WikifitaGitHub live67e8de5
outro · unit-distance/unit-distance-grpo-inference

GRPO Tournament Inference: Evolutionary Hypothesis Selection

GRPO (Group Relative Policy Optimization) applied to tournament inference: how hypotheses compete, evolve, and converge in the Co-Scientist multi-agent architecture.

Baixar raw

GRPO Tournament Inference

GRPO (Group Relative Policy Optimization) Tournament Inference is the mechanism by which hypotheses are generated, debated, ranked, and evolved in the unit-distance research project. It combines two distinct concepts:

  1. GRPO -- a reinforcement learning algorithm that optimizes model behavior by comparing outputs within a group (relative ranking) rather than against a fixed reward signal
  2. Tournament Inference -- the application of tournament-style competition to select the best reasoning path at inference time (test-time compute)

Together, they form the engine of the Co-Scientist architecture's hypothesis evolution system, as described in the DeepMind Nature 2026 paper and implemented in the Antigravity 2.0 harness.


What Is GRPO?

Group Relative Policy Optimization

GRPO is a reinforcement learning algorithm that differs from traditional RLHF (Reinforcement Learning from Human Feedback) in a fundamental way:

PropertyTraditional RLHFGRPO
Comparison basisFixed reward modelRelative ranking within a group
Training signalAbsolute scoreRelative preference
Group dynamicsIndependent samplesCompetitive selection
Optimization targetMaximize rewardWin against peers

In GRPO, the model generates a group of candidate outputs, ranks them relative to each other, and uses the ranking to update the policy. The "group" is the key: instead of asking "is this output good?", GRPO asks "is this output better than the others?".

Why Relative Ranking Works Better

  1. Robustness to reward hacking. A fixed reward model can be gamed. Relative ranking within a group is harder to exploit because the "target" moves with each group.

  2. Natural exploration. By generating multiple candidates, the model explores the solution space more broadly than single-sample generation.

  3. Implicit curriculum. As the model improves, the groups it generates become more competitive, creating a natural difficulty progression.


Tournament Inference

What It Is

Tournament inference applies the tournament selection mechanism from evolutionary algorithms to the inference process. Instead of generating a single answer, the model:

  1. Generates multiple candidates (the tournament pool)
  2. Evaluates them against each other (the matches)
  3. Selects the winner (the champion)
  4. Optionally evolves the winner (mutation/refinement)

This is the "tournament and evolution" process described in the DeepMind Co-Scientist paper as one of the system's principal contributions.

The Co-Scientist Agent Roles

The tournament maps directly to the agent roles in the Co-Scientist architecture:

Tournament PhaseAgent RoleFunction
Pool generationGeneration AgentProposes initial hypotheses
Diversity checkProximity AgentMaps and clusters hypotheses for diverse exploration
Match playReflection AgentEvaluates hypotheses against each other
SeedingRanking AgentOrganizes the tournament bracket
Champion refinementEvolution AgentRefines and improves the winning hypothesis
Tournament oversightMeta-review AgentSynthesizes insights to optimize the system

Test-Time Compute Scaling

A key finding from the Co-Scientist paper: hypotheses improve with increased computation dedicated to the task (test-time compute). More tournament rounds, more candidates per round, and more evolution cycles all produce better results -- up to the diminishing returns ceiling.

This is directly analogous to how chess engines improve with more search depth. The tournament is the search mechanism; GRPO is the evaluation function.


The Elo System

How Elo Maps to Tournament Inference

The unit-distance project used an Elo rating system (borrowed from chess) to calibrate the quality of hypotheses. This is a natural fit for tournament inference because:

  1. Elo is relative. Like GRPO, Elo ranks participants against each other, not against an absolute scale.
  2. Elo is dynamic. A hypothesis's rating changes as it "plays" against other hypotheses.
  3. Elo is interpretable. The rating maps to a meaningful quality level.

The Elo Calibration Scale

Elo RangeApproximate MeaningUnit-Distance Examples
~1800Existential proof (barely above trivial)OpenAI's original proof (delta ~10^-38)
2150First valid constructionH7: Full Multivariate
2200Matches human SOTA (Sawin)H15: Central CM Tower (delta = 0.014)
2350Surpasses human SOTAH16: Multi-Quadratic CM16 (delta = 0.0196)
2500+Significant improvement over known boundsTarget for Season 3
2700+Human expert levelCurrent human SOTA (delta > 0.036)

Tournament Match Structure

Each "match" between hypotheses evaluates:

  1. Mathematical correctness. Is the proof valid? Does the construction satisfy the Golod-Shafarevich inequality?
  2. Novelty. Does this approach explore a new region of the solution space?
  3. Reproducibility. Can the result be verified independently?
  4. Elegance. (Subjective but important) Is the construction clean and minimal?

The winner of each match advances; the loser is either eliminated or sent to the EVOLVE state for refinement.


The Evolutionary Selection Mechanism

Generate-Debate-Evolve Loop

The full tournament inference cycle implements an evolutionary algorithm:

graph TD
    A[INIT: Parse research question] --> B[GENERATE: Produce 2-4 hypotheses]
    B --> C[DEBATE: Critique each hypothesis]
    C --> D[RANK: Elo tournament selection]
    D --> E{Winner mature enough?}
    E -->|Yes| F[RESPOND: Output result]
    E -->|No| G[EVOLVE: Refine winner]
    G --> H{Iteration < 5?}
    H -->|Yes| I[LOOP: Increment counter]
    I --> B
    H -->|No| J[HALT or RESPOND]
    G --> K[DELEGATE: Sub-agent verification]
    K --> L[SYNTHESIZE: Merge findings]
    L --> E

Mutation Operators

In evolutionary algorithm terms, the EVOLVE state applies mutation operators to the winning hypothesis:

OperatorDescriptionExample
Parameter tuningAdjust numerical parametersChange the polydisc radius in H2
Structural extensionAdd new componentsAdd more primes to the class field tower
Dimensional shiftChange the base field degreeMove from degree 8 to degree 16 (H16)
Constraint relaxationRemove unnecessary restrictionsAllow non-abelian extensions
Hybrid mergingCombine elements from multiple hypothesesMerge the prime selection from H4 with the base field from H8

Selection Pressure

The tournament creates selection pressure by:

  • Eliminating weak hypotheses (they lose matches)
  • Rewarding strong hypotheses (they accumulate Elo)
  • Forcing evolution (the EVOLVE state refines the winner)
  • Capping iteration (the 5-loop maximum prevents infinite refinement)

This selection pressure drives the system toward increasingly better solutions. In the unit-distance project, it drove the system from the existential proof (delta ~10^-38) to H16 (delta = 0.0196) across 16 hypotheses.


Connection to Multi-Agent Methodology

The Swarm Model

The unit-distance project used a cognitive swarm model with specialized roles:

Swarm RoleTournament FunctionAntigravity Implementation
Research DirectorTournament organizerMain agent (Gemini 3.1 Pro High)
ELO-RANKERMatch judgeSpecialized sub-agent with web access
CITATION-VERIFIERFact-checkerRead-only sub-agent
PROVENANCE-TRACKERAudit trailPostToolUse hook or sub-agent
WIKI-MAINTAINERKnowledge base syncuv run skill script

Delegation and Synthesis

The DELEGATE state allows the Research Director to offload specific tournament functions to specialized sub-agents:

  1. Verification delegation: A sub-agent checks whether a hypothesis's mathematical claims are valid, without the main agent losing context.
  2. Parallel exploration: Multiple sub-agents explore different evolutionary branches simultaneously.
  3. External calibration: The ELO-RANKER sub-agent provides objective rating by comparing against external benchmarks (Sawin's bound, OpenAI's proof).

After delegation, the SYNTHESIZE state integrates sub-agent findings back into the main reasoning thread.

Why Multi-Agent Matters

Single-agent tournament inference is limited by:

  • Context window: One model cannot hold all hypotheses, all critiques, and all historical context simultaneously
  • Bias: A single model may favor its own earlier outputs
  • Specialization: Different tasks (verification, ranking, evolution) benefit from different prompting strategies

The multi-agent swarm distributes these functions across specialized sub-agents, each operating with its own context and constraints.


Alefita's Proposed Extension: Steering Vector Optimization

The DeepSeek transcript discussed Alefita's proposal to use GRPO tournament inference for persona steering vector optimization:

The Protocol

  1. Prefill a steering vector in the first block of the CoT
  2. Map a base list of desired behavioral traits (curiosity, skepticism, precision, etc.)
  3. Run 5 cycles of GRPO with Elo tournament to select the best steering configuration
  4. Apply the winning steering to subsequent reasoning

Why This Works

  • GRPO naturally optimizes for relative quality. The "best" steering vector is the one that produces the best reasoning relative to alternatives.
  • Tournament selection prevents local optima. By maintaining a diverse pool of steering configurations, the system avoids converging on a single (potentially suboptimal) configuration.
  • The 5-loop cap matches the channel protocol. The same iteration limit that governs hypothesis evolution also governs steering optimization.

The Testability Claim

The DeepSeek transcript noted: "Your theory about prefill of steering with GRPO and Elo tournament is testable." This is a concrete, implementable protocol that could be evaluated empirically.


Historical Evolution in the Unit-Distance Project

The Hypothesis Tournament

The unit-distance project itself was a tournament of 16 hypotheses:

HypothesisEloStatusKey Innovation
H1: Valuation Optimization--ValidatedAuthentic OpenAI derivation
H2: Continuous Polydisc Radius--ValidatedAuthentic OpenAI derivation
H3: Pro-2 Class Towers--DisqualifiedExternal contamination
H4: Higher Valuation Powers--Validated (below benchmark)Pro-3 towers
H5-H6: Global Optimization--Validated (below benchmark)Combined optimization
H7: Full Multivariate~2150ValidatedFull multivariate approach
H8: Imaginary Quadratic Base~2650ValidatedImaginary quadratic foundation
H9: Multi-Quadratic CM--DisqualifiedCompleteness failure
H10: Galois Symmetry--DisqualifiedCompleteness failure
H11: Imaginary Quadratic 2-Tower1500ProvenFirst formal proof
H12-H14--DisqualifiedCompleteness failures
H15: Central CM Tower2200ProvenMatches Sawin bound
H16: Multi-Quadratic CM162350ProvenSurpasses Sawin

The Tournament Arc

The progression from H1 to H16 demonstrates the evolutionary selection mechanism in action:

  1. H1-H2: Establish the base (OpenAI derivation)
  2. H3: First disqualification (contamination detection -- the anti-contamination protocol worked)
  3. H4-H6: Incremental improvements (below benchmark but valid)
  4. H7-H8: Major jumps (entering competitive range)
  5. H9-H10: Failed experiments (necessary for completeness)
  6. H11: First formal proof (milestone but low delta)
  7. H12-H14: Rapid iteration (three failures in quick succession)
  8. H15: Breakthrough (matches Sawin)
  9. H16: Surpasses Sawin (the tournament champion)

GRPO vs. Other Optimization Methods

MethodComparison BasisExplorationAdaptationUsed In
GRPORelative within groupHigh (multiple candidates)Fast (group dynamics)Co-Scientist, unit-distance
RLHFFixed reward modelLow (single sample)Slow (reward model is static)Standard LLM training
PPOAdvantage estimationMedium (clipped updates)MediumGeneral RL
Evo-SearchPopulation-basedHigh (mutation/recombination)Medium (no policy gradient)Genetic algorithms

GRPO's advantage for multi-agent research: it combines the exploration of evolutionary search with the optimization efficiency of policy gradient methods, all operating at inference time rather than training time.


Cross-References


This page documents the theoretical framework connecting GRPO, tournament inference, and the multi-agent research methodology. The core concepts (GRPO, Elo, evolutionary selection) are established in the literature. The application to persona steering vector optimization is a proposal discussed in the DeepSeek transcript and awaits empirical validation.