on very short, apparently easy tasks. A statistical model that assumes success becomes almost certain as tasks get shorter can therefore distort the high-reliability tail.

So a claim such as “the 80% horizon is exactly X minutes” should be treated as a versioned estimate, not a timeless constant.

6. METR’s low-context limitation matters more than the headline number

METR explicitly notes that its human task-duration baselines are closer to what a low-context skilled contractor, new hire, or freelancer could complete than what a deeply embedded employee could do.

Real workplace performance often depends on undocumented history, internal conventions, previous conversations, tacit knowledge, codebase familiarity, and organizational context.

A strong self-contained benchmark result therefore does not automatically translate into equivalent high-context workplace autonomy.

7. GDPval changes the target from successful completion to valuable deliverables

METR asks how long a task an agent can complete. GDPval asks whether an experienced professional judges the resulting work product as comparable to professional human work.

GDPval’s first version covers:

  • 9 U.S. GDP sectors
  • 44 occupations
  • 1,320 full-set tasks
  • 220 tasks in the open-source gold subset
  • tasks created and reviewed by professionals with more than 14 years of experience on average

These are professional deliverables such as legal documents, engineering plans, spreadsheets, support artifacts, care plans, presentations, diagrams, and other work products.

8. GDPval is not a 320-task benchmark

The lecture materials describe GDPval as roughly 320 tasks. The official first-version specification is 1,320 tasks, with 220 in the gold subset.

The lecture materials also list an incorrect arXiv identifier.

The verified GDPval paper is:

arXiv:2510.04374 — GDPval: Evaluating AI Model Performance on Real-World Economically Valuable Tasks

9. GDPval’s early frontier result is a historical snapshot, not the current ceiling

The original GDPval release found frontier models approaching professional deliverable quality. Claude Opus 4.1 was the strongest model in that release and was judged as good as or better than human expert work in just under half of comparisons.

That historical snapshot should not be presented as the current frontier ceiling.

Later OpenAI reporting on the same GDPval family showed:

  • GPT-5.2 Thinking: 70.9% wins-or-ties
  • GPT-5.4: 83.0% wins-or-ties

Those later numbers use newer models and evaluation settings.

Professional deliverable quality has continued to improve rapidly, but model version, benchmark setup, reasoning effort, and scaffolding must be recorded with every score.

10. Faster and cheaper does not mean autonomous

Raw model generation can be dramatically faster and cheaper than expert labor. But real use often requires task specification, file preparation, review, retry, correction, integration, and accountability.

GDPval explicitly analyzes workflows that include human oversight and repair.

The practical economic question is not only model cost. It is:

How much total human time remains after AI generation, review, and repair?

11. DeepScholar-Bench tests a different failure mode

DeepScholar-Bench asks systems to generate a paper’s related work section using recent research literature.

Its evaluation has three major dimensions:

  1. Knowledge synthesis
  2. Retrieval quality
  3. Verifiability

A system must retrieve the right foundational literature, identify relevant technical contributions, connect papers correctly, and attach citations that actually support the claims being made.

12. DeepScholar-Bench is arXiv:2508.20033

The lecture materials list DeepScholar-Bench as arXiv 2501.09876.

The verified paper is:

arXiv:2508.20033 — DeepScholar-Bench: A Live Benchmark and Automated Evaluation for Generative Research Synthesis

The early arXiv version reported that no evaluated system exceeded 19% across all metrics.

A later ICLR revision reports that no system surpassed a 31% geometric mean across metrics.

Those numbers should not be interpreted as a simple 19-to-31 performance trend because the benchmark version, evaluated systems, and aggregation changed.

Generative research synthesis remains far from saturated.

13. The fluency trap is one of the most important evaluation failures

Research agents can produce writing that looks highly credible: polished prose, coherent structure, and many citations.

None of those guarantee that the citations actually support the claims.

DeepScholar-Bench is valuable because it separates presentation quality from retrieval and verifiability.

The more fluent an output becomes, the more important it is to verify the evidence beneath it.

14. The three benchmarks measure three different things

METR

Question: How difficult a task can the agent complete at a given reliability level?

Primary concern:

  • duration
  • autonomy
  • failure probability
  • low-context robustness

GDPval

Question: Is the deliverable comparable to professional human work?

Primary concern:

  • quality
  • usefulness
  • multimodal work products
  • economic relevance

DeepScholar-Bench

Question: Can the system retrieve, synthesize, and verify evidence correctly?

Primary concern:

  • source selection
  • synthesis
  • citation grounding
  • verifiability

A strong result on one benchmark does not imply a strong result on the others.

15. The strongest counterargument: benchmarks may measure the benchmark environment more than deployment readiness

All three benchmark families simplify reality.

METR uses mostly self-contained, well-specified digital tasks.

GDPval v1 is largely one-shot and cannot fully capture repeated interaction, organizational context, changing priorities, or stakeholder negotiation.

DeepScholar-Bench focuses specifically on generative research synthesis.

None directly measures organizational politics, hidden incentives, permissions, changing objectives, legal accountability, multi-month memory, or long-term ownership.

Each benchmark isolates one deployment-relevant dimension. Deployment readiness requires several of those dimensions to hold at the same time.

16. What this article does not claim

  • The lecture did not establish that a 60-minute METR horizon means 60 minutes of literal autonomous wall-clock work.
  • It did not establish that the seven-month doubling trend must continue indefinitely.
  • It did not establish that 50% success is a production-ready reliability threshold.
  • It did not establish that GDPval’s original near-50% frontier result is the current frontier ceiling.
  • It did not establish that low model inference cost equals low total deployment cost.
  • It did not establish that DeepScholar-Bench remained permanently capped at 19%.
  • It did not establish that success on software-heavy benchmarks generalizes to all professions.
  • It did not establish that benchmark performance alone proves job replacement.

17. The real thesis of Part 8

Agent evaluation is becoming multidimensional because agent capability itself is multidimensional.

A system may operate over longer tasks yet fail too often; produce professional-looking deliverables yet require heavy review; retrieve large amounts of literature yet attach weak citations; or perform well in low-context benchmark settings yet struggle inside a high-context organization.

The key transition is from asking whether a model is “smart” to asking whether an agent is delegable.

That requires at least:

capability + reliability + context handling + output quality + verifiability + recovery

The frontier is not simply how long an agent can work. The frontier is how much work we can delegate before the expected cost of failure and supervision overwhelms the value of autonomy.

Part 9 closes the series with Future Research Areas.

Claim Map

Speaker / course claims

  • long-horizon agent evaluation is more informative than saturated short benchmarks
  • reliability falls as task difficulty and horizon increase
  • economic-quality and research-synthesis benchmarks expose failures not visible in short correctness tests

Verified facts

  • METR’s original study found an approximately seven-month historical doubling trend in 50% task-completion time horizon
  • Claude 3.7 Sonnet was around 55–59 minutes at 50% reliability depending on setup
  • METR defines time horizon using human expert task duration, not agent wall-clock time
  • METR’s current methodology warns that estimates above 16 hours are unreliable with the current task suite
  • GDPval v1 covers 9 sectors, 44 occupations, 1,320 tasks, with a 220-task gold subset and experts averaging more than 14 years of experience
  • GDPval paper: arXiv:2510.04374
  • later GDPval-family results materially exceed the original near-50% frontier snapshot
  • DeepScholar-Bench paper: arXiv:2508.20033
  • early DeepScholar-Bench version: no system exceeded 19% across all metrics
  • later ICLR revision: no system surpassed 31% geometric mean across metrics

Editorial interpretation

  • task duration, reliability, economic usefulness, and verifiability should be tracked separately
  • benchmark success is evidence about a controlled dimension, not direct proof of workforce replacement
  • the practical metric of agent progress is delegability, not benchmark length alone

Lecture Map

  • 00:00–00:28 — METR, time horizons, reliability, and low-context limits
  • 00:28–00:52 — GDPval and professional deliverable quality
  • 00:52–01:08 — DeepScholar-Bench and research verifiability
  • 01:08–01:15 — long-horizon reliability and deployment implications

Primary Sources