LLM Quality Benchmark Landscape, May 2026 — Survey for tinygpt

Research compiled by an Explore subagent on 2026-05-29 to cover the gap between my Jan 2026 knowledge cutoff and current state. Includes URLs.

1. Open LLM Leaderboard

HuggingFace’s Open LLM Leaderboard v2 (MMLU-Pro, IFEval, BBH, GPQA, MuSR, MATH-Lvl5) has effectively been deprecated. Reports indicate the official Space has been intermittently broken in May 2026, and a v3-style overhaul (“Open LLM Leaderboard 2026”) rolled out earlier in 2026 with anti-contamination controls causing the largest reshuffle in its history. MuSR and MATH-Lvl5 saturated; current composition leans on MMLU-Pro, GPQA-Diamond, IFEval, LiveCodeBench, AIME 2025/2026, MATH-500, SWE-bench Verified. Plain MMLU is now considered “outdated/saturated” and excluded from non-saturated leaderboards.

2. Contamination-Resistant

LiveBench still the gold standard; 21 tasks across 7 categories, monthly additions + 6-month full refresh, drawn from post-cutoff arXiv/news/IMDb/contests.
ZebraLogic (1000 logic-grid puzzles, programmatically generated, controllable complexity via Z3 conflict count) — actively maintained, last updated May 2026.
New entrants: BeyondBench (Sep 2025) and ThinkBench (dynamic OOD reasoning) explicitly target contamination resistance via synthesis.

3. Frontier Reasoning

ARC-AGI-2: GPT-5.5 leads at 85%, GPT-5.4 Pro 83.3%, Gemini 3.1 Pro 77.1% (May 2026). Still separates frontier models.
ARC-AGI-3 launched March 25, 2026 — interactive games, no instructions. All frontier models <1%, humans 100%.
HLE (Humanity’s Last Exam): still low single-to-low-double digits for frontier models; agi.safe.ai hosts the live board.
FrontierMath: T1–T3 nearing >95% per 2026 forecasts; T4 still discriminating.

4. Code/Agent

SWE-Bench Verified: Claude Opus 4.7 leads at 87.6% (Apr 2026).
LiveCodeBench: contamination-free, continuously scraped from LeetCode/AtCoder/Codeforces — preferred over HumanEval/MBPP which are saturated.
GAIA: Claude Sonnet 4.5 leads at 74.6% (Princeton HAL).
τ-bench, AgentBench, WebArena, OSWorld: Berkeley RDI (2026) found 8 of these gameable to near-perfect without solving tasks — treat with caution.

5. Long-Context

RULER (NVIDIA, 13 tasks × 4 categories) — still the credible synthetic bar. Frontier models effectively use 50–65% of advertised context.
LongBench v2: Claude Opus 4.5 leads at 64.4% — nearing saturation.
MRCR v2 and NoLiMa are the newer harder needle variants. Advertised vs effective context diverges by 30–60 points past 200K tokens.

6. Runnable on a Single Mac

Fully local, no judge model:

LiveBench (math/coding/IF programmatic graders; “language” subset uses LLM judge — skip or use a local Qwen3 judge)
ZebraLogic (Z3-verified, deterministic)
RULER (synthetic, exact-match)
LongBench v2 (multiple choice — exact match)
GSM8K, MATH-500, AIME, GPQA-Diamond, MMLU-Pro, IFEval, BBH, HumanEval/LiveCodeBench — all via lm-evaluation-harness (supports MLX through local-chat-completions; watch the stop-sequence bug in 0.4.11)
TinyStories PPL and bits-per-byte stay valuable for byte-level models. The NeurIPS 2025 E2LM competition explicitly addresses early-training (<200B tokens) evaluation for sub-7B models — directly relevant to tinygpt.

Requires paid API judge (cost gate):

MT-Bench, AlpacaEval, Arena-Hard, GAIA-text-judge, HLE auto-grader, GPQA open-ended — all need GPT-4/Claude grader
SWE-Bench Verified runs Dockerized test suites locally but is heavy (50+GB images, hours per model)
τ-bench requires a user-simulator LLM

7. New since Jan 2026

ARC-AGI-3 (Mar 2026) — interactive
AIME 2026 — fresh, contamination-free math
ReasonBench (Dec 2025) — variance-aware reasoning evaluation across 11 methods × 4 models × 7 tasks
BeyondBench (Sep 2025) — synthesis-based contamination resistance
LongBench Pro — bilingual realistic long-context extension
Berkeley RDI agent benchmark exploit paper (2026) — caveat on trusting τ/SWE/WebArena scores

Recommendation for tinygpt

Wire in (all local, no judge cost):

bits-per-byte on held-out TinyStories/Shakespeare (already have it) — extend with TinyStories-eval split
HellaSwag-easy + ARC-Easy via lm-eval-harness for byte-level signal early
GSM8K + IFEval + GPQA-Diamond for HF-loaded Llama models
RULER short-context (4K/8K) — gives a credible context-effectiveness number
LiveBench programmatic subset — math + coding + reasoning, no judge

Skip until external budget: MT-Bench, Arena-Hard, GAIA, HLE, full SWE-Bench.