Deep Dive · 5 of 5
Evaluation and the Open Problems
The final deep dive in this series. The field handbook covered long-horizon evaluation and the field's remaining open problems in a few paragraphs each. This piece expands both — METR's time-horizon methodology, GDPval, and DeepScholar-Bench with verified citations, plus a closer look at meta-verification, self-generated curricula, and continual learning, the three problems nobody in this series has solved yet.
On the citations below: all three benchmarks' identities are high confidence, corroborated across multiple independent sources. METR's original 7-month doubling-time figure (from their March 2025 paper) is high confidence; a claimed acceleration to a roughly 3-4 month doubling time in 2026, and any specific 2026 model results, come only from secondary forum/blog commentary on METR's own updates and are explicitly flagged as unverified — check metr.org directly for current figures before citing them. DeepScholar-Bench is a joint Stanford/UC Berkeley effort, not solely Stanford, despite living in the same intellectual neighborhood as this course.
Why Evaluation Gets Hard
METR's Time-Horizon Method
GDPval
DeepScholar-Bench
The Three Unsolved Problems
Closing the Series
Confidence Notes & Gaps
Full Reference List
Why Evaluation Gets Hard
Every technique across this five-part series was justified with a benchmark number — HumanEval pass rates, Codeforces percentiles, AIME scores, QA benchmark gains. Those numbers share a property that makes them easy to trust and easy to over-trust: a short task, a fixed correct answer, graded in isolation. An agent that has to complete a task made of dozens or hundreds of sequential decisions, using tools, under ambiguous instructions, with no single "check the final answer" moment, doesn't fit that mold — and the three benchmarks below were each built specifically because the old mold stopped being enough.
METR: Measuring Capability as a Time Horizon
Measuring AI Ability to Complete Long Tasks HIGH
Defines a "50%-task-completion time horizon": the length of a task, measured by how long it takes a human professional, that a model can complete with 50% success. Rather than asking what fraction of fixed questions a model answers correctly, this reframes capability as how much real-world task duration is actually reachable — exposing something accuracy scores hide, since small per-step error rates compound multiplicatively across a long task. The original paper reported this time horizon doubling roughly every 7 months since 2019, with Claude 3.7 Sonnet at publication reaching around 50 minutes.
The trend is a real, well-sourced curve; the exact current doubling rate and most recent data points are contested and should be checked at metr.org directly, not taken from this diagram's shape.
Why "reliably" matters more than "capable." An agent that completes a hard task 20% of the time and one that completes it 95% of the time both demonstrate the task sits within the model's reach — but only the second is something you'd actually delegate. If per-step success is 98% and a task needs 100 sequential correct steps, overall success probability is roughly 0.98¹⁰⁰ ≈ 13%. That's why METR's methodology tracks a 50%-success threshold specifically, and why a rising time horizon doesn't automatically mean an agent is safe to hand real, unsupervised work to at that duration.
GDPval: Measuring Economic Usefulness Directly
GDPval: Evaluating AI Model Performance on Real-World Economically Valuable Tasks HIGH
1,320 tasks across 44 occupations spanning the top nine GDP-contributing sectors, built from real work products — reports, spreadsheets, slide decks, diagrams — created by professionals averaging around 14 years of experience, graded by expert comparison against actual human deliverables rather than a single "correct answer." The paper reports frontier model quality closing in on human-expert level on this benchmark and improving roughly linearly over time; a 220-task "gold" subset is openly released with automated grading.
DeepScholar-Bench: Evaluating Research Synthesis
DeepScholar-Bench: A Live Benchmark and Automated Evaluation for Generative Research Synthesis HIGH
A "live" benchmark that draws its queries and reference material directly from recent arXiv papers, testing whether an agent can write a paper's related-work section by retrieving, synthesizing, and correctly citing prior research — graded on knowledge synthesis quality, retrieval quality, and verifiability, alongside a reference pipeline called DeepScholar-Base. The headline finding: the benchmark is far from saturated — no evaluated system, including OpenAI's Deep Research, exceeded roughly 19% on the combined metrics at time of publication, a sharp contrast with how close frontier models get on shorter-form benchmarks.
📄 arXiv:2508.20033
📺 No specific dedicated talk/interview independently confirmed
Why DeepScholar-Bench's low ceiling matters more than METR's or GDPval's numbers: METR shows agents can complete meaningfully long tasks, and GDPval shows they're closing in on human-expert quality on well-scoped professional deliverables. DeepScholar-Bench, evaluating the same kind of open-ended research synthesis this whole course keeps returning to, shows the opposite — a task where even the strongest available systems are still under a fifth of the way to good. Read together, these three don't tell one consistent story about "how close are we" — they tell you the answer depends entirely on which kind of task you're asking about.
The Three Unsolved Problems
Every deep dive in this series named a place where the field genuinely doesn't have an answer yet. Pulled together, they're really three problems, not many:
1
Meta-verification. Deep Dive 1's Constitutional AI showed a verifier can itself be an AI judge rather than a human — but nothing in this series checks whether that judge is trustworthy. Nearly every technique across all five deep dives depends on some external reference to check against: a test suite, a reward model, a known answer. A system that has to judge reasoning with no reference answer at all — is this internally coherent, does the cited evidence actually support the conclusion, can an independent model reproduce the result — runs into a verifier-of-the-verifier regress nobody in this literature has fully closed.
2
Self-generated tasks and curricula. Every benchmark in this series, including the three above, was built by humans. If humans have to keep hand-writing every new training or evaluation task, improvement stays bottlenecked on human task-authoring capacity. A more autonomous loop would have the system generate its own next task, attempt it, verify the outcome, and generate a harder one next — with the genuine difficulty being keeping generated tasks in a productive zone: hard enough to teach something, not so hard every attempt fails and there's no signal at all.
3
Continual learning and memory. Every technique across all five deep dives — STaR, GRPO, DAPO, Search-R1, all of it — follows the same train-then-freeze-then-deploy cycle. None of them let a deployed agent keep learning from its own live experience. Doing that safely opens a long list of problems this series hasn't touched: catastrophic forgetting, instability from noisy or adversarial feedback, and the practical question of what a long-lived agent should actually remember versus discard across working memory, episodic memory, and learned procedural strategies.
Closing the Series
🎯 The Bottom Line — For the Whole Series
Five deep dives, thirteen techniques, one recurring idea: generate, verify, keep what works, then train the system to reproduce the good version directly — applied to single answers (Deep Dive 1), multi-step plans (Deep Dive 2), the model's own weights (Deep Dive 3), and which candidate or which search query to trust (Deep Dive 4). Every technique that worked, worked because its verification step was reliable — execution results for RLEF, test cases for AlphaCode, a group's own relative scores for GRPO. Every open problem this deep dive just named is a case where that verification step doesn't exist yet: judging reasoning with no reference answer, generating a task worth learning from, deciding what a system should remember. If you take one idea out of this entire five-part series, take this one: the ceiling on self-improving agents was never generation. It was always verification, and it still is.
⚠️ Confidence Notes and Gaps
All three benchmarks' identities, authors, and core methodology are high confidence, corroborated across multiple independent sources. METR's original 7-month doubling-time figure is high confidence, directly from their March 2025 paper; any claim of an accelerated ~3–4 month doubling rate in 2026, and any specific recent model's time-horizon score, comes only from secondary forum and blog commentary on METR's later updates, not independently confirmed against metr.org directly in this research pass — treat those specific numbers as unverified and check the source before citing them. DeepScholar-Bench is correctly a joint Stanford/UC Berkeley project; don't attribute it to Stanford alone. None of the three papers were fetched directly from arxiv.org in this research environment.
🔗 Full Reference List
Self-Improving AI Agents: A Field Handbook (base article)
Deep Dive 1 — Agentic Feedback Loops in Practice
Deep Dive 2 — Planning as Search
Deep Dive 3 — Training on What Search Finds
Deep Dive 4 — Competitive Programming and Retrieval Agents
METR — arXiv:2503.14499
METR — blog post
GDPval — arXiv:2510.04374
DeepScholar-Bench — arXiv:2508.20033