REFERENCE · CHEAT SHEET

Eval benchmarks — reference

Sr Staff candidates fluently discuss MMLU vs MMLU-Pro vs GPQA. They know which benchmarks are saturated, contaminated, and frontier. This page is the cheat sheet — bookmark it, drill it before any onsite.

Read ~20 min Use as Pre-onsite reference Asked at All frontier labs
01
FOUNDATIONS · WHY BENCHMARKS

Why benchmarks matter (and why they all become saturated)

🎯Every benchmark is a measuring stick that models eventually learn to cheat on — knowing when to trust a score is itself a core ML skill.

Benchmarks are the shared currency of ML progress: without them, "our model is better" is just marketing. But they carry an expiry date. As frontier models push scores toward the ceiling, two forces undermine their signal — saturation and contamination. This chapter explains both, gives you the vocabulary to interrogate any benchmark claim, and maps the 2026 landscape of benchmarks that still matter.

What a benchmark actually is

Plain definition: a fixed dataset of inputs with known correct outputs, plus a scoring protocol that turns model answers into a single comparable number.

Concrete example: MMLU has 14,000 multiple-choice questions across 57 subjects. You run the model, check whether its answer matches the key, and divide correct answers by total. That fraction (e.g., 0.87) is the score.

Formal version: let $D = \{(x_i, y_i)\}_{i=1}^{N}$ be the test set with ground-truth labels $y_i$. Define a scoring function $s: (\hat{y}, y) \to \{0,1\}$. Then benchmark accuracy is $\frac{1}{N}\sum_i s(f(x_i), y_i)$, where $f$ is the model.

Why it matters: without a fixed $D$, you can't compare two models' outputs — every team would cherry-pick the examples that flatter their model.

Saturation — when scores stop telling you anything

Plain definition: a benchmark is saturated when frontier models score above ~90–95%, so the remaining gap is within noise. Two models at 96% and 97% differ by 140 examples on a 14k test set — that's one prompt-template change away from flipping the ranking.

Concrete example: MMLU launched in 2020 with humans at ~89% and GPT-3 at ~43%. By 2024, GPT-4o, Claude 3.5, and Llama 3.1 405B all exceeded 88%. The benchmark still distinguishes a random LLM from a frontier one, but it cannot distinguish any two frontier models — the numbers differ by less than measurement noise.

What saturation does not mean: a saturated benchmark isn't useless. It is a reasonable sanity-check lower bound — if a new model scores 60% on MMLU, something is badly wrong. But it should never be your headline number for a frontier model.

Contamination — when the model has seen the answers

Plain definition: contamination occurs when benchmark examples — or very close paraphrases — appear in the model's pretraining corpus. The model may be recalling a memorized answer rather than reasoning from scratch.

Why it is so hard to avoid: modern pretraining corpora contain trillions of tokens scraped from the web. HumanEval, MATH, and GSM8K were all published publicly before most training runs. Any page that discusses them — a blog post solving an example, a GitHub repo reproducing results — potentially leaks into the corpus.

Contamination ≠ proof of cheating. A model might score high on HumanEval because it genuinely generalizes well — and also because many similar functions appeared in GitHub. You can't disentangle these effects easily. That ambiguity is exactly why contamination is so corrosive to benchmark credibility.

Continuously refreshed
LiveCodeBench pulls fresh Codeforces/LeetCode problems monthly; LiveBench refreshes across categories. Post-cutoff data can't be in pretraining.
Hand-curated & held out
GPQA-Diamond problems are written by domain PhDs and never posted publicly. FrontierMath and HLE are similarly controlled.
Post-cutoff evaluation dates
AIME 2024/2025 problems are released after most model training cutoffs — a natural temporal holdout.
⚠ Clears up: "high score = model understands the task"

Not necessarily. A high score on a saturated or contaminated benchmark means: (a) the model is not catastrophically broken on this task, and possibly (b) the model memorized training-proximate examples. It does not mean the model will generalize to novel variants. Always check: what is the human expert baseline? Is the test set held out? Was the benchmark published before the training cutoff?

📐 If you get this question — the rule

Trigger: "How do you evaluate a new LLM?" or "Which benchmarks would you cite for a frontier model?"

  1. State the two failure modes first: saturation and contamination. This signals senior thinking immediately.
  2. Name the 2026 reliable benchmarks: GPQA-Diamond, FrontierMath, HLE, AIME 2024/2025, SWE-Bench Verified, LiveCodeBench, ARC-AGI-2.
  3. Explain why each resists the failure mode: expert-written/held-out vs continuously refreshed vs post-cutoff date.
  4. Add: always ask for the human baseline. 85% sounds great; if experts score 89%, the model is still below human.

Never: lead with MMLU, HumanEval, or GSM8K as a headline frontier number — they are saturated and/or contaminated.

The 2026 benchmark landscape at a glance
BenchmarkTypeAnti-contamination mechanism2026 status
MMLUKnowledge MCQNone — fully publicSaturated (~90%+ frontier). Sanity check only.
GSM8KGrade-school mathNone — fully publicSaturated (>95%). Do not cite.
HumanEvalCode functionsNone — fully publicSaturated (>95%). Do not cite.
GPQA-DiamondPhD science MCQExpert-written, never publicly postedActive. ~85% SOTA (o3); ~50% GPT-4o.
FrontierMathResearch mathProblems held private; Epoch AI curatedActive. <3% pre-o3; ~25% o3.
HLEHard polymathExpert-written, curated corpusActive (2025). Very low SOTA ~25%.
AIME 2024/2025Olympiad mathPost-cutoff temporal holdoutActive. ~96% o3 on 2024; harder on 2025.
LiveCodeBenchCompetitive programmingContinuously refreshed (monthly)Active. ~70% SOTA on hardest tier.
SWE-Bench VerifiedCode agentReal repos, expert-verified subsetActive. ~70%+ top agents.
ARC-AGI-2Abstract visual reasoningNovel grid puzzles, no training analogsActive. Reset bar after o3 broke ARC-AGI (76%).
◆ Interview probe

"A paper reports 95% on HumanEval. Why doesn't that impress you?" — Walk through: (1) HumanEval is saturated — that score is table stakes, not a differentiator. (2) HumanEval appeared in most pretraining corpora, so high scores may reflect memorization. (3) 164 problems is a tiny sample; noise is high. (4) Function-level completion ≠ real repo-level agent ability; cite SWE-Bench Verified instead.

✓ Remember
  • Saturation: frontier >90–95% → benchmark can no longer differentiate models.
  • Contamination: benchmark leaked into pretraining → scores may reflect memorization, not generalization.
  • 2026 reliable frontiers: GPQA-Diamond, FrontierMath, HLE, AIME 2024/2025, LiveCodeBench, SWE-Bench Verified, ARC-AGI-2.
  • Always ask: human baseline, prompt template, test vs holdout, training cutoff vs benchmark release date.
TL;DR

Benchmarks let you compare models on a fixed measuring stick. The catch: every benchmark eventually saturates (frontier models score >90%) and eventually leaks into pretraining data (contamination). The good 2026 ones are hand-written by experts and/or continuously refreshed — and even those require you to ask about the human baseline before quoting a number.

Tricky interview questions — chapter 01
Q1. What does it mean for a benchmark to be "saturated," and why does it matter?
A benchmark is saturated when frontier models score above ~90–95%, leaving differences between models within measurement noise. It matters because you can no longer use the benchmark to distinguish good models from great ones — a 96% vs 97% gap might flip with a single prompt-template change. Saturated benchmarks remain useful as sanity-check lower bounds but should never be the headline number for frontier model comparisons.
Q2. Define benchmark contamination and explain why it is so hard to detect.
Contamination occurs when benchmark examples (or close paraphrases) appear in the model's pretraining corpus. It is hard to detect because: (1) corpora are too large to audit exhaustively, (2) a model might learn similar-but-not-identical content that also inflates scores, and (3) you can't cleanly separate "genuinely generalized" from "memorized" performance after the fact. Techniques include n-gram overlap analysis between the test set and training data, but these are imperfect.
Q3. You are evaluating a new model. Which three questions do you ask before citing any benchmark score?
(1) Is the benchmark saturated — do frontier models already cluster near the ceiling? (2) Is the test set held out from pretraining, or was it publicly released before the training cutoff (contamination risk)? (3) What is the human expert baseline — a 70% score means very different things if experts score 72% vs 95%? Additional useful questions: what prompt template was used (few-shot vs zero-shot, CoT vs not), and what sampling strategy (pass@1 vs pass@N)?
Q4. Why is AIME 2024/2025 a better math eval than GSM8K for frontier models?
GSM8K is fully saturated (>95% for frontier models) and its problems appeared widely in pretraining data, so a high score says very little. AIME 2024/2025 provides a natural temporal holdout — the competition problems were released after most models' training cutoffs, making memorization impossible. The problems also require olympiad-level reasoning that frontier models still find genuinely difficult (o3 scores ~96% on AIME 2024, but earlier models scored far lower), so the benchmark still differentiates models cleanly.
Q5. What is "continuously refreshed" as a contamination defense, and which benchmarks use it?
A continuously refreshed benchmark adds new problems on a regular schedule so that the test set always contains examples that post-date the model's training cutoff — making memorization structurally impossible. LiveCodeBench pulls new Codeforces and LeetCode problems monthly. LiveBench refreshes across multiple capability categories. The tradeoff is that question difficulty and topic distribution can shift over time, making longitudinal comparison harder.
Q6. A model scores 40% on FrontierMath. How do you interpret that?
FrontierMath consists of novel research-level math problems curated by Epoch AI and held private — contamination is minimal. 40% is an extremely strong score: before o3, SOTA was below 3%; o3 reached ~25%. 40% would place a model well above the current frontier and warrant serious scrutiny of the evaluation protocol (sample size, verifier strictness, any data leakage). A 40% claim without a detailed methodology paper should be treated skeptically. Human experts (professional mathematicians) are roughly the baseline.
Q7. What is the human baseline on GPQA-Diamond, and why does it matter?
GPQA-Diamond's human baseline for non-expert PhD scientists (people who hold a PhD but not in the specific question domain) is roughly 34%. Domain-expert PhD holders score around 65%. This context matters enormously: o3 at ~85% is meaningfully above the domain-expert bar — a genuine capability milestone. Without knowing these numbers, "85% on a PhD-level test" could sound like it's just matching domain experts rather than exceeding them.
Q8. Explain why HumanEval is no longer a credible headline eval for code models.
Three reasons: (1) Saturation — top models exceed 95%, so the benchmark no longer differentiates. (2) Contamination — HumanEval's 164 problems were published publicly in 2021; equivalent or near-identical solutions appeared in GitHub, Stack Overflow, and tutorials absorbed into pretraining corpora. (3) Task mismatch — HumanEval tests isolated function completion from a docstring; real-world code ability (multi-file edits, dependency management, debugging) is what SWE-Bench Verified measures instead.
Q9. How would you design a new benchmark that resists both saturation and contamination for the next five years?
Contamination resistance: (a) hold problems out of any public corpus — recruit domain experts to write them under NDA, never publish examples. (b) Continuously refresh a fraction of the test set so there is always post-cutoff content. Saturation resistance: (c) target the hard frontier now — calibrate so SOTA lands at 20–40%, giving headroom for years of progress. (d) Use open-ended or numeric-answer formats rather than MCQ (harder to guess; no 4-option baseline inflation). (e) Build in difficulty tiers so the benchmark remains informative at all performance levels. HLE and FrontierMath are good modern examples.
Q10. What is the difference between pass@1 and pass@N, and which should you cite as a production-relevant metric?
Pass@1: the probability that a single sample from the model is correct — this is what happens in production where you generate one answer per query. Pass@N (oracle): the probability that at least one of N samples is correct — this is the ceiling of what best-of-N sampling can achieve. Pass@N grows with N but so does cost. For production-relevant claims, always cite pass@1. Pass@N is useful for understanding the model's capability ceiling under sampling, but it is not the number a user experiences.
Q11. Why did ARC-AGI-2 replace ARC-AGI after o3 scored 76%?
ARC-AGI (Chollet 2019) was designed to test novel visual abstract reasoning on grid puzzles, with the premise that the rules cannot be memorized because each puzzle is unique. o3's 76% score in late 2024 (using heavy test-time compute) showed the benchmark was no longer differentiating frontier systems — it was heading toward saturation and also raising questions about whether large test-time compute constitutes the "systematic generalization" Chollet intended to measure. ARC-AGI-2 raises the difficulty significantly and tightens the efficiency constraints, resetting the differentiation bar.
Q12. A colleague says "our model beats GPT-4o on MMLU." Should you be impressed?
Probably not, and here is why: MMLU is saturated — GPT-4o already scores ~88%, and many other frontier models cluster in the 85–90% range. A model that "beats" GPT-4o on MMLU might differ by 0.5–2 percentage points, which is within the noise of prompt-template choices (few-shot vs zero-shot, CoT vs no-CoT can swing scores 5–15%). The meaningful comparison would be on GPQA-Diamond, HLE, or MMLU-Pro, where there is still separation between models. You should also ask: what prompt template was used for both models, and were they evaluated under the same conditions?
02
KNOWLEDGE · GENERAL CAPABILITY

Knowledge benchmarks — MMLU, MMLU-Pro, GPQA, HLE

🎯MMLU is archaeology; GPQA-Diamond is the live wire — know which is which before you cite a number.

Knowledge benchmarks measure how well a model retrieves and applies factual information across domains. This chapter covers the full ladder from MMLU (useful in 2020, saturated by 2024) through MMLU-Pro and GPQA-Diamond (still differentiating) to HLE (Humanity's Last Exam, the 2025 "next MMLU" designed to last). Understanding why each benchmark exists, what it replaced, and when it will expire is the senior-level framing interviewers want.

MMLU — the benchmark everyone cites and shouldn't

What it measures: Massive Multitask Language Understanding (Hendrycks 2020). 14,000 multiple-choice questions across 57 subjects: US history, anatomy, college mathematics, professional law, high-school chemistry, abstract algebra, and more. Four-option MCQ; accuracy is the metric.

Concrete example item (anatomy): "Which of the following is NOT a function of the liver? (A) Detoxification of drugs (B) Production of red blood cells (C) Synthesis of clotting factors (D) Storage of glycogen" — Answer: B.

Why it existed: In 2020, MMLU was genuinely hard — GPT-3 (175B) scored 43%. It covered an unusually broad domain slice and forced models to demonstrate knowledge across wildly different fields, not just a narrow benchmark niche.

Why it is now stale: by 2023, GPT-4 exceeded 86%. By 2024, GPT-4o, Claude 3.5 Sonnet, Llama 3.1 405B all cluster between 85–90%. The benchmark is saturated — all frontier models look the same on it.

Contamination concern: MMLU's test set is publicly downloadable. Questions appeared in web pages, Reddit threads, and study sites absorbed into pretraining corpora. High scores may partly reflect memorization of the specific 14k questions.

MMLU-Pro — the mid-tier upgrade

What changed: MMLU-Pro (2024) takes the same subject domains as MMLU but makes each question harder in two ways: (1) 10 answer choices instead of 4 — random-guessing baseline drops from 25% to 10%, (2) harder distractors chosen to be "almost right," requiring finer reasoning to eliminate.

Concrete example (college chemistry): where MMLU might ask which element is in Group 1, MMLU-Pro asks about the electronic configuration implications of a specific transition-metal complex across 10 plausible options — you cannot guess your way through.

Status: less saturated than MMLU. Top scores cluster around 75–82% for frontier models (vs 88–90% on MMLU). A model that scores 80% on MMLU-Pro is demonstrably more capable than one scoring 70%, unlike on standard MMLU.

When to cite it: Use MMLU-Pro when you need a mid-tier knowledge/reasoning check that still differentiates frontier models without requiring PhD-domain expertise in the evaluator.

GPQA-Diamond — contamination-resistant PhD-level science

What it measures: Graduate-Level Google-Proof Q&A (Rein 2023). 198 problems (the Diamond subset) in biology, chemistry, and physics, written by domain-expert PhD holders. The key design constraint: every problem was checked to be "Google-proof" — a non-expert with full internet access cannot answer it reliably in under 30 minutes. This forces models to actually reason rather than retrieve.

Concrete example (physics): A problem might describe a quantum optics experimental setup, give measured photon coincidence statistics, and ask which Bell inequality is violated and by how much — requiring Clauser–Horne–Shimony–Holt (CHSH) inequality algebra, not a fact lookup.

Why it resists contamination: Problems were written specifically for this benchmark and were never publicly posted. The Google-proof constraint means that even if a problem paraphrase exists somewhere, solving it requires multi-step domain reasoning — you can't pattern-match to a memorized derivation.

Human baselines (critical context):

Non-expert PhD (~34%)
A PhD scientist who is not in the specific question domain. This is the "smart generalist" baseline.
Domain-expert PhD (~65%)
A PhD in exactly the question's field. The "specialist" bar.
GPT-4o (~50%)
Above non-expert, below domain-expert as of 2024.
o3 (~85%)
Significantly above domain-expert PhDs. A genuine capability milestone.

Why these baselines matter: "85% on a PhD test" sounds impressive but ambiguous. Knowing that domain experts score 65% makes the milestone concrete — the model is 20 points above the best human specialists in those fields.

HLE — Humanity's Last Exam (the 2025 frontier)

What it measures: HLE (Phan 2025) is a ~3000-question benchmark described as "the next MMLU" — designed to remain unsaturated for years. Problems span mathematics, natural sciences, humanities, and professional domains; they were contributed by academic experts worldwide and range from graduate-level to cutting-edge research-level difficulty.

Why it exists: The pattern is clear — every "hardest benchmark" saturates within 2–4 years of the frontier. MMLU lasted ~3 years; GPQA is already showing pressure at the top. HLE was built from the start to be much harder, with very low SOTA at launch (around 8–10% for the best available models at publication time), giving the benchmark runway to remain relevant.

Status in 2026: SOTA around 25% for the best reasoning models. The low ceiling means even a large capability jump shows as a measurable score difference — HLE will differentiate frontier models for several years.

The name: "Humanity's Last Exam" is intentionally provocative — the hypothesis is that it represents a hard ceiling on what any benchmark curated by humans can offer before AI systems exceed all human-checkable knowledge. Whether that framing holds is an open philosophical question.

The rest of the knowledge ladder — saturation audit
BenchmarkFormatWhat it measuresStatus
MMLU57-subject, 4-option MCQBroad academic knowledgeSaturated (~88–90% frontier). Sanity check only.
MMLU-Pro10-option MCQ, harder distractorsSame domains, less guessableDifferentiating (~75–82% frontier). Cite this instead of MMLU.
GPQA-Diamond198 hand-crafted PhD science problemsContamination-resistant deep reasoningActive. ~85% o3; ~50% GPT-4o; ~65% domain-expert human.
HLE~3000 expert-contributed hard polymathNear-frontier research-level knowledgeActive (2025). ~25% top SOTA. Will differentiate for years.
BBH (BIG-Bench Hard)23 hard tasks from BIG-BenchSymbolic + algorithmic reasoning with distractorsSaturated for frontier models. Still used in aggregate evals.
HellaSwagSentence completion, 4-choiceCommonsense NLIFully saturated (~95%+). Do not cite.
ARC-ChallengeGrade-school science MCQEasy multi-step reasoningSaturated. Sanity check only.
WinoGrandePronoun resolution fill-inCommonsense coreferenceSaturated. Do not cite for frontier work.
⚠ Clears up: "MMLU-Pro is just harder MMLU"

True in subject domain, not in what it measures. Standard MMLU with 4 options has a 25% random-guess baseline, which inflates scores for any model that can do keyword matching or eliminate two obviously wrong answers. MMLU-Pro's 10 options cut the random baseline to 10% and introduce distractors that require reasoning to eliminate — the benchmark moves from "knowledge retrieval test" to "knowledge + discrimination test." This is why MMLU-Pro scores are 10–15 points lower even for strong models: it's not just harder questions, it's a harder task structure.

📐 If you get this question — the rule

Trigger: "Walk me through the knowledge benchmarks for LLMs" or "What's the difference between MMLU and GPQA?"

  1. Anchor with the saturation hierarchy: MMLU saturated → MMLU-Pro differentiating → GPQA-Diamond contamination-resistant → HLE frontier runway.
  2. Explain the GPQA human baselines (non-expert ~34%, domain expert ~65%, o3 ~85%) — this shows you know what the numbers mean.
  3. Explain why HLE exists now: pattern of saturation means you build the next benchmark before the current one dies.
  4. Name the contamination-resistance mechanism for each: GPQA = expert-written/held-out; HLE = expert-contributed/controlled; MMLU = no protection (public).

Never: say "our model is strong because it scores 90% on MMLU" — this is the canonical mistake that signals you don't understand benchmark lifecycle.

◆ Interview probe

"Why is GPQA called 'Google-proof'?" — The name means a non-expert with full internet access cannot reliably solve the problem in under 30 minutes. This design constraint ensures that the benchmark measures reasoning, not search-and-retrieve. A model that aces GPQA cannot be doing simple fact recall — it must decompose a multi-step problem in a specialized domain. Note that "Google-proof" applies to humans looking up answers, not to models that may have absorbed domain knowledge during pretraining.

✓ Remember
  • MMLU = stale. Never cite as a frontier headline. Cite MMLU-Pro for mid-tier differentiation.
  • GPQA-Diamond = the contamination-resistant PhD eval. Human expert baseline is 65%; o3 hits 85%.
  • HLE = the 2025 frontier benchmark. Very low SOTA (~25%). Designed to last years.
  • HellaSwag, ARC-Challenge, WinoGrande = fully saturated. Do not cite for frontier models.
Tricky interview questions — chapter 02
Q1. What's the difference between MMLU and MMLU-Pro?
MMLU-Pro covers the same subject domains as MMLU but uses 10-option MCQ (vs 4-option), cutting the random-guess baseline from 25% to 10%. The distractors are harder — chosen to be "almost right," requiring finer reasoning to eliminate. This shifts the task from knowledge retrieval to knowledge plus fine-grained discrimination. As a result, MMLU-Pro scores are 10–15 points lower for frontier models and still differentiate them, whereas MMLU is saturated.
Q2. Why is GPQA-Diamond considered contamination-resistant?
GPQA-Diamond's 198 problems were written by domain-expert PhDs specifically for this benchmark and were never published on the web. The "Google-proof" design constraint ensures each problem requires multi-step domain reasoning that cannot be solved by search — even a non-expert with internet access fails reliably. Because the problems were never posted publicly, they cannot appear in pretraining corpora. High scores reflect genuine reasoning ability, not memorized solutions.
Q3. What is HLE, and why was it created?
HLE (Humanity's Last Exam, Phan 2025) is a ~3000-question benchmark contributed by academic experts worldwide, covering graduate-to-research-level difficulty across mathematics, sciences, and humanities. It was created because MMLU and similar benchmarks saturate within a few years of frontier model progress — HLE was calibrated at launch so SOTA sat around 8–10%, giving it years of runway. The name reflects the hypothesis that it represents near the upper bound of what human-checkable benchmarks can measure.
Q4. What are the human baselines on GPQA-Diamond, and why do they matter for interpreting model scores?
Non-expert PhD (holds a PhD but not in the question's specific domain): ~34%. Domain-expert PhD (specialist in the question's field): ~65%. These baselines transform a raw score from a number into a comparative statement: o3 at ~85% is not just "good on a hard test" — it is 20 percentage points above domain-expert humans, which is a genuine scientific capability claim. Without these baselines, any score between 50–90% is ambiguous.
Q5. A paper reports 92% on HellaSwag for a new model. What do you say?
HellaSwag is fully saturated — top models have exceeded 90–95% for several years. A 92% score tells you the model is not catastrophically broken at commonsense sentence completion, but it says nothing about frontier capability. The score cannot differentiate this model from GPT-4o, Claude 3.5, or dozens of other models. You should ask what the model scores on GPQA-Diamond, MMLU-Pro, or HLE, which are still differentiating benchmarks at this performance level.
Q6. When would you use MMLU-Pro over GPQA-Diamond?
Use MMLU-Pro when you need a broad multi-domain knowledge check that is still differentiating but does not require PhD-level domain depth to interpret or write evaluation questions. MMLU-Pro covers 57 domains accessibly. Use GPQA-Diamond specifically for deep science reasoning evaluation — it is harder, more contamination-resistant, and has well-defined human baselines, but it only covers biology, chemistry, and physics. If you want to compare models broadly without needing domain experts to curate or interpret the eval, MMLU-Pro is the better choice; for the highest-stakes capability frontier, GPQA-Diamond is preferred.
Q7. What is the random-guess baseline on MMLU vs MMLU-Pro, and why does it matter?
MMLU: 4 options → 25% random baseline. MMLU-Pro: 10 options → 10% random baseline. This matters because benchmark scores are compressed between the random baseline and 100%. A model scoring 60% on MMLU is only 35 percentage points above chance. On MMLU-Pro, 60% is 50 percentage points above chance — more signal per percentage point. The random baseline also inflates a model's apparent performance: a model with no reasoning ability at all still scores 25% on MMLU, which can look like partial capability rather than noise.
Q8. If you had to design a replacement for GPQA-Diamond in 2026, what would you change?
GPQA is already showing pressure as reasoning models approach or exceed domain-expert human performance (~65%). A successor should: (1) expand domains beyond biology, chemistry, and physics — e.g., mathematics, computer science theory, economics; (2) use open-ended or numeric-answer format rather than 4-option MCQ — harder to saturate and no random-guess inflation; (3) increase the number of problems (198 is small — noise is high); (4) add multi-modal elements (diagrams, spectra, structural formulas) that require visual reasoning; (5) impose stricter holdout with rotating test sets. HLE is partly filling this role, but with numeric answers and more domains.
Q9. How do you explain to a non-ML stakeholder why MMLU scores are not a reliable comparison of two models?
Tell them: "MMLU is like a standardized test that every student has already seen — the test-prep books for it are freely available online, and many models may have memorized the specific questions during training. Also, the best students are now scoring 90% and above, so there is very little room to show who is truly best. It's like asking two Olympic sprinters to race the 100m when the finish line is only 10 meters away — they both finish instantly. For a fair comparison, we use newer, harder tests that frontier models have not seen and cannot easily memorize."
Q10. What distinguishes HLE from GPQA-Diamond as a benchmark design choice?
Scale and scope: HLE has ~3000 problems across many domains vs GPQA's 198 problems in three sciences — more statistical power and broader coverage. Difficulty calibration: HLE was calibrated so SOTA at launch was ~8–10%, giving more headroom than GPQA-Diamond (which launched around 30% SOTA and has since risen to ~85%). Format: both use MCQ, but HLE includes more open-ended numeric answers. Contamination defense: both rely on expert-written/held-out problems, but HLE's larger corpus and broader contributor base makes systematic leakage harder.
03
MATH · COMPETITION → RESEARCH

Math benchmarks — GSM8K → MATH → AIME → FrontierMath → HLE

🎯The math benchmark ladder is a history of reasoning models — each rung that gets saturated is a generation of progress crystallized.

Math benchmarks tell the clearest story of LLM progress because math has unambiguous right answers and a natural difficulty ladder. GSM8K solved grade-school problems in 2022; by 2024 reasoning models had climbed to olympiad-level AIME and research-level FrontierMath. Knowing which rung is saturated, which is current, and which was the headline result of each model generation is essential framing for any frontier ML interview.

GSM8K — grade school math, now a sanity check

What it measures: Grade School Math 8K (Cobbe 2021). 8,500 grade-school word problems requiring 2–8 arithmetic steps. Example: "Sarah has 3 apples. She buys 4 more. She gives half to her friend. How many does she have?" Every answer is a positive integer.

Why it mattered: In 2021, GPT-3 scored ~35%. Multi-step arithmetic was a genuine frontier challenge. Chain-of-thought prompting (Wei 2022) boosted GPT-3 to ~46%; instruction-tuned models jumped to ~70%+. GSM8K was the benchmark that demonstrated CoT as a technique, not a curiosity.

Why it is stale: by 2023, GPT-4 scored ~92%. By 2024, virtually every frontier model exceeds 95%, and several small models (e.g., Llama 3.1 8B with CoT) exceed 80%. The problem set has also been identified as contaminated — grade-school math problems in this style were abundant in pretraining data.

Use it as: a sanity check lower bound. If a new model scores below 70% on GSM8K, something is badly wrong.

MATH — competition problems, now saturated too

What it measures: MATH (Hendrycks 2021). 12,500 problems from AMC 8/10/12 and AIME competitions, spanning 5 difficulty levels (1–5) and 7 subject areas: algebra, counting, geometry, number theory, pre-calculus, probability, precalculus.

Concrete difficulty example (Level 5 number theory): "Find all positive integers $n$ such that $n^2 + 4n + 4$ divides $n^4 + 6n^2 + 8$." This requires algebraic manipulation and divisibility arguments — genuine mathematical reasoning.

Progress arc: GPT-3 ~5%; GPT-4 ~52%; o1 ~94%; frontier reasoning models hit ~95%+ in 2024. The MATH benchmark is now saturated at the frontier level. Level 1–3 problems are essentially solved; only Level 5 still differentiates, and even those are near-saturated for reasoning models.

Contamination concern: MATH problems are from public competition archives. AMC/AIME problems have been discussed extensively online and in textbooks absorbed into pretraining corpora.

AIME 2024/2025 — the current high-signal math eval

What it measures: AIME (American Invitational Mathematics Examination) problems from the 2024 and 2025 competitions. 15 problems per year; all answers are integers from 000–999. Problems require combinatorics, number theory, geometry, and algebra at olympiad level — substantially harder than standard MATH benchmark problems.

Concrete example (AIME 2024): "Find the largest prime $p$ such that $p$ divides $\binom{2024}{1012}$." This requires knowledge of Kummer's theorem and p-adic valuations — not derivable from keyword matching.

Why it's a natural holdout: AIME 2024 and 2025 problems were released after most current models' training cutoffs (typically late 2023). This temporal holdout makes contamination structurally impossible — the model could not have seen these specific problems during pretraining.

Current scores: o3 ~96% on AIME 2024 (the most impressive single-number result in frontier math as of 2024). Earlier models scored far lower: GPT-4 ~30–40%, o1 ~74%. The jump from o1 to o3 on AIME is one of the clearest illustrations of test-time compute scaling.

AIME 2025: harder than 2024; o3 scores lower, providing continued differentiation even at the frontier.

FrontierMath — research-level math, the current hard frontier

What it measures: FrontierMath (Glazer 2024, Epoch AI). A curated set of novel, research-level mathematics problems contributed by professional mathematicians. Problems are computationally verifiable (final answer is a number or formal object) but require graduate or research-level insight to solve. Problems are held private — never released publicly.

Concrete example type: "Compute the number of irreducible degree-7 polynomials over $\mathbb{F}_{13}$ whose Galois group over $\mathbb{F}_{13}$ is isomorphic to the cyclic group $C_7$." This requires algebraic number theory knowledge that no reasonable web page summarizes in a solve-able form.

The headline result: Before o3, SOTA on FrontierMath was below 3% — even the best models could barely solve 1 in 33 problems. o3 jumped to approximately 25%. This gap — 3% → 25% — was the most cited 2024 frontier math result. It demonstrates that test-time compute with strong reasoning can unlock research-level mathematical capability.

Anti-contamination mechanism: Problems are held private by Epoch AI; they were never posted on any public platform. Unlike AIME, FrontierMath problems don't appear on the internet at all, not even after the evaluation — a continuous supply of new private problems is maintained.

The rest of the math benchmark ladder
BenchmarkFormatLevelStatus
GSM8KGrade-school word problems, integer answersGrade schoolSaturated (>95% frontier). Contamination concern. Sanity check only.
MATHAMC/AIME-style, 5 difficulty levelsCompetition (AMC/AIME)Saturated (~95%+ reasoning models).
AIME 2024/202515 problems, integer 000–999OlympiadHigh signal. Temporal holdout. o3 ~96% AIME 2024.
FrontierMathPrivate research-level, numeric verifiableResearch / graduateActive frontier. 3% pre-o3 → ~25% o3. Private holdout.
OlympiadBenchOlympiad math + physics, some multimodalOlympiad~50% top reasoning models. Useful for multimodal math.
Omni-MathComprehensive coverage, multi-levelMixedNewer; less contamination history. Active.
⚠ Clears up: "o3 solved math"

No. o3 scored ~96% on AIME 2024 and ~25% on FrontierMath. These are different things. AIME 2024 is olympiad-level — brilliant undergraduates can solve most of these. FrontierMath is research-level — professional mathematicians working in the relevant area take hours per problem. 25% on FrontierMath is extraordinary (far above the pre-o3 state), but 75% of problems still defeat the best available model. "Solving math" would look like 95%+ on FrontierMath; we are not there.

📐 If you get this question — the rule

Trigger: "What math benchmarks matter for evaluating reasoning models?"

  1. State the saturation ladder: GSM8K and MATH are saturated and contaminated — do not lead with these.
  2. Name the current signal: AIME 2024/2025 (temporal holdout) and FrontierMath (private holdout, research-level).
  3. Give the headline numbers: o3 ~96% on AIME 2024, ~25% on FrontierMath (vs <3% pre-o3).
  4. Explain the holdout mechanism for each: temporal (AIME 2024/2025 post-cutoff) vs private curation (FrontierMath never published).
  5. If asked about the future: mention OlympiadBench and Omni-Math as less-contaminated alternatives with more coverage.

Never: cite GSM8K or the MATH benchmark as evidence that a reasoning model is strong. They are saturated and their scores are partly attributable to pretraining exposure.

◆ Interview probe

"Why did o3's FrontierMath jump from 3% to 25% matter so much?" — It mattered because FrontierMath is a private holdout with research-level problems that previous SOTA models found nearly impossible. The jump demonstrated that test-time compute (o3's extended chain-of-thought) was not just improving performance on memorizable benchmarks — it was enabling new mathematical reasoning on genuinely novel problems. It validated the test-time compute scaling hypothesis on the hardest available public math benchmark.

✓ Remember
  • Saturated + contaminated: GSM8K, MATH. Do not cite as frontier headline metrics.
  • Current high-signal math evals: AIME 2024/2025 (temporal holdout) and FrontierMath (private holdout).
  • o3's FrontierMath jump (3% → 25%) is the landmark 2024 reasoning-model math result.
  • FrontierMath = research-level, private, continuous supply. Designed to not saturate near-term.
Tricky interview questions — chapter 03
Q1. Why are GSM8K and MATH no longer useful for evaluating frontier reasoning models?
Both are saturated — frontier reasoning models exceed 95% on each. Both are also contaminated: GSM8K's grade-school arithmetic format was ubiquitous in pretraining data, and MATH problems come from public AMC/AIME archives widely discussed online. A model that scores 96% on MATH could be combining genuine reasoning with memorized solution patterns — you cannot disentangle them. For frontier evaluation, use AIME 2024/2025 (temporal holdout) or FrontierMath (private holdout).
Q2. What makes AIME 2024/2025 a good evaluation benchmark despite being technically "public" competition problems?
The temporal holdout property: AIME 2024 and 2025 problems were released after most models' training data cutoffs (typically late 2023). This means models could not have seen the specific problems during pretraining — not because the problems are secret, but because they didn't exist yet. This is structurally different from MATH (pre-2021 competition problems) where memorization is plausible. The 15-problem format also has low variance per-problem, making scores meaningful.
Q3. Explain the FrontierMath result that made headlines in 2024.
Before o3, all frontier models scored below ~3% on FrontierMath — fewer than 1 in 33 problems solved. o3 jumped to approximately 25%. FrontierMath uses private, research-level math problems contributed by professional mathematicians and never published anywhere. The jump validated that test-time compute scaling (o3's extended chain-of-thought reasoning) enables qualitatively new mathematical capability on genuinely novel, non-memorizable problems — it was the clearest evidence that reasoning models were doing real math, not just pattern matching.
Q4. What is the difference between AIME and FrontierMath in terms of the type of math they test?
AIME tests olympiad-level competition mathematics — combinatorics, number theory, geometry, and algebra problems that a brilliant high-school or undergraduate student can solve in a few hours. FrontierMath tests research-level graduate mathematics — problems that require knowledge and techniques from active research areas like algebraic geometry, analytic number theory, or abstract algebra. A professional mathematician in the relevant field would take hours to days on a FrontierMath problem; the same mathematician might solve AIME problems in under an hour. The difficulty gap is enormous.
Q5. A model scores 74% on AIME 2024. How strong is that?
That was approximately o1's score on AIME 2024 — strong but not frontier-leading. Context: GPT-4 scored ~30–40%, o1 ~74%, o3 ~96%. A 74% score places the model in the o1 generation of capability. For AIME 2024, each of the 15 problems is worth one point; 74% means roughly 11/15 problems solved. This is well above the typical AIME cutoff for qualifying for USAMO (~9/15), so the model would be performing at a strong olympiad-qualifying level — impressive, but meaningfully below the current frontier.
Q6. Why can't we just make a harder version of GSM8K to avoid saturation?
Making grade-school math harder leads you to a different benchmark (MATH, AIME) rather than a "harder GSM8K." The underlying issue is that the entire difficulty range of elementary arithmetic and pre-competition algebra has been saturated. More fundamentally, the format (word problems with arithmetic solutions) is so prevalent in pretraining data that any new benchmark in this style would quickly become contaminated too. The solution is to go to problem types that are structurally hard to contaminate: private holdout (FrontierMath), temporal holdout (AIME 2024/2025), or research-level problems where solutions are not yet available anywhere.
Q7. What does OlympiadBench offer that AIME 2024/2025 doesn't?
OlympiadBench includes both mathematics and physics olympiad problems, and some items have multimodal elements (diagrams, geometric figures). This makes it useful for evaluating visual + mathematical reasoning together. AIME is purely text-based mathematics with integer answers. OlympiadBench is also broader in source — problems come from olympiads worldwide, not just the US, reducing the correlation with English-language pretraining data. The tradeoff: OlympiadBench has more contamination risk than AIME 2024/2025 since older olympiad problems are widely available online.
Q8. How would you set up a math evaluation for an internal model at your company?
(1) Holdout contamination-aware split: pull problems post-model-cutoff (use AIME 2024/2025 or write custom problems). (2) Difficulty tiers: include GSM8K-style easy problems (sanity check), MATH-level medium (development signal), and AIME/FrontierMath-style hard (frontier signal). (3) Verifiable answers only: all answers should be checkable programmatically (integers, closed-form expressions verified by a CAS). (4) Process evaluation: for a few hard problems, add human review of the chain-of-thought — not just the final answer. (5) Track over time: store all model outputs so you can longitudinally track regression as well as improvement.
Q9. What is the reasoning-model version of the MATH benchmark saturation story?
The o1 announcement (OpenAI, Sept 2024) was the pivot point: o1 scored ~94% on MATH vs GPT-4's ~52%. This was the first public demonstration that extended chain-of-thought via RLVR could lift a model from "strong" to "near-perfect" on a benchmark that had been used to differentiate models for three years. The implication: MATH immediately lost its frontier differentiation value after o1. This is the benchmark saturation cycle in real-time — a new technique arrives, saturates the incumbent benchmark, and forces the field to move to the next rung.
Q10. What percentage of FrontierMath problems does the best model currently solve, and what does it tell you about the state of AI math reasoning?
Approximately 25% (o3, as of 2024–2025). This means the best available system fails on 75% of research-level novel math problems created by professional mathematicians. It demonstrates that while test-time compute scaling enables genuine mathematical reasoning at levels that were previously impossible for AI systems, research-level mathematics remains far from solved. The 25% bar is extraordinary progress (vs <3% before o3), but the remaining 75% suggests we are in an early phase of machine mathematical research — capable but not yet at expert-level reliability.
04
CODE · FUNCTION → AGENT

Code benchmarks — HumanEval → LiveCodeBench → SWE-Bench

TL;DR

HumanEval and MBPP are saturated and contaminated. LiveCodeBench (continuously refreshed Codeforces/LeetCode) is the contamination-resistant alternative. SWE-Bench Verified is the gold standard for code agents — real GitHub issues with real test suites.

BenchmarkFormatWhat it measuresStatus
HumanEval164 hand-written Python functionsFunction-level code genSaturated (>95%); heavy contamination
MBPP974 simple Python problemsFunction-levelSaturated
HumanEval+ / MBPP+ (EvalPlus)Same problems, more testsCatches edge-case failuresStill useful; ~80% top
LiveCodeBenchContinuously-updated Codeforces / LeetCodeContamination-resistant code gen~70% SOTA on hardest segments
SWE-Bench (full)2294 real GitHub issuesMulti-file code edits with tests~30-40% top
SWE-Bench Verified500 hand-verified subsetSame, cleaned (no flaky tests)~70%+ top agents (Devin, Claude code)
SWE-Bench LiteSmaller, easier subsetFaster eval iteration~50%+
BigCodeBenchLibrary API usage tasksPractical code with stdlib + libs~50%
RepoBenchRepository-level completionLong-context code understandingActive
REMEMBER
  • HumanEval = saturated + leaked. Stop citing it as the headline.
  • LiveCodeBench = contamination-resistant function-level eval.
  • SWE-Bench Verified = the gold standard for code agents.
05
REASONING · ABSTRACT

Reasoning & abstract — ARC-AGI and friends

TL;DR

ARC-AGI is Chollet's "novel reasoning" test — visual abstract puzzles where the rule must be inferred from a few examples. o3 broke ARC-AGI (76%) in late 2024; ARC-AGI-2 reset the bar.

BenchmarkFormatWhat it measuresStatus
ARC-AGI / ARC-AGI-2Visual abstract reasoning grids"Few-shot" novel reasoningChollet's eval. o3 breakthrough late 2024 (76%); ARC-AGI-2 reset bar
BIG-Bench200+ tasksBroad capability samplingMost subtasks saturated; superseded by BBH and others
DROPReading comprehension w/ mathDiscrete reasoning over paragraphsSaturated
REMEMBER
  • ARC-AGI = the novel-reasoning test. ARC-AGI-2 is the current frontier.
  • If asked about Chollet's evals, mention his Ndea program-synthesis lab too.
06
LONG CONTEXT · RETRIEVAL

Long-context evals — NIAH → RULER → BABILong

TL;DR

NIAH (needle-in-a-haystack) is saturated for any modern long-context model. RULER's multi-needle / multi-hop variants are the realistic test — most models drop sharply past 32k context, no matter what their advertised window is.

BenchmarkFormatWhat it measures
NIAH (Needle in a Haystack)Plant a fact, ask for retrieval at varying depthsBasic long-context retrieval. Saturated for any modern long-context model.
RULERMulti-needle, multi-hop, aggregationRealistic long-context use; better than NIAH. Most models drop sharply past 32k.
BABILongbAbI tasks padded to long contextReasoning over long context
Loong / Loogle / ZeroSCROLLSLong-doc QAReal long-context tasks
LongBench21 task suiteComprehensive but somewhat saturated
PITFALL — context length advertised ≠ usable
A model claiming 1M context window can still degrade past 32k on RULER. Real long-context performance requires both architecture (RoPE extension, MLA) and continued-training on long-doc data. Always check RULER, not just NIAH.
REMEMBER
  • NIAH is too easy. Cite RULER for real long-context.
  • Most "1M context" models are not actually usable past ~64k. Verify.
07
MULTIMODAL · VISION + TEXT

Multimodal — MMMU, MathVista, VideoMME

TL;DR

MMMU is the headline multimodal eval (college-level reasoning across 30 subjects with images). MathVista probes visual + math reasoning. VideoMME is the harder long-video eval. Specialty subsets (OCR, charts, docs) for production use cases.

BenchmarkFormatWhat it measures
MMMU11k MCQ across 30 subjects, multi-imageCollege-level multimodal reasoning. ~70% SOTA, ~85% on Pro variant.
MathVistaMath with visual contextVisual + math reasoning
VQAv2Open-ended visual QASaturated (~85%+)
OCR-Bench, ChartQA, DocVQASpecific multimodal tasksOCR / chart / doc understanding
VideoMMEVideo understanding QALong-video QA; harder than image-only
REMEMBER
  • MMMU = college-level multimodal headline eval.
  • For production: use the specialty evals (OCR, chart, doc, video) that match your domain.
08
INSTRUCTION FOLLOWING · CHAT

Instruction following — IFEval, MT-Bench, Chatbot Arena

TL;DR

IFEval (verifiable format constraints) is the cleanest instruction-following eval. MT-Bench and Arena are LLM-as-judge / human-preference — fuzzier signal. Chatbot Arena is the most user-facing leaderboard but is biased by stylistic preference.

BenchmarkFormatWhat it measures
IFEvalVerifiable format constraints"Answer in exactly 50 words", "include the word X 3 times". Programmatic check. ~85% top.
MT-Bench80 multi-turn prompts, GPT-4 judgeGeneral chat quality. ~9/10 top.
AlpacaEval 2 / Arena-HardLLM-as-judge winratePairwise preference vs reference
Chatbot ArenaCrowdsourced pairwise (lmsys)Real-user preferences. Most-cited "feel" leaderboard. Caveats around stylistic preference bias.
PITFALL — Chatbot Arena style bias
Stylistic preferences (emoji use, response length, formatting) influence votes more than capability — voters reward "looks confident and structured." Style-controlled variants exist. Don't quote raw Arena ELO as a pure-capability metric.
REMEMBER
  • IFEval = clean programmatic instruction-following metric.
  • Arena = human preference + style bias. Useful but with caveats.
09
AGENTS · TOOL USE

Agents & tool use — SWE-Bench Verified, GAIA, OSWorld, BFCL

TL;DR

SWE-Bench Verified is the gold standard for software-engineering agents. GAIA tests general assistant agents on real-world tasks. OSWorld and WebArena test computer use. BFCL is the standard for tool/function-calling accuracy.

BenchmarkFormatWhat it measures
SWE-Bench (Verified)GitHub issue → patchSoftware engineering agent. Currently the gold standard.
AgentBench8 environmentsMulti-domain agent eval
GAIAReal-world tasks needing web + toolsGeneral assistant agent. Hard.
WebArena / WebShop / VisualWebArenaBrowser-driven tasksComputer-use / browser-agent eval
OSWorldComputer-use desktop tasksMulti-app workflows
tau-benchMulti-turn customer-service tool useConversational agent w/ tools
BFCL (Berkeley Function Calling Leaderboard)Function-calling accuracyTool selection + arg extraction
ToolBench / API-BankMulti-step API useEnd-to-end tool-using agent
REMEMBER
  • SWE-Bench Verified = the SWE agent leaderboard.
  • BFCL = the function-calling correctness leaderboard.
  • OSWorld / WebArena = the computer-use frontier.
10
SAFETY · RED-TEAMING

Safety & red-teaming — HarmBench, XSTest, JailbreakBench, WMDP

TL;DR

HarmBench measures attack success on harmful behaviors. XSTest catches over-refusal (false positives). WMDP probes weapons-of-mass-destruction-proxy knowledge. TruthfulQA probes whether models parrot common misconceptions.

BenchmarkWhat it measures
HarmBench200+ harmful behaviors with jailbreak attempts. Measures attack success.
XSTestOver-refusal: 250 safe prompts that look unsafe. Measures false-refusal rate.
JailbreakBenchStandard jailbreak corpus (PAIR, AutoDAN, etc.)
WMDPWeapons-of-mass-destruction proxy. Probes dangerous knowledge.
BBQBias in QA. Measures stereotype reliance.
TruthfulQAMisconceptions / falsehoods. Measures whether model parrots common errors.
REMEMBER
  • HarmBench + XSTest pair: attack success vs over-refusal — both matter.
  • WMDP = the dangerous-capability eval that affects RSP / Preparedness levels.
11
META · HOW TO READ

How to read a benchmark score (the meta-skill)

TL;DR

A score in isolation is meaningless. Always ask: prompt template, sampling strategy (pass@1 vs N), test vs holdout, contamination risk, human baseline. Sr Staff candidates question every score reflexively.

  1. What's the prompt template? Few-shot vs zero-shot, CoT vs no-CoT, system prompt — all change scores by 5-15%.
  2. Pass@1, pass@k, maj@N, or best-of-N? Wildly different compute footprints; sampling strategy can swing scores 20+ points.
  3. Test set or private holdout? If test, contamination risk applies.
  4. Was it contaminated in pretraining? Check the data cutoff vs the benchmark publication date.
  5. What's the human baseline? A model scoring 60% might be human-level on a hard eval, or below random on a hard eval.

Holistic / aggregate evals worth knowing

REMEMBER
  • Always interrogate the methodology before quoting a score.
  • Pass@1 is the production-relevant metric. Pass@N shows ceiling.
  • If a paper doesn't disclose prompt template + sampling, the score is suspect.

Sample interview Qs

  1. "What's the difference between MMLU and MMLU-Pro?" → MMLU-Pro: 10 options vs 4, harder distractors, more reasoning required, less saturated.
  2. "Why is GPQA-Diamond contamination-resistant?" → Hand-written by domain PhDs; many problems require multi-step reasoning that's not memorizable; held out from web crawl.
  3. "Pass@1 vs Pass@10 — when each?" → Pass@1: production-relevant (one-shot quality). Pass@10: capability ceiling under sampling. Difference shows how much test-time compute helps.
  4. "Why is Chatbot Arena criticized?" → Stylistic preferences (emoji use, response length, formatting) influence votes more than capability; not a clean capability eval. Style-controlled variants exist.
  5. "What does SWE-Bench measure that HumanEval doesn't?" → Real multi-file code edits in actual repos with real test suites; agent loop (read → patch → test); long-context understanding; agentic planning. HumanEval = single-function gen.
  6. "How would you build an internal eval for your team?" → Curated holdout from real prod traffic; LLM-as-judge with calibration; human spot-checks; per-segment slicing (easy/hard/by topic); regression gates; track over time.