Technical Summary: GrowLoop – Self-Evolving Conversation Evaluation Seeded by Human

1. Problem Statement

The evaluation of Large Language Models (LLMs) in open-ended conversation is shifting from verifiable correctness (e.g., math, code) to human-likeness (naturalness, personality, empathy). However, evaluating these qualities faces three fundamental challenges rooted in tacit knowledge (knowledge humans possess intuitively but struggle to articulate):

1. Low Inter-Annotator Agreement: Unlike verifiable tasks, human judgments on human-likeness vary widely. While some cases yield consensus, others involve legitimate disagreement based on differing cultural backgrounds, communication styles, and personal preferences. Existing methods often force a single ground truth or blindly match annotation distributions, failing to distinguish between evaluation errors and legitimate divergent opinions.
2. Inability to Formalize Criteria: Tacit knowledge resists explicit formulation. Expert-authored benchmarks rely on intuition that is difficult to codify into stable rubrics, while Reward Models trained on preferences often act as black boxes, lacking interpretability and the ability to be debugged or repaired.
3. Static Benchmarks vs. Evolving Targets: As AI capabilities advance and human expectations shift, static benchmarks become obsolete. They fail to cover new failure modes or adapt to changing cultural and technological contexts. Current "self-evolving" approaches typically only scale case difficulty but do not evolve the evaluation criteria (rubrics) themselves.

2. Methodology: The GrowLoop Framework

GrowLoop is a self-evolving conversation evaluation system driven by minimal human seed annotations and model progress. It operates on two distinct regimes based on annotator agreement and utilizes a Rubric-Case dual-loop co-evolution mechanism.

2.1 Consensus-Divergence Aware Evaluation

The system partitions evaluation cases into two zones:

• Consensus Zone: Cases where human annotators agree. Here, the system is evaluated on human-AI agreement.
• Divergence Zone: Cases where annotators legitimately disagree. Here, the system is not required to match a specific ground truth but must produce a plausible judgment within the range of reasonable human opinions. This acknowledges that multiple valid judgments can exist simultaneously.

2.2 Rubric Generation (Externalizing Tacit Knowledge)

The system converts minimal human seed annotations (50 cases) into operational evaluation instruments through Heuristic Learning, a process treating the rubric as a free-form textual variable optimized by LLM agents. The process involves three phases:

1. Cold-Start Initialization: The agent analyzes seed annotations bottom-up, guided by structural priors from existing benchmarks (e.g., safety-quality separation, hierarchical dimensions). It extracts candidate dimensions and "fatal issues" (e.g., safety breaches) to bootstrap two complementary rubrics: Rubric_safety (binary fatal-issue detection) and Rubric_quality (multi-dimensional scoring).
2. Heuristic Learning: An iterative optimization loop refines the rubrics.
   • Zone-Aware Comparison: Optimization signals are derived only from the consensus zone to avoid chasing noise in ambiguous cases.
   • Natural-Language Gradients: The system performs meta-reflection to trace scoring errors back to structural causes in the rubric definition (e.g., ambiguous anchors), acting as a regularizer against overfitting.
   • Priority-Ordered Updates: Revisions follow a strict hierarchy (definitions > anchors > edge-case patches) with compression rules to prevent unbounded growth.
3. Cascaded Integration: The final evaluation function applies Rubric_safety first (zeroing out responses with fatal issues) before applying Rubric_quality.

2.3 Case Generation

The system generates a diverse, discriminative, and calibrated case set using a Conversation Situation Package (CSP), a 15-field structured template that controls variables like context, user persona, ambiguity patterns, and cognitive traps.

• Multi-Agent Generation: Four agents (Reference Analyzer, Diversity Agent, Case Generator, Critic) transform CSPs into multi-turn dialogues, ensuring coverage of all rubric dimensions and authentic interaction rhythms.
• Verification & Targeted Re-Generation: The generated set is evaluated against five hard gates (Diversity, Ranking Consistency, Discriminability, Adjacent Tier Gap, Difficulty Calibration). If gates fail, a targeted re-generation loop adjusts case difficulty or structure until all criteria are met.

2.4 Rubric-Case Co-Evolution

GrowLoop closes the gap between static rubrics and evolving models via a dual-loop mechanism:

• Rubric $\to$ Cases: The rubric drives case generation. If certain dimensions are under-represented or saturated by top-tier models, the system generates targeted cases to probe these gaps.
• Cases $\to$ Rubric: Evaluation results expose deficiencies in the rubric.
  • Anchor-level signals (low consistency) trigger refinement of existing dimensions via Heuristic Learning.
  • Structural signals (unattributed failure patterns) trigger human-in-the-loop injection of new seeds, which expands the rubric's scope to cover new failure modes.

3. Key Contributions

1. Unified Evaluation Paradigm: GrowLoop is the first system to simultaneously address consensus-divergence awareness, tacit knowledge externalization, and continuous self-evolution.
2. Heuristic Learning: A novel method for externalizing tacit knowledge into explicit, editable rubrics and cases from human seeds, leveraging LLM agents to iteratively refine criteria without full retraining.
3. Dual-Loop Co-Evolution: A mechanism where rubrics and cases iteratively drive each other's improvement, allowing the benchmark to autonomously detect blind spots and expand its scope when evaluation targets shift.
4. Living Infrastructure: The system shifts the benchmarking paradigm from manual updates or difficulty scaling to a comprehensive, continuous self-evolution process.

4. Experimental Results

The framework was validated on human-likeness evaluation in open-ended conversation using a 500-case benchmark and a four-tier model pool.

• Rubric Quality:
  • Consensus Zone: The optimized rubric achieved 86.0% agreement with human judgments on the consensus zone, significantly outperforming baselines.
  • Divergence Zone: The system demonstrated the ability to produce reasonable judgments in the absence of human consensus and, crucially, to identify role-boundary violations (e.g., medical advice) that human annotators overlooked.
  • Cross-Model Consistency: The rubric transferred reliably across different judge LLMs (Gemini and Claude), confirming its stability.
• Comparison with Baselines:
  • GrowLoop outperformed nine baselines (including zero-shot judges, manual prompts, training-free methods, and reward models) on Tie-aware Accuracy (0.78), Pair-Accuracy (0.87), and Spearman's $\rho$ (+0.78).
  • Unlike reward models, GrowLoop provided interpretable and editable criteria.
  • Unlike static benchmarks, it successfully identified cases where baselines failed due to implicit priors (e.g., equating verbosity with quality).
• Case Quality:
  • The generated 500-case set passed all five hard gates, demonstrating high diversity (entropy > 0.97), stable ranking consistency (Kendall $\bar{\tau} \ge 0.7$), and clear discriminability between model tiers (Cliff's $\delta \ge 0.32$).
  • The cases extended human seeds in depth and trap design, covering 100% of the trap taxonomy compared to the limited coverage of human-authored sets.
• Evolvability:
  • Rubric Evolution: The system converged on rubric quality (86.6%) and safety (91.5%) within 10 and 6 iterations, respectively, without additional human annotation. Updates generalized to unseen cases with similar structural patterns.
  • Case Evolution: The targeted re-generation loop converged autonomously, with metrics improving monotonically across five rounds.

5. Significance and Claims

The paper positions GrowLoop not merely as a benchmark, but as a living evaluation infrastructure. Its significance lies in:

• Addressing the Tacit Knowledge Gap: It provides a mechanism to externalize human intuition into explicit, debuggable criteria, solving the "we know more than we can tell" problem in subjective evaluation.
• Handling Legitimate Disagreement: By distinguishing between consensus and divergence, it avoids the arbitrariness of forcing a single ground truth on subjective tasks while maintaining rigorous standards.
• Continuous Adaptation: It solves the obsolescence problem of static benchmarks by enabling the evaluation criteria and test cases to co-evolve with model capabilities and shifting human expectations.
• Foundation for Non-Verifiable RL: The authors suggest that GrowLoop lays the necessary groundwork for applying Reinforcement Learning with Verifiable Rewards (RLVR) to non-verifiable tasks like conversation, where the evaluation criteria themselves must be interpretable and stable.

The authors conclude that while the current implementation is text-based, the framework is modality-agnostic and offers a path toward evaluating voice, visual, and embodied interactions as multimodal foundation models mature. They also identify the distillation of the multi-agent evaluator into a compact, editable Reward Model as a critical next step for scalable deployment.