Scope and coverage
RLTasks focuses on large language models and language-model agents, including language models that accept images or documents. It does not compare image, speech, or video generators. Interactive 3D work appears as an LLM coding task.
The library contains 469 benchmark entries spanning reasoning, coding, science, tool use, long context, knowledge, behavior, and finance. Coverage starts with Artificial Analysis’s evaluation directory and model comparison page, then extends to benchmarks cited in official model launches and primary research. Ulam’s published evaluations are included with their affiliation disclosed.
Named versions, subsets and metric views have separate entries where their contracts differ. The entry count therefore does not represent that many independent datasets. Composite indexes and personality traits are labeled as such.
What each benchmark contains
Each article explains the capability being measured, the task and scoring design, interpretation limits, and a reproducibility contract. Primary sources and dated model-launch references identify where the benchmark is documented and used. The original metric remains visible.
Published comparisons
Individual benchmark leaderboards collect recent-model measurements from Artificial Analysis’s public evaluation pages, official model-launch reports and technical model cards. Each source and evaluation setting stays in its own cohort. The publisher-selected model subset is a dated snapshot rather than a claim to include every historical run. Raw scores, units, model labels, tool annotations and review dates remain attached to the source. Numeric ordering applies within the selected cohort; it does not establish matched budgets or statistical significance. Missing scores are never filled with estimates. Publisher-estimated values, if present, are explicitly labeled.
The model comparison page also lets readers inspect complete launch-report tables. The same model can have different reported scores because of source dates, task versions, tools, effort, fallbacks or grading. The site preserves those differences instead of picking whichever number is highest. Personality percentiles describe style and do not define a capability ranking.
Research comparisons transcribe published aggregate values from Ulam’s benchmark pages. STEMBench’s audit snapshot is dated 14 October 2026; ErdősBench and SimoBench were retrieved on 25 October 2026. These snapshots are not a live feed. The source site may subsequently change.
Different agent workflows and compute budgets prevent a broad causal interpretation of the ordering. Filtering a cohort makes the comparison easier to inspect but does not prove every underlying setting is matched. Unpublished speed, token cost, uncertainty intervals, and model revisions remain unavailable rather than receiving estimated values.
Affiliation and editorial judgment
RLTasks is powered by Ulam. Ulam builds reasoning data and evaluation environments, and its evaluations are identified in the library. RLTasks’s interpretations are editorial analysis, not endorsements by the authors of external benchmarks. Inclusion does not imply a partnership with Artificial Analysis or any benchmark publisher.
Model colors identify the reporting model family. Score bars show the range within the selected source, while the original numeric scale stays visible. Per-model explanations use published breakdowns when available; aggregate-only results receive metric interpretation rather than invented diagnoses of individual failures.
Reading the evidence coverage
Every benchmark page includes a sourced result table. Research corpora and training-data releases belong in the research reference. A table can contain a historical cohort, a named subset, or reference baselines; those labels are part of the result. Chart readings carry an approximation mark and retain their source figure. Public research-solution credits are counts of documented outcomes, with no implied success rate or common attempt budget.
Review and corrections
Source checks and content updates are manually initiated. Public citation availability is checked separately from factual and protocol review; a successful HTTP request only establishes that the page is reachable. A changed scoring version, task set, or evaluation harness should produce an explicit revision rather than overwriting the meaning of an existing score. Report a correction through Ulam’s contact form with the benchmark URL and supporting source.