← All AI/ML open problems

Statistical learning theory · Learning dynamics & scaling / Open question

Predicting AI Beyond Measured Scale

A compute-allocation recipe can be misleading if its uncertainty omits curve-family ambiguity or the adaptive way pilots were chosen.

OpenEditorially scopedFramework needs definitionCompiled 2026-11-19

Seed catalog · Showing the bundled record.

Problem statement

For a finite set of RL recipes, model expected success as a logistic curve in log compute with bounded intercept, slope and ceiling, plus a discrepancy bounded by eta on a stated pilot-and-extrapolation interval. Observations have conditionally sub-Gaussian noise and are selected adaptively. Determine minimax prediction intervals and pilot designs that distinguish slope/efficiency changes from ceiling changes; quantify how the interval width grows with extrapolation distance and discrepancy, including indistinguishable-curve lower bounds.

Assumptions and scope

  • Declare the observable metric and fitted/structural curve family; reward, error rate, negative log-success and Elo-like scores are not interchangeable.
  • Include training-seed variation and reuse of evaluation tasks. A finite-range fit does not identify arbitrary behavior at infinite compute.
  • Any extrapolation guarantee must name a bounded range and assumptions excluding indistinguishable alternatives; absent those assumptions, seek an impossibility bound.

What would settle it?

A specified curve/noise model with valid adaptive confidence or prediction bounds and an informative design rule, paired with identifiability lower bounds. Re-fitting a sigmoid and a power law is insufficient.

Selected for a theoretical or computational target. A fitted scaling curve or improved benchmark score alone does not establish the stated learning-theoretic guarantee.

Status and evidence

ScaleRL, Tan et al. and actor-critic scaling studies demonstrate useful empirical fits with different metrics/regimes. The residual target is calibrated adaptive extrapolation and its limits, not choosing one universal fitting function. This is an editorial, source-motivated research target. The preliminary review through 2026-11-19 does not certify novelty or exclude an unlocated resolution.

Preliminary source check; not an expert-verified open status or an independent audit of proposed proofs. The compilation date is not a literature-review cutoff.

Sources and further reading

  1. Khatri et al. — The Art of Scaling Reinforcement Learning Compute for LLMs ↗empirical motivation · source checked 2026-11-19

    Abstract checked: empirical sigmoidal fits, recipe-dependent efficiency/ceilings and held-out-scale prediction; not a universal asymptotic theorem.

  2. Tan et al. — Scaling Behaviors of LLM Reinforcement Learning Post-Training ↗empirical motivation · source checked 2026-11-19

    ACL 2026 publisher abstract checked: Qwen2.5 scaling, model-size efficiency saturation and data reuse in the investigated regimes.

  3. Zheng et al. — An Empirical Study of Reinforcement Learning Scaling Laws ↗empirical motivation · source checked 2026-11-19

    Abstract checked: actor-critic experiments in Starpilot/Craftax, Elo-like metric and nonmonotone compute-optimal model size. This is not an LLM post-training theorem.

  4. He et al. — How Data Scales in Agentic Reinforcement Learning ↗empirical motivation · source checked 2026-11-19

    Full-text Sections 3–6 checked: distinguish worlds, environment structures, tasks, trajectories and reward density. Exponents and synthesis discounts are empirical, regime-specific fits; allocation assumes a stated cost model.