Pairwise-Experience-Evolutionary Meta-Prompting

• The paper advances meta-prompting by integrating evolutionary algorithms with pairwise evaluations to yield adaptive reasoning improvements.
• PE-EMP employs tournament selection, crossover, and mutation to refine reasoning strategies in tasks such as LLM reasoning and multimodal segmentation.
• This approach circumvents static prompting limitations by dynamically evolving prompts that enhance inductive bias, self-correction, and compositionality.

Pairwise-Experience-Evolutionary Meta-Prompting (PE-EMP) is an inference-time optimization paradigm designed to evolve prompt-based agent reasoning through online, pairwise, competition-driven meta-prompt adaptation. Rather than relying on static prompts or pre-trained, fixed policies, PE-EMP injects evolutionary search and continual adaptation at the meta-prompt level, leveraging pairwise evaluation, self-critique, and evolutionary operators to select, modify, and propagate high-fitness reasoning strategies. This mechanism enables agents to dynamically accumulate and refine inductive biases, driving stepwise improvements in task performance, compositionality, and self-correction across diverse domains, including symbolic reasoning and multimodal segmentation (Lu et al., 4 Feb 2026, Ye et al., 31 Dec 2025).

1. Conceptual Overview and Motivation

PE-EMP functions as a reflexive optimizer within various agentic frameworks (e.g., Empirical-MCTS for LLM reasoning, EVOL-SAM3 for zero-shot segmentation). Its core insight is to turn each agent decision into a micro-evolutionary experiment, where multiple candidate meta-prompts are generated, paired, and competed. At every decision point, the system:

• Samples or generates a population of candidate prompts.
• Evaluates each by producing agent responses and critiquing them (e.g., via SPCT-style self-principled critique or multimodal scoring).
• Assigns fitness via pairwise tournament-based feedback, often formalized through Bradley–Terry or Elo-like models.
• Applies evolutionary operators (selection, crossover, mutation) to propagate successful traits and introduce diversity.
• Distills critical insights or principles from successful prompts as explicit memory blocks for reuse or global policy adaptation.

Intuitively, PE-EMP instantiates a “prompt-genotype” evolutionary search loop within the agent’s reasoning process, fostering prompt-level meta-learning that steers agents toward context-optimized, high-fidelity policies (Lu et al., 4 Feb 2026, Ye et al., 31 Dec 2025).

2. Mathematical Formulation

PE-EMP formalizes prompt evolution through a set of adaptive, pairwise-evaluated operators:

• Let $\mathcal{P}_i$ denote the $i$th prompt in population $\{\mathcal{P}_1,\ldots,\mathcal{P}_N\}$.
• Each prompt produces a response $r_i$ under current task context; scoring is achieved using adaptive criteria $\{c_j\}_{j=1}^m$ with weights $\{w_j\}$ ($\sum_j w_j=1$):

$S_i = \sum_{j=1}^m w_j\, \mathrm{score}_j(r_i)$

• Pairwise win probabilities are obtained via the Bradley–Terry model:

$$p_{\mathrm{win}(i \text{ beats } j)} = \frac{e^{S_i}}{e^{S_i} + e^{S_j}}$$

• Prompt fitness is defined as mean win-rate:

$$F(\mathcal{P}_i) = \frac{1}{N-1}\sum_{j\ne i} \frac{e^{S_i}}{e^{S_i} + e^{S_j}}$$

with complexity-regularized extension $$\tilde F(\mathcal{P}_i) = F(\mathcal{P}_i) - \lambda C(\mathcal{P}_i)$$ for complexity metric $C$ and penalty $\lambda$.

• Evolutionary cycling includes:
  • Tournament selection of parents by fitness.
  • Crossover: Interleaving “self-principle” clauses from parent prompts to synthesize offspring.
  • Mutation: Stochastic alteration (e.g., weighting criteria, clause re-phrasing) at low rate $\mu$.
• Low-fitness prompts are eliminated; newly generated high-fitness prompts are inserted, enforcing continual population improvement (Lu et al., 4 Feb 2026).

A comparable formulation is adopted in EVOL-SAM3 for segmentation, where prompt fitness $\mathcal{F}(z)$ results from Visual Arena tournaments and semantic mutation is guided by language/vision model proposals (Ye et al., 31 Dec 2025).

3. Algorithmic Realizations and Pseudocode

Empirical-MCTS Internal Loop

PE-EMP is integrated into Monte Carlo Tree Search as an evolutionary subloop over meta-prompts:

def PE_EMP_Expansion(q, E_prior, P_pop): parent1 = TournamentSelect(P_pop, fitness=F) parent2 = TournamentSelect(P_pop, fitness=F) r1 = LLM_Generate(q, prompt=parent1, context=E_prior) r2 = LLM_Generate(q, prompt=parent2, context=E_prior) (S1, S2), E_new = SelfPrincipledCritique(r1, r2, E_prior) p_win = exp(S2)/(exp(S1)+exp(S2)) delta = int(S2 > S1) child = Crossover(parent1, parent2) if Random() < mu: child = Mutate(child) if fitness(child) > min(fitness(P) for P in P_pop): replace_worst(P_pop, child) return (S2, S1), E_new, child

This expansion embeds the evolutionary prompt update within each node-expansion, returning reward vectors, distilled insights, and evolved prompts for subsequent rollouts (Lu et al., 4 Feb 2026).

Visual Arena and Semantic Mutation (EVOL-SAM3)

EVOL-SAM3 employs a population-based “Generate–Evaluate–Evolve” loop. Algorithmic highlights include:

• Filtering by segmentation confidence and semantic gates.
• Elo-like fitness scoring from pairwise mask comparison.
• Semantic mutation operator for diversity and self-correction.
• Optional geometric-branch adjudication for hallucination guardrails.

Pairwise-tournament updates are formalized as: $$\mathcal F(z_a) \leftarrow \mathcal F(z_a) + \Delta,\quad \mathcal F(z_b) \leftarrow \mathcal F(z_b) - \Delta$$ for a winner $z_a$, with challenger upweighting when $z_b$ wins (Ye et al., 31 Dec 2025).

4. Mechanisms for Real-Time Prompt Adaptation

The evolutionary cycle within PE-EMP is characterized by the rapid assimilation of successful prompt traits and systematic pruning or mutation of ineffective strategies. Each winning prompt’s distilled “self-principles” are:

• Reinjected into the prompt population for future rollouts.
• Encoded as modular insights or rules to global or long-term memory agents when present (as in Empirical-MCTS).

Prompt mutation maintains exploration and mitigates premature convergence or stagnation, ensuring adaptation as task requirements shift. In multimodal tasks, semantic mutation systematically refines ambiguity and improves alignment between text prompts and spatial/geometric priors (Lu et al., 4 Feb 2026, Ye et al., 31 Dec 2025).

5. Application Domains and Empirical Findings

LLM Reasoning and Mathematical Tasks

Empirical-MCTS demonstrates substantial improvements on complex reasoning benchmarks such as AIME25, ARC-AGI-2, and MathArena Apex by coupling MCTS with PE-EMP compared to traditional stateless search or purely memory-driven baselines. The evolved meta-prompts rapidly converge towards high-utility strategies incorporating problem-specific knowledge, verified proofs, and error-checking clauses (Lu et al., 4 Feb 2026).

Zero-Shot Reasoning Segmentation

EVOL-SAM3 applies PE-EMP to prompt search for pixel-level segmentation guided by complex linguistic queries. On the ReasonSeg benchmark, EVOL-SAM3 achieves 72.5 gIoU / 67.4 cIoU test performance, surpassing fully supervised SFT/RL methods (e.g., LISA-13B at 65.0 gIoU) and static prompt agents (Ye et al., 31 Dec 2025). Key effects include recovery from language-model hallucinations, correction of ambiguous referents, and robust performance on low-saliency or compositional concepts. Ablation studies demonstrate that PE-EMP’s evolutionary loop yields monotonic gains with additional generations.

Empirical Results Table

Model/Method	Val gIoU / cIoU	Test gIoU / cIoU	Notable Setting/Comment
SAM 3 Agent (7B)	62.2 / 49.1	—	Static zero-shot baseline
LISA-13B SFT/RL	65.0 / —	—	Supervised SOTA
EVOL-SAM3 (ours)	70.7 / 63.4	72.5 / 67.4	Qwen2.5-VL 7B, PE-EMP, $T=2$

6. Illustrative Example

In Empirical-MCTS, a classic summation problem illustrates PE-EMP’s adaptation:

• Initial prompts: $\mathcal{P}_A=$“Think step-by-step.”, $\mathcal{P}_B=$“Focus on algebraic manipulation.”
• After successive pairwise competitions and evolutionary updates, the prompt population evolves toward $\mathcal{P}_E=$“Compute $n(n+1)/2$; then verify by induction & boundary checks.” This progression demonstrates that PE-EMP can synthesize, within a handful of inner rollouts, sophisticated meta-prompts robust to subtle indexing errors, integrating higher-level mathematical reasoning and verification strategies beyond the initial prompt set (Lu et al., 4 Feb 2026).

In EVOL-SAM3, PE-EMP systematically corrects segmentation failures (e.g., mis-labeling oars as “boat,” grounding abstract regions like “activity area”) by evolving prompt hypotheses, guided by pairwise visual-linguistic scoring and semantic/geometric priors (Ye et al., 31 Dec 2025).

7. Significance and Implications

PE-EMP represents a non-parametric, inference-time alternative to gradient-based meta-learning, providing task-adaptive optimization exclusively at the meta-prompt level. This approach bypasses issues endemic to SFT (catastrophic forgetting, domain overfitting) and RL (reward design, instability) by leveraging direct experience accumulation and pairwise evolution during agent reasoning. The generic nature of PE-EMP’s evolutionary prompt adaptation loop underpins its applicability across symbolic LLM problem-solving, zero-shot multimodal understanding, and likely broader domains where dynamic adaptation without retraining is critical. A plausible implication is that PE-EMP’s integration with structured search (e.g., MCTS, arena tournaments) is essential for scaling agent capabilities on open-ended, compositional tasks beyond what static prompting or pre-trained policy priors can achieve (Lu et al., 4 Feb 2026, Ye et al., 31 Dec 2025).