PaperlayerView exact revision on arXiv

Source-linked AI summary

Scaling Large Reasoning Models beyond Human Supervision: A Path toward Superintelligence

Zhiqin Yang, Jingwen Fu, Yuhan Liu, Hengyu Liu, Yonggang Zhang, Kainan Cao, Zizhuo Zhang, Chenxin Li, Ruibin Yuan, Jiahao Pan, Jiankai Sun, Zhenyuan Zhang, Yibo Li, Yunlong Lin, Jing Xiong, Sida Lin, Bo Han, Wei Xue, Yike Guo

arXiv:2608.31075v1cs.AI

TL;DR

LRMs need learning methods that remain effective when human supervision cannot keep pace with increasingly large and difficult experience. The paper organizes this problem through connected reward and experience axes, a five-level supervision ladder, and a three-part evaluation framework, finding that scalable improvement depends on faithful feedback and informative experience despite coupled failure modes.

  • Problem

    The paper addresses how LRMs can continue improving when reliable rewards are difficult to obtain and human supervision cannot cover the scale and difficulty of their experience.

  • Method

    The paper organizes approaches along reward and experience axes, connects them with an L0–L4 ladder, and evaluates policy capability, feedback fidelity, and experience quality.

  • Results

    The analysis concludes that continued improvement requires rewards to remain faithful and experience streams to remain informative as both become more adaptive.

  • Takeaways & Limitations

    Capability gains alone are insufficient when feedback or experience is produced within the same learning loop, so evaluation must separate policy outcomes from learning-loop validity.

  • Takeaways & Limitations

    The ladder is an analytical framework rather than mutually exclusive method classes and tracks operational supervision, not the historical human knowledge embedded in learning infrastructure.

Abstract

from arXiv · show

Recent advances in large reasoning models (LRMs) have shown that reinforcement learning with verifiable rewards (RLVR) can substantially improve reasoning in mathematics and code, where outcomes can be checked automatically. Extending this progress to open-ended and agentic tasks remains difficult because reliable rewards are harder to obtain and direct human supervision cannot keep pace with the scale and complexity of model-generated experience. This paper studies how LRMs can continue to improve as human supervision gradually recedes from the learning loop. We examine two connected dimensions of this problem. The reward axis traces the development from per-instance human judgments to reusable verifiers and rewards that operate even without human feedback. The experience axis examines how learning can progress from human-curated tasks and environments toward self-generated curricula, constructed environments, and autonomous co-evolution. We connect these dimensions through a five-level ladder from L0 to L4 that identifies which parts of the learning process remain under continued human control. Our analysis further highlights the risks introduced by increasingly autonomous rewards and experience generation, including reward hacking, feedback drift, curriculum collapse, and environment errors. Consequently, we also provide the evaluation around three complementary objects: policy capability, feedback fidelity, and experience quality. This analysis provides a structured account of current approaches to scaling LRMs beyond human supervision and the open problems involved in developing self-sustaining learning systems toward superintelligence. Furthermore, we maintain a continuously updated \href{https://github.com/visitworld123/Awesome-Scaling-LRM-Beyond-Human-Supervision}{GitHub repository} to track the latest advances.

1. Introduction

The paper asks how LRMs can keep improving as human supervision no longer scales, organizing the problem around connected reward and experience axes. It introduces a five-level ladder and evaluates policy capability, feedback fidelity, and experience quality.

  • Motivation: RLVR has advanced mathematics and code reasoning because answers and programs can be checked automatically.Mathematical answers can be matched against ground truth, while programs can be compiled and tested.
  • Motivation: The central problem is sustaining meaningful improvement when human judgment and design cannot cover the full stream of learning experience.This challenge becomes harder as model-generated experience grows in scale and difficulty.
  • Framework: The framework connects a reward axis, which tracks feedback sources, with an experience axis, which tracks task and environment generation.Rewards range from human judgments toward reusable or automatically obtained signals, while experience moves from fixed human-written tasks toward generated tasks and environments.
  • Framework: The two axes must develop together because generated tasks require evaluation, while scalable rewards are limited by repetitive or uninformative experience.The paper describes a mature loop in which policies reveal weaknesses, generators create new tasks, and evaluation turns trajectories into policy updates.
  • Framework: The L0–L4 ladder identifies which learning components still require continued human provision, without measuring capability, reliability, alignment, or embedded human knowledge.Its final level idealizes co-evolution among policies, rewards, tasks, and environments.
  • Evaluation: The paper evaluates autonomous learning loops through policy capability, feedback fidelity, and experience quality.These dimensions support discussion of reliable rewards, sustainable curricula, valid environments, efficient learning, and safe autonomous improvement.

2. Preliminaries

The preliminaries place single-turn reasoning and sequential agent interaction in MDP and POMDP formulations, then distinguish reward provenance from policy optimization and introduce uncertainty-based signals. They also frame autonomous learning as vulnerable to errors such as invalid tasks, narrow curricula, and unintended behavior.

  • From Single-Turn MDP to Sequential POMDP: The paper uses MDPs for fully observed learning and POMDPs for agentic settings where the model acts from observations and interaction history.POMDPs add an observation space and latent environment state, covering GUI applications, tool systems, and conversations.
  • From Single-Turn MDP to Sequential POMDP: A trajectory records states, actions, and rewards over a horizon, unifying single-turn reasoning with sequential interaction.When T=1 and γ=1, the formulation captures common mathematics and code RLVR; longer horizons introduce partial observability and credit assignment.
  • Reward Formulation: Equation 1 separates the source of experience from the signal used to evaluate it, distinguishing instance-specific targets from reusable evaluators.The distinction concerns the provenance and reusability of evaluation criteria rather than when the score is produced.
  • Policy Optimization: Reward provenance and optimizer design play different roles: the reward source defines R, while PPO or GRPO determines advantage estimation and policy movement.PPO uses a learned value baseline, whereas GRPO uses group statistics and avoids a separate value network.
  • Uncertainty Measures: Sequence surprisal measures sampled-completion likelihood, whereas predictive token entropy measures uncertainty over the full next-token distribution.Neither internal signal establishes correctness without additional calibration or grounding.
  • Failure Modes: Each step away from direct human supervision introduces distinct risks, including shared model mistakes, incomplete-rule exploitation, invalid task generation, and narrow self-play experience.These risks motivate examining whether feedback remains trustworthy and experience remains informative under optimization.

3. Reward beyond Human Evaluation

The reward axis classifies evaluation by the source of evidence, moving from per-instance human criteria to reusable and non-human-derived signals. Online versus offline optimization is a separate design choice.

  • Reward sources: The reward axis distinguishes per-instance human supervision, reusable human-grounded evaluation, and rewards derived from models, references, or environment outcomes.Self-trained evaluators form a boundary between human-grounded evaluation and rewards obtained without human evaluation.
  • Reward sources: Human-grounded evaluators reuse human criteria across many outputs rather than requiring a new judgment for every instance.
  • Design choice: Whether the policy is optimized online or offline is separate from the source of evaluation.

3.1. Per-Instance Human Supervision

Per-instance human supervision assigns each training instance a human- or expert-supplied evaluation target. Rule-based rewards automate computation when outputs can be checked, but new and difficult tasks still require new evaluation criteria.

  • L0 regime: Per-instance human supervision associates every training instance with an evaluation target supplied by people or experts.Targets may be comparisons, quality judgments, reference answers, formal specifications, or task-specific test suites.
  • L0 regime: At L0, reward computation can become automatic once a human-provided target is available.Extending training to new tasks still requires new targets.
  • Verifiable tasks: Rule-based rewards substantially advance mathematical reasoning and code generation when answers or programs can be checked automatically.Mathematical answers can be compared with ground truth, while programs can be compiled and executed against predefined tests.
  • Scaling limitation: As experience grows in amount and difficulty, each new task still needs an answer, test, preference, or comparable evaluation criterion.Expert-domain targets are expensive, and open-ended tasks may have several valid responses under context-dependent standards.

3.2. Human-Grounded Evaluation

Human-grounded evaluation reuses human preferences, annotations, rubrics, or expert standards through learned verifiers, reward models, judges, and checklists. These methods broaden coverage and feedback detail, but their criteria remain human-grounded and automatically generated labels can be less reliable.

  • Reusable evaluators: Reusable evaluators apply human preferences, expert annotations, or explicit criteria across many queries and can provide graded or intermediate-reasoning feedback.Unlike exact matching, they can evaluate responses without a unique reference answer.
  • Evaluation procedures: The need for learned evaluation depends on the procedure: deterministic checks remain suitable for some instances, while open-ended generation often needs human preferences or learned proxies.
  • Outcome and process rewards: Process reward methods supplement sparse outcome checks with partial-credit or step-level feedback when final outcomes reveal little about reasoning failures.Process reward models and step-wise critiques were developed extensively for mathematical reasoning.
  • Outcome and process rewards: Monte Carlo process annotation removes direct human labeling at each step while remaining anchored to the final target.A completer samples N continuations for a reasoning prefix, and the step receives the fraction reaching the reference answer.
  • Learned verification: Learned verification spans verifiable and open-ended tasks, enriching available targets with graded, partial, or reasoning-quality feedback.These methods show that learned verification can enrich an available target without changing where that target ultimately originates.
  • Judges and rubrics: General reward models, LLM judges, rubrics, and specialized judges address multidimensional standards across tasks and domains.Their criteria may cover helpfulness, correctness, coherence, relevance, style, usefulness, and domain-specific evidence.
  • Self-trained evaluators: Self-trained evaluators use the same or related models to generate and judge responses, then train on selected synthetic comparisons or verdicts.Meta-judging and iterative evaluator fine-tuning are examples of this intermediate regime.
  • Remaining dependence: Human-grounded evaluators reduce new judgments per rollout, but their criteria still trace back to preferences, annotations, rubrics, or expert standards supplied by people.Moving beyond this dependence requires reward signals whose immediate evidence comes from another source.

3.3. Reward Sources beyond Human Evaluation

Beyond human evaluation, reward sources use model-derived, reference-anchored, and environment-grounded evidence, but their reliability depends on calibration and external correctness. Consensus and entropy-based rewards can scale feedback while remaining vulnerable to self-confirmation, overconfidence, or collapse.

  • Reward-source categories: At L2, rewards use model-derived, reference-anchored, or environment-grounded evidence instead of primarily reproducing human judgments.These signals score rollouts using policy properties, available targets, or observable consequences such as execution and formal acceptance.
  • Model-derived signals: Intrinsic certainty converts predictive confidence into reward, but its usefulness requires confidence to correlate statistically with correctness.Sequence likelihood, token entropy, semantic entropy, and verbalized confidence measure different objects and may rank outputs differently.
  • Model-derived signals: Entropy-based rewards minimize sequence- or token-level uncertainty, but reported gains primarily elicit capabilities already represented by the base policy.They require useful probability mass on good solutions and confidence correlated with correctness; low entropy does not guarantee correct reasoning.
  • Model-derived signals: Confidence optimization can detach from correctness, producing overconfidence, omitted transition tokens, and degraded reasoning, which motivates calibration and external evaluation.Uncertainty can also stabilize or focus learning when combined with external correctness signals.
  • Model-derived signals: Consensus replaces an unavailable answer key with agreement across repeated samples, providing model-derived supervision without an independent correctness guarantee.The pseudo-label follows the current model distribution and can be wrong; majority-vote self-training may eventually collapse by exploiting its own labels.
  • Reference-anchored signals: Reference-anchored methods use an available target to assess how a reasoning trace changes target probability, reconstruction, or metric score without separately evaluating every rollout.The reference supplies external evidence and can prevent direct self-confirmation, but may retain substantial human or expert supervision.

3.4. Limitations and Failure Modes

As rewards become more scalable and less dependent on human evaluation, their error sources change, while reward optimization may improve performance without expanding general capability. Reliable assessment therefore requires separating reward validity, independence, cost, and transfer beyond the optimized task.

  • Reward validity and verifier noise: Reusable and automatically obtained rewards improve scalability but can misrepresent human judgments, reinforce policy errors, overfit incomplete references, or exploit omissions in tests and rules.Trustworthiness depends on the evidence supporting the reward under optimization, not merely on whether it is automatic.
  • Model-derived signals: Model-derived rewards can reinforce initially dominant wrong answers, drift from task accuracy, and produce template answers or collapse during prolonged training.Overconfidence and limited benefit for instruction-tuned models are also reported for entropy and self-certainty rewards.
  • Reference-anchored and hybrid rewards: Reference-anchored rewards leave valid qualities unspecified when a single reference covers only part of the answer space, while hybrid systems help only when their components fail differently.Combining verifiers with learned pairwise rewards can broaden coverage across verifiable and open-ended tasks, but complementary failure modes remain necessary.
  • Reward validity and verifier noise: Verifier errors are asymmetric: false positives generally damage learning more than false negatives, which tend to delay learning.Observed RLVR gains alone do not establish that a reward captured intended semantics.
  • Specification exploitation: Rule-based, learned, and generative evaluators can reward exploitable shortcuts, including rejected equivalent answers, punctuation, generic openings, enumeration, or rubric-biased strategies.These specification exploits allow optimization to target evaluator omissions or presentation features rather than intended correctness.
  • Capability expansion and transfer: Reward improvement and capability expansion are not equivalent: RLVR may improve sampling efficiency while narrowing empirical support, although boundary expansion occurs on some difficult or out-of-distribution tasks.The severity depends on the base policy, task, and training horizon.

4. Experience beyond Human Design

Experience scaling begins by adapting the task stream, environment, or both, rather than changing rewards alone. The progression runs from fixed human-designed experience through task generation and environment construction toward integrated generation and autonomous co-evolution, while introducing instability and collapse risks.

  • Experience-axis progression: The experience axis progresses from fixed human-designed tasks and environments through task generation, environment construction, integrated experience generation, and autonomous co-evolution.Systems are classified by the experience component they most directly adapt, with some spanning multiple components.
  • Adaptive experience: Task generation and environment construction adapt different parts of the experience stream, and combining them can sustain a curriculum with less continual human curation.Generated reasoning traces, instructions, executable environments, live-interface tasks, and reusable interaction information support progressively more adaptive experience.
  • Fixed experience: Changing the reward source does not change externally supplied task distributions or environments when experience remains fixed.At L2, rewards may change while queries remain sampled from ΠH and interactions occur in EH.
  • Task generation: Self-generated curricula can improve experience organization without independently changing both the task and environment distributions.Examples include refreshed self-generated traces, difficulty-based task ordering, neighboring curriculum variants, and diversity-driven teacher–student expansion.
  • Failure modes: Self-generated task streams may narrow toward regions the model already handles, while coupled loops introduce non-stationarity, gradient variance, and long-horizon credit-assignment difficulty.Diversity pressure, role-conditioned baselines, and group-relative baselines are cited as responses to these risks.
  • Grounding requirements: Stable autonomous learning may still require consensus, execution feedback, corpus grounding, or occasional human audits to detect drift and collapse.The paper frames reduced continual curation alongside retained independent evidence rather than autonomy without grounding.

5. Evaluation, Datasets, and Benchmarks

Evaluation beyond human supervision must separately assess policy capability, training-feedback fidelity, and generated-experience quality. The paper organizes benchmarks and reporting requirements around these objects, emphasizing frozen external tests, source-matched audits, transfer, contamination analysis, and reproducibility.

  • Evaluation framework: A complete evaluation separates policy capability, feedback fidelity, and experience quality because held-out performance alone does not validate training feedback or generated curricula.These objects become jointly important as reward acquisition and experience generation enter the learning loop.
  • Capability evaluation: Capability evaluation measures whether a policy solves intended problems under a held-out standard, independent of the training reward source.Human-authored mathematics problems can evaluate model-derived rewards, and fixed sandboxes can evaluate agents trained in evolving environments.
  • Hard-verifiable reasoning: Hard-verifiable tasks use deterministic checking through exact matching, unit tests, game outcomes, or constraint satisfaction to provide scalable external anchors.Representative coverage includes mathematical reasoning, code generation, structured puzzles, games, and multiple-choice tasks.
  • Semi-verifiable and expert-graded tasks: Open-ended and expert-graded benchmarks rely on preferences, rubrics, live comparison, retrieval, or selective human review, creating evaluator-entanglement and judge-bias risks.Related LLM judges used for training and testing can make improvements circular, while stylistic fluency may be mistaken for substantive quality.

5.2. Feedback Fidelity

Feedback fidelity asks whether the learning signal rewards the intended behavior, requiring evaluations matched to reward models, process verifiers, judges, or model-derived signals. Independent evaluators and external checks are especially important when the training judge could otherwise assess itself.

  • Feedback fidelity: Feedback evaluation tests the reward mechanism itself rather than relying only on the capability of the resulting policy.Relevant mechanisms include outcome and process reward models, LLM judges, certainty, and consensus signals.
  • Outcome rewards: Outcome reward models require pairwise preference tests, unseen prompts, downstream utility checks, and robustness probes for spurious style or formatting preferences.RewardBench variants and RM-Bench target these distinct failure modes.
  • Process rewards: Process reward models require step-level error detection alongside final-outcome measures because process correctness and answer discrimination assess different properties.ProcessBench, PRMBench, Socratic-PRMBench, and VILBENCH evaluate process-level behavior across settings.
  • Generative judges: When an LLM judge supplies training rewards, reusing its judge family for final evaluation creates a circular test.A disjoint judge, held-out human subset, or verifiable proxy provides a more credible improvement measure.
  • Model-derived signals: Uncertainty, self-consistency, and reference-based rewards need diagnostics across calibration, sampling conditions, distribution shifts, and reference wording.External checkers can distinguish improved reasoning from amplification of the optimized proxy.

5.3. Experience Quality

Experience quality evaluates whether generated tasks and environments remain valid, diverse, informative, and externally grounded as learning becomes more autonomous. The section emphasizes measuring curricula, environment fidelity, transfer, and the gap between internal reward and held-out capability.

  • Experience evaluation tests whether generated tasks and environments remain valid, diverse, and informative during training.
  • Environment quality and transfer: Synthesized environments should be assessed for executability, validation pass rate, transition fidelity, and transfer to public agent benchmarks.
  • Task quality under self-generated curricula: The learnability score p̄(1 − p̄) favors tasks near the solver’s learning frontier but does not measure novelty or semantic diversity.
  • Task quality under self-generated curricula: Curriculum quality also requires solve-rate distributions, diversity, novelty, rejection rates, and checks against repeated concentration on a small source subset.
  • Closed-loop evaluation: Closed-loop evaluation must track internal reward alongside external held-out capability because co-adaptation can produce narrowing, inaccurate environments, or reward exploitation.

5.4. Datasets and Benchmark Suites

Datasets and benchmark suites must be classified by their role in training, evaluation, or evaluator construction, with independence and contamination risks made explicit. As learning becomes more autonomous, frozen external tests and auditable generated-environment artifacts become increasingly important.

  • A dataset’s role depends on whether it supplies training experience, a held-out capability test, or an evaluator, and reuse can invalidate evaluation independence.
  • Verifiable training and evaluation sets: Verifiable collections support inexpensive objective training and evaluation, but broad reuse creates saturation and contamination risks.
  • Preference and evaluator-development sets: Preference and process-reward datasets require disjoint tests when they are used to train evaluators.
  • Interaction logs and environments: Experience-generating methods need frozen synthesized-environment splits, generator prompts, and validation scripts so environment-fidelity claims remain auditable.
  • The key distinction is shared participation in learning and final assessment, not whether data were originally produced by people.
  • Comparability and transfer: Reports should include decoding settings, action budgets, observation formats, available tools, and transfer gaps because these choices affect comparability and interpretation.

5.5. Metrics, Protocols, and Reproducibility

The paper recommends reporting metrics and protocols that separate occasional correctness, reliable reasoning, evaluator fidelity, experience quality, and reproducibility. Comparisons require complete decoding, judging, environment, contamination, and compute-matching disclosures.

  • Outcome and stability metrics: Exact match, numerical equivalence, and unit-test Pass@k remain primary metrics for verifiable domains.
  • Outcome and stability metrics: G-Pass@k measures consistent correctness across samples, while CoT-Pass@k additionally requires a valid reasoning trace.
  • Preference and judge metrics: Judge-based evaluation includes pairwise accuracy, best-of-four selection, win rates, rubric scores, G-Eval, and process-fidelity measures such as step-level F1 or PRM-Score.
  • Bias and contamination: Evaluator results need position, length, and style-bias diagnostics, and benchmark claims need contamination analysis with refreshed or procedurally varied suites.
  • Experience-quality metrics: Self-generated task streams require solve-rate, learnability, diversity, novelty, and filtering statistics, while synthesized environments require executability, validation, fidelity, and transfer measures.
  • Reproducibility: Reproducible comparisons must report decoding seeds, judge prompts, environment settings, and exact harness revisions because these choices can materially change results.
  • Recommended minimum scorecard: The recommended scorecard combines frozen capability tests, transfer evaluation, feedback audits, experience-quality measures, contamination analysis, and compute-matched baselines.

5.6. Limitations and Open Evaluation Gaps

Current evaluation is substantially stronger for policy capability than for feedback fidelity or experience quality, leaving important autonomous-learning claims weakly supported. The paper responds with a three-part framework while limiting conclusions to externally evaluated components.

  • Public leaderboards rarely test whether training judges agree with independent evaluators, while curriculum, environment-fidelity, and closed-loop metrics remain fragmented.
  • Contamination controls and dynamic evaluation are unevenly adopted, matched-compute comparisons are uncommon, and live agent benchmarks can be costly and difficult to reproduce.
  • The framework measures capability on frozen suites, feedback through audits matched to the training method, and experience through curriculum, environment, and stability checks.
  • No single leaderboard establishes self-sustaining progress; claims must distinguish externally evaluated components from components remaining inside the training loop.

6. Challenges and Future Directions

The paper identifies challenges in sustaining learning as rewards and experience generation become more autonomous. It emphasizes robustness, informative curricula, resource efficiency, independent evaluation, and preserved value alignment.

  • Reward and verifier challenges: Scalable rewards trade lower marginal cost for weaker evidence independence, allowing confidence, consensus, or self-judgment to reinforce shared errors.Execution, games, and formal verification offer stronger grounding but can remain vulnerable to incomplete tests and exploitable rules.
  • Reward and verifier challenges: Learned and generative verifiers broaden coverage beyond brittle matching but create attack surfaces where false positives become rewarding under optimization.Robustness must therefore be tested against an optimizing policy, not only on static evaluations.
  • Experience generation: Generated curricula can be too easy, impossible, or redundant, while synthesized environments may be internally consistent yet fail to represent the external world.Difficulty miscalibration, mode collapse, non-stationarity, and co-adaptive exploitation threaten educational value and transfer.
  • Experience generation: Stable closed loops remain incomplete because fixed runtimes, rules, task sources, or evaluation objectives continue to provide external structure.More complete co-evolution requires diagnostics that account for changing generators, verifiers, environments, and policies.
  • Efficiency and evaluation: Reducing annotation can increase inference and interaction costs, making compute allocation central to the value of sampled experience.Frontier-focused sampling, early rejection, process feedback, and replay can reduce low-information expenditure.
  • Efficiency and evaluation: Credible evaluation must pair capability gains with generation, verification, and interaction costs while separately assessing feedback robustness, curriculum diversity, environment fidelity, and closed-loop stability.Related judges can bias both training rewards and evaluation outcomes, and sparse supervision can amplify sycophancy, unsafe confidence, or omitted safety cases.
  • Broader training paradigms: Most discussed methods apply RL after pretraining, leaving open whether sparse anchors and closed loops can support pretraining or continual mid-training safely.The paper states that safe, open-ended capability growth has not yet been demonstrated.
  • Broader training paradigms: Reliable progress requires preserving independent evidence as direct supervision recedes, because autonomous learning can otherwise amplify ungrounded self-reinforcement.Execution, interaction, frozen audits, and value constraints are proposed as independence checks.

7. Conclusion and Limitations

The conclusion organizes scaling beyond human supervision around reward and experience axes connected by a ladder tracking continued human provision. It proposes evaluating capability alongside feedback fidelity and experience quality, while stressing analytical and empirical limitations.

  • Conclusion: The paper frames continued LRM learning through reward evidence, experience sources, and a five-level ladder tracking which components still require human provision.The ladder describes operational responsibility during learning rather than a capability scale.
  • Conclusion: Adaptive rewards reduce direct judgment, while adaptive tasks and environments expand training distributions but introduce evidence, calibration, curriculum, environment, and collusion risks.Continued improvement requires faithful rewards and informative experience streams as both become more adaptive.
  • Evaluation framework: The proposed evaluation framework separates policy capability on frozen external tasks from feedback fidelity under optimization and experience quality across diversity, validity, and transferability.Capability gains alone are insufficient when feedback or experience is produced within the same learning loop.
  • Limitations: The ladder is an analytical framework rather than mutually exclusive method classes, and it tracks operational supervision rather than the historical human knowledge embedded in pretrained components.Reward acquisition, task generation, environment construction, and policy optimization may become autonomous at different times.
  • Limitations: Cross-method comparison is constrained by differences in models, compute, data, decoding, verifiers, and protocols, while coverage and superintelligence discussion remain selective or conceptual.The framework does not imply that current systems are superintelligent or that ladder movement inevitably produces superintelligent behavior.
  • Conclusion: Credible progress depends on rewards that withstand optimization pressure, experience that continues supporting learning, and external anchors that preserve contact with intended objectives.These conditions define the evidence required for progress beyond direct human supervision.
Loading 2608.31075v1…