Evals & Benchmarks, in order.

Map of Content for the evals-and-benchmarks domain — 11 concepts. The science of measuring models: benchmark validity, contamination, saturation, LLM judges, and production-sourced evaluation. Curated entry point; see Home for all domains.

• Benchmark Contamination and Decontamination — Sun, Zhan & Gales (Cambridge, arXiv 2606.23313): benchmark data contamination inflates reported LLM scores when test samples leak into training, and existing decontamination methods are judged only by aggregate accuracy and mostly need a clean reference model. Two contributions: (1) a sample-level evaluation framework — per-sample distribution distances (D_KL and D*_L1 between the decontaminated and an uncontaminated model's output distributions) that expose a fundamental disconnect (paraphrase+permutation cuts dataset-level residual contamination 17.2→8.4 while D_KL actually rises >13%); (2) Uncertainty-Based Decontamination (UBD), which needs no clean reference model and no knowledge of which samples are contaminated — deep ensembles of the contaminated model (5 LoRA seeds differing only in batch ordering) expose memorized samples as high-confidence-but-high-variance (epistemic/knowledge uncertainty), because memorization is batch-order-sensitive. A per-sample scalar α̂ ≈ 1−2σ (σ = ensemble std of the ground-truth probability, above a confidence threshold T_p) drives either UBD-Debiasing (post-hoc suppression of the inflated ground-truth mass, redistributed proportionally, no weight update, MCQ-only) or UBD-Unlearning (fine-tune toward the debiased soft target; no-op when α̂≈1). On MMLU-Pro/MATH-MCQA it gives >40–60% relative D_KL reduction and beats paraphrase/permutation and even the reference-model DeconIEP baseline; ensemble uncertainty tracks oracle contamination at PCC 0.8–0.9 where log-probability stays <0.4
• Benchmark Score Redundancy — Zeng & Papailiopoulos (Microsoft Research, arXiv 2606.24020): an 84-model × 133-benchmark public score matrix (2,604 cells, 23.3% filled) is effectively rank-2 — a model's scores across all 133 benchmarks are largely two numbers (two diagnostics agree: Soft-Impute completion bottoms at rank 2, and top-2 SVD components explain >90% of cross-model variance in every fully-observed submatrix), echoing prior g-factor / 'general-capability + provider-residual' findings on a heterogeneous frontier-era matrix. BenchPress (logit-space rank-2 bias-decomposed ALS matrix completion) predicts every held-out cell to 4.6 MedAE points at 100% coverage; a compact 5-benchmark probe set {GPQA-D, HLE, Codeforces, MMLU-Pro, ARC-AGI-1} recovers a model's full scorecard to 3.93 points (low-cost set 4.55), predictions preserve 92.1% of ≥5-point pairwise model rankings, and a strict-temporal-cutoff new model needs only 5 seed scores to reach 4.83 points — so you can infer most of a scorecard from a handful of probes rather than run every eval, with a hybrid ensemble-spread + matrix-support reliability layer flagging which predictions to trust; but scores are inferable, not benchmarks unnecessary, the rank-2 geometry is snapshot-specific, and the matrix is the same uncontrolled public grid the compute-control critique warns about
• Compute-Controlled Benchmarking — Noam Brown's critique that the single-number 'benchmark grid' is broken because it doesn't control for test-time compute; the fix is to plot performance against a cost/token/time x-axis; benchmark-maxxing, held-out private sets, the Goodhart bad-equilibrium that keeps the grid alive, why routing/consensus must be judged at equal budget — and Gemma 4's headline table as the worked example, benchmarking a thinking model against a non-thinking predecessor
• DRACO Benchmark — Perplexity's benchmark of 100 production-sourced deep-research tasks (10 domains, 40 countries) graded by 26-expert rubrics on accuracy/completeness/objectivity/citation; Perplexity Deep Research leads every domain and axis, Claude Opus 4.6 is the strongest non-Perplexity system, factual accuracy is the universal weak spot
• LLM-as-a-Judge — Using one LLM to grade another's outputs against criteria/rubrics; DRACO's protocol is per-criterion binary MET/UNMET + justification, weight-aggregated into normalized score and pass rate; key properties — rankings stay stable across judge models while absolute magnitudes vary, and adaptive per-case rubrics (Google's AutoRaters) detect failures but blend them away, motivating stable custom metrics for the behavior under change; a 21-judge / ~541K-judgment audit finds raw exact-match agreement overstates chance-corrected reliability by 33–41pp (kappa deflation) and high test-retest can mask severe position bias, so judges need chance-correction, bias, and cross-benchmark validation before thresholded use
• LLM-Judge Validation — UC Berkeley's 21-judge / 9-provider / ~541K-judgment audit (Norman et al., 2026): LLM-as-a-judge validation is systematically under-rigorous — exact-match agreement overstates chance-corrected κ by 33–41pp (kappa deflation, universal across every judge), judge rankings shift up to 14 positions across benchmarks, and high test-retest reliability masks severe position bias (the consistency–bias paradox); distilled into a 5-step Minimum Viable Validation Protocol
• Measuring Beyond Accuracy Saturation — Nadgir, Kapoor, … Narayanan (Princeton-led, 14 authors, arXiv 2606.26158): when a benchmark's accuracy saturates (top agents statistically indistinguishable), the field's reflex is retire-and-replace with a harder successor — but that privileges accuracy and throws away six other measurable axes. Using CORE-Bench Hard (computational reproducibility of scientific code) as a case study, they decouple accuracy saturation from benchmark saturation: (1) saturation itself surfaces construct-validity threats invisible to weaker agents — log analysis via Docent found 15 task-level errors + 20 exploitable shortcuts, yielding corrected CORE-Bench v1.1 (39 tasks) and an OOD suite CORE-Bench OOD (19 tasks, new fields), yet accuracy still saturates (top agent 100%, next four tie ~97.4%); (2) statistically-indistinguishable agents still differ sharply in reliability (more-accurate agents are more consistent; all are massively underconfident, ~93% pass vs 32.1% stated confidence, and none beats random at telling its own correct from incorrect runs), efficiency (GPT-5.3-Codex ~60% cheaper than an equal-accuracy peer; token-cost and dollar-cost tell different stories), and model-vs-scaffold contribution (scaffold swings accuracy ~44pp on one model; two scaffolds on the same model disagree on 31% of tasks but an oracle router hits 100%; direct fixes beat rewrites 95%/68%); (3) a small randomized human study finds agent collaboration more than halves reproduction time (2.11×, p≈0.002, likely an underestimate since 5/25 manual runs hit the 3-hour wall and no collaborative run did). Thesis: don't retire a saturated benchmark — re-instrument it
• Production-Sourced Evaluation — Building benchmarks from de-identified real production usage rather than synthetic or hand-authored tasks; DRACO's central method — difficulty-proxied sampling, PII-stripping, augmentation, automatable refresh with a human QA gate; representativeness vs. over-specification tradeoff; production traffic as a proprietary eval asset
• Reference-Free Judge Over-Crediting — Kranti & Vajjala (arXiv 2607.12885): the presence and placement of a reference answer in the prompt is a first-order determinant of an LLM judge's correct/incorrect verdicts — in no-reference settings judges systematically over-credit incorrect answers, and adding the reference flips up to 85% of verdicts (most NR→RV flips are CORRECT→INCORRECT over-credit retractions), an effect that holds across English/Arabic/Telugu but is far worse in low-resource Telugu; a two-stage calibration (C1 vs C2: can the judge tell right from wrong) + sensitivity (NR/RV/RC: does reference presence flip verdicts) pipeline diagnoses judge suitability before reference-free deployment, and human annotation confirms the reference-driven stricter verdicts align better with people (agreement jumps ~0.34→0.85 NR→RV for Gemma), so reference-free absolute scores are genuinely inflated not merely different — a reliability-without-validity instance orthogonal to Norman et al.'s, on three open/open-weight judge models at temperature 0 in binary QA
• Scale-Dependent Prompt Sensitivity — Large models underperform small ones on 7.7% of standard benchmarks due to overthinking; brevity constraints recover 26pp and fully reverse hierarchy on GSM8K/MMLU-STEM
• Task Time-Horizon Scaling — METR's measure of the task length AI can complete reliably on its own, doubling roughly every 4 months (up from every 7): Opus 3 ~4min (Mar 2024) → Opus 4.6 ~12hr (2026) → weeks projected for 2027; paired with benchmark saturation (SWE-bench, CORE-Bench)