Lark: Biologically Inspired Neuroevolution for Multi-Stakeholder LLM Agents

AI summary

Overview

• Research area: Multi-agent systems (MAS), LLM-driven reasoning, and biologically inspired neuroevolution for multi-stakeholder decision-making.
• Technical level: Intermediate. The paper uses formal notation (Borda scoring, logistic duplication probability, token-penalty equations) but the underlying ideas are explained in accessible terms.
• Scope: A proof-of-concept workshop paper introducing Lark, a four-mechanism evolutionary framework for generating stakeholder-aligned strategies, evaluated over 30 synthetic rounds against 14 systems.

What This Paper Is About

Multi-agent systems built on large language models are good at reasoning and generating strategies, but they tend to produce verbose, costly output, explore too few alternative strategies, and lack structured ways to reconcile conflicting stakeholder preferences. Lark addresses these three problems by wrapping an LLM in a discrete, generation-by-generation evolutionary loop that refines candidates, duplicates and specializes the best ones, aggregates stakeholder rankings through influence-weighted Borda voting, and penalizes excessive token use. The goal is a practical, compute-aware framework that improves stakeholder alignment without sacrificing cost competitiveness against commercial models.

Key Contributions

1. A four-mechanism neuroevolutionary framework. Lark combines (i) plasticity (concise, context-sensitive refinements to candidates), (ii) duplication and maturation (copying high-performing candidates and specializing them into new modules), (iii) ranked-choice stakeholder aggregation using influence-weighted Borda scoring, and (iv) compute awareness via token-based penalties that reward brevity and update an efficiency metric each generation.
2. An explicit framing as evolutionary search rather than an MDP. The authors argue that multi-stakeholder strategy generation lacks sequential state transitions and immediate rewards, making discrete population-based search better suited than reinforcement learning, and they enumerate four reasons (no MDP structure, sparse ordinal learning signal, exploration efficiency in discrete language spaces, and natural alignment with LLM-based population sampling).
3. A controlled 30-round evaluation against 14 systems. Lark Full (built on DeepSeek-V3.1) was benchmarked against nine contemporary LLMs plus four ablation variants, achieving a mean rank of 2.55 and a mean composite score of 29.4/50.
4. Ablation evidence for each mechanism. Paired Wilcoxon signed-rank tests isolate the contribution of every component, with duplication/maturation producing the largest deficit when removed.

Main Findings

• Overall standing: Lark Full achieved a mean rank of 2.55 (95% CI [2.17, 2.93]) and a mean composite score of 29.4/50 (95% CI [26.34, 32.46]) across 30 rounds, finishing Top-3 in 80% of rounds.
• Strong baselines trailed: GPT-o3 reached mean rank 4.30 and score 28.8/50; Qwen3-Next-80B reached mean rank 4.40 and score 27.9/50. Differences versus these two were small and non-significant (GPT-o3: ΔMean 0.6, Cohen's dz = 0.20, p = 0.290; Qwen3: ΔMean 1.5, dz = 0.28, p = 0.548).
• Cost competitiveness: Lark Full cost $0.016006 per task versus $0.016424 for GPT-o3. No token caps were enforced; costs reflect natural output lengths under model defaults.
• Duplication/maturation was the largest contributor: Ablating it produced ΔScore = 3.5 (95% CI [3.00, 4.00]), Cohen's dz = 2.53, p = 1.83e-06, and a rank penalty of 6.40 positions (95% CI [5.53, 7.27]).
• Plasticity was the second largest: ΔScore = 3.4 (95% CI [2.74, 4.06]), dz = 1.86, p = 5.93e-06, rank penalty 5.75 positions (95% CI [4.72, 6.78]).
• Ranked-choice voting mattered: Removing it (replacing Borda with simple averaging) gave ΔScore = 2.4 (95% CI [1.69, 3.11]), dz = 1.20, p = 1.94e-05, rank penalty 4.00 positions (95% CI [2.82, 5.18]).
• Token penalties helped rather than hurt: Removing the compute-efficiency penalty gave ΔScore = 2.2 (95% CI [1.72, 2.68]), dz = 1.63, p = 5.93e-06, rank penalty 3.60 positions (95% CI [2.89, 4.31]).
• Ablations were cheaper but lower quality: Variant costs ranged from $0.006053 (NoMutationNoDuplication) to $0.011163 (NoPenalty), while giving up 2 to 3.5 points of quality.
• Large gaps versus some commercial models: Versus GPT-4.1 (ΔMean 2.5, dz = 0.90, p = 1.40e-04) and GPT-4o (ΔMean 4.0, dz = 2.54, p = 1.83e-06), and versus the base DeepSeek-V3.1 model (ΔMean 5.7, dz = 1.12, p = 1.83e-06).
• Practical magnitude: The authors state that a +2 to 3.5 point lift on a 50-point rubric corresponds to roughly 4 to 7 percent of the scale and a mean-rank improvement of +3.6 to 6.4 positions.

Methodology in Plain English

The researchers set up a discrete, generation-based evolutionary loop. A population of candidate strategies is first sampled from an LLM given a scenario (context, objectives, and a set of stakeholders). Each generation then runs through the same steps: plasticity applies small, targeted refinements to candidates with a certain probability, using a temperature parameter that decays over generations to shift from exploration toward exploitation. Stakeholders then rank all candidates, and those rankings are combined into influence-weighted Borda scores, where each stakeholder carries a weight and the weights sum to one. A token-based penalty softens scores for strategies that exceed a target token budget, and an efficiency metric (average Borda score per token) is tracked each generation. High-performing candidates get duplicated with a probability set by a logistic function of how far their score sits above the population mean, then undergo maturation, in which the LLM specializes the copy for a particular stakeholder subgroup or objective dimension. The next generation is built from the top candidates plus the duplicated ones. The authors justify this over reinforcement learning by noting that stakeholder feedback is sparse and ordinal, candidates are evaluated holistically rather than step by step, and LLM generation naturally matches population sampling.

The evaluation used an LLM-as-a-judge protocol with two independent judges to reduce single-evaluator bias. Judging used a 50-point rubric with five equally weighted 10-point criteria: Coverage/Completeness, Feasibility/Realism, Specificity/Thoroughness, Constraint Adherence, and Clarity/Structure. The judge model was openai/gpt-oss-120b. System identities were blinded with anonymized IDs, input order was randomized per judge for each scenario, and judge aggregation used ranked-choice voting with temperature = 0.1.

The benchmark consisted of 30 unique synthetic scenarios across six decision-making domains, five scenarios each: Multi-Stakeholder Trade-offs, Policy Proposal, Product Roadmap, Campaign Plan, Infrastructure Siting, and Clinical Decision-Making. Scenarios were generated with prompt templates specifying stakeholder profiles, constraints, and conflicting objectives. Lark was implemented with DeepSeek-V3.1 as the base model and compared against nine LLMs (DeepSeek-V3.1 baseline, Qwen3-Next-80B-A3B-Thinking, GPT-5-nano, GPT-5-mini, GPT-4o, GPT-4.1, o4-mini, o3-mini, and o3); Claude and other proprietary families were excluded due to funding limits. Four ablations were run under matched conditions: Lark-NoPlasticity, Lark-NoRankedChoiceVoting, Lark-NoMutationAndNoDuplication, and Lark-NoPenalty. Full Lark results represent performance after five evolutionary generations. Statistics used paired Wilcoxon signed-rank tests (two-sided, alpha = 0.05) with Holm's step-down adjustment and paired Cohen's dz effect sizes, treating rounds as the unit of analysis (n = 30).

Why This Matters

Impact on research. The paper provides a concrete, ablation-backed demonstration that biologically motivated operators (plasticity, duplication with specialization) plus ordinal social-choice aggregation can be plugged into an LLM agent loop and produce measurable gains over both the unmodified base model and several commercial systems. It also offers an explicit argument for choosing evolutionary search over reinforcement learning when stakeholder feedback is ordinal and evaluation is holistic, which is a design choice other multi-agent researchers can reuse or contest.

Real-world applications:

• Policy proposal drafting, where multiple constituencies with conflicting priorities must be reconciled.
• Product roadmap planning across competing internal stakeholders.
• Campaign planning, where messaging must satisfy several audience segments.
• Infrastructure siting decisions, which typically involve residents, regulators, and operators with opposed preferences.
• Clinical decision-making scenarios with multiple parties (patients, clinicians, administrators) holding different objectives.

Industry relevance. The cost figures matter commercially: Lark Full at $0.016006 per task sits essentially level with GPT-o3 at $0.016424 while ranking better on these rounds, and the token-penalty mechanism directly targets the verbosity that inflates inference bills. The framework's compute-aware framing and its claim of sitting near a Pareto frontier for these tasks make it relevant to teams deploying LLM agents where both quality and per-call cost are constrained.

Future Directions

1. Real-world validation. The authors identify ecological validity on genuine multi-stakeholder problems as the biggest open question, with planned studies in policy, healthcare, and organizational settings.
2. Compute and energy auditing. Direct measurements of wall-clock time and energy were not taken; the authors plan compute/energy audits and comparisons against multi-objective baselines.
3. Scalability of ranked aggregation. The stability of Borda-based aggregation with larger stakeholder sets is untested; planned scalability tests target 10 to 50 stakeholders with richer preference structures.
4. Broader model coverage and baselines. The study excluded Claude and other proprietary families due to funding limits, and the authors plan hyperparameter sweeps plus expanded baselines including specialized multi-agent frameworks, alongside longitudinal deployment to study iterative decision cycles.

Target Audience

Researchers and practitioners working on LLM-based multi-agent systems, evolutionary computation, and multi-objective or social-choice-based AI. It is also relevant to engineers building stakeholder-facing decision-support tools who need to weigh quality against per-task inference cost, and to readers interested in how biological metaphors (synaptic plasticity, gene duplication and subfunctionalization) translate into working algorithmic operators. Readers looking for large-scale or real-world validated results should note that the authors explicitly frame this as proof-of-concept work inviting community feedback.

Authors’ abstract

We present Lark, a biologically inspired decision-making framework that couples LLM-driven reasoning with an evolutionary, stakeholder-aware Multi-Agent System (MAS). To address verbosity and stakeholder trade-offs, we integrate four mechanisms: (i) plasticity, which applies concise adjustments to candidate solutions; (ii) duplication and maturation, which copy high-performing candidates and specialize them into new modules; (iii) ranked-choice stakeholder aggregation using influence-weighted Borda scoring; and (iv) compute awareness via token-based penalties that reward brevity. The system iteratively proposes diverse strategies, applies plasticity tweaks, simulates stakeholder evaluations, aggregates preferences, selects top candidates, and performs duplication/maturation while factoring compute cost into final scores. In a controlled evaluation over 30 rounds comparing 14 systems, Lark Full achieves a mean rank of 2.55 (95% CI [2.17, 2.93]) and a mean composite score of 29.4/50 (95% CI [26.34, 32.46]), finishing Top-3 in 80% of rounds while remaining cost competitive with leading commercial models ($0.016 per task). Paired Wilcoxon tests confirm that all four mechanisms contribute significantly as ablating duplication/maturation yields the largest deficit ({\Delta}Score = 3.5, Cohen's d_z = 2.53, p < 0.001), followed by plasticity ({\Delta}Score = 3.4, d_z = 1.86), ranked-choice voting ({\Delta}Score = 2.4, d_z = 1.20), and token penalties ({\Delta}Score = 2.2, d_z = 1.63). Rather than a formal Markov Decision Process with constrained optimization, Lark is a practical, compute-aware neuroevolutionary loop that scales stakeholder-aligned strategy generation and makes trade-offs transparent through per-step metrics. Our work presents proof-of-concept findings and invites community feedback as we expand toward real-world validation studies.