LLM evaluation
14 published articles
Special Report: AI Evaluation
The State of AI Benchmarking in 2026: Inside the Collapse of Static Tests and the Systems Built to Replace Them
A comprehensive tour of the 2026 AI benchmarking landscape: why MMLU, GSM8K, and HumanEval broke, how dynamic benchmarks and expert exams like HLE and GPQA Diamond replaced them, what agentic and jagged-intelligence testing reveals, and how human preference, LLM judges, and production observability now round out the full evaluation stack.
2026-08-03
Benchmark Analysis
LiveBench Refuses to Sit Still. That's the Whole Point
LiveBench replaces a fixed answer key with a continuously refreshed pool of coding problems, repositories, and prediction questions, sidestepping the contamination that undermined MMLU and GSM8K. It's part of a broader shift toward dynamic evaluation alongside LiveCodeBench, ForecastBench, and LLMEval-Fair.
2026-08-03
Benchmark Analysis
5 Million Votes, 25 Elo Points: Inside the Chatbot Arena Leaderboard Nobody Can Shake
LMSYS Chatbot Arena ranks models by blind human preference on an Elo scale, with nearly five million votes now packing the top labs into a 25-point band. The LLM-as-a-Judge methods that scale this kind of evaluation carry their own documented biases toward verbosity and position.
2026-08-02
Benchmark Analysis
On SWE-bench Verified, Top Models Hit 96%. On Private Enterprise Code, They Barely Clear 23%
SWE-bench Verified makes frontier models look close to solving real-world software engineering, with top scores above 95%. SWE-bench Pro, run on private enterprise repositories, drops those same models to 23% or lower, exposing how much of the Verified score depended on public data exposure.
2026-08-02
Benchmark Analysis
GPQA Diamond Was Built to Be Google-Proof. Frontier Models Are Now Clearing 95% Anyway
GPQA Diamond was designed so search-equipped humans can't reliably answer its expert-level science questions. Frontier models are now clearing 89% to 96%, pushing the benchmark toward the same saturation that retired MMLU.
2026-08-01
AI research
Over half your AI's reasoning is froth, and nobody noticed until now
LLMs often generate reasoning chains that are correct but padded with unnecessary steps. A new diagnostic benchmark shows current evaluators miss this inefficiency entirely, and half of human-written reasoning steps may be compressible.
2026-07-31
Benchmark Analysis
Humanity's Last Exam Opened With a 2.7% Score. The Best Models Still Can't Break 65%
Humanity's Last Exam was built by CAIS and Scale AI to succeed MMLU as the hardest general LLM benchmark. Two years on, even Claude Opus 5's leading 64.7% score sits well below the 90% human expert baseline.
2026-07-31
Benchmark Analysis
HumanEval Measured Whether AI Could Code. It Never Asked Whether the Code Was Real Work
HumanEval's 164 function-completion problems became the standard test for AI coding ability, but memorization and its narrow scope left a wide gap between passing the benchmark and handling a real codebase, a gap SWE-bench was built to expose.
2026-07-30
Grade-School Math Benchmark Quality Audit
GSM8K Was Supposed to Test Grade-School Math. Up to 42% of Its Questions Don't
GSM8K became a standard test of arithmetic reasoning in LLMs, but audits have found error rates in its question set as high as 42%, undermining what its scores actually measure.
2026-07-29
Benchmark Analysis
MMLU Ruled AI Benchmarking for Years. Then Models Started Acing the Answer Key
MMLU defined a generation of AI benchmarking, but training-data contamination and a flawed answer key have pushed frontier labs toward dynamic successors like HLE and GPQA Diamond.
2026-07-28
Foundations of AI
DeepMind's new framework shows why AI is brilliant at some things and terrible at others
A Google DeepMind paper introduces a three-part framework for thinking about rigor in AI: conceptual, epistemic, and operational. It argues that deep learning's rapid progress has come from prioritizing performance-driven iteration over scientific understanding, and that closing the gaps will require more than better benchmarks.
2026-07-20
Agent evaluation
Sandbox benchmarks are hiding how agents really fail. HKU just built the fix.
UniClawBench evaluates proactive agents across five fundamental capabilities in 400 bilingual real-world tasks, using live Docker containers and a three-agent closed-loop evaluation. It disentangles base model abilities from framework choices, revealing where agents truly break.
2026-07-13