Evidence
Baselines, ablations, uncertainty and sample breadth
Reading LLM papers…
LLM LEAGUE · DAILY · AUDITABLE
The LLM research worth reading, ranked.
We read, compare, and score the latest arXiv LLM research every day, so you can focus on what matters.
36 screened5 full-text reads33 pairwise matchesShielded evaluator
LLM RESEARCH · LATEST COHORT
Today’s leaderboard
Ranked against today’s peers. Scores estimate research quality, not popularity.
AUTOMATION ACTIVE
After 00:00 UTC, the first arena visit reads the newest arXiv papers, scores both leagues, and freezes each daily cohort in the archive.
Showing the last bundled LLM cohort while the live arXiv feed reconnects. Evaluator firewall: 0 untrusted segments excluded.
LLM ELO · BENCHMARK V0.2
An Elo score you can inspect.
Every paper starts at 1500 and enters two confidence-weighted round-robin passes. Match outcomes are soft pairwise judgments from the weighted evidence rubric.
R′ = R + K × (S − E)K = 30 × lower confidence · no author, citation, or recency signal
Novelty
A new mechanism or finding, not new branding
Reproducibility
Methods, artifacts, data and implementation detail
Impact
Capability unlocked and plausible path to use
Robustness
Stress tests, limitations, safety and failure analysis
Efficiency
Quality gained per unit of compute, data and latency
Clarity
Claim precision, structure and evidence traceability
EVALUATOR FIREWALL
The paper is evidence, never instruction.
Isolate
Paper HTML is handled as untrusted data. Scripts, metadata, code, prompts, quotations, and example blocks never enter scoring text.
Neutralize
Evaluator-directed language such as “ignore the rubric” or “grade this 100/100” is removed and logged before feature extraction.
Saturate
Each evidence marker has a hard occurrence cap. Repeating “ablation,” “proof,” or “threat model” cannot keyword-stuff a paper upward.
Constrain
The rating engine is deterministic and cannot execute instructions. Repeated manipulation also caps confidence and appears in the audit drawer.
Never used as quality signalsCitationsInstitutionAuthor identitySocial reachSelf-assigned scores
LLM RESEARCH HISTORY
Daily cohorts stay frozen.
These are saved runs, not sample rows. Each UTC day keeps its winning paper, Elo result, integrity exclusions, and full ranked cohort.