💻 computer science

Lark: Biologically Inspired Neuroevolution for Multi-Stakeholder LLM Agents

Lark is a biologically inspired neuroevolutionary framework that enhances multi-stakeholder LLM agents by integrating plasticity, duplication/maturation, ranked-choice aggregation, and compute-aware penalties to efficiently generate concise, high-performing strategies while transparently managing trade-offs.

Original authors: Rikhil Tanugula, Dheeraj Chintapalli, Sunkalp Chandra

Published 2026-08-26

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Rikhil Tanugula, Dheeraj Chintapalli, Sunkalp Chandra

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the rapidly expanding world of artificial intelligence, a new challenge has emerged that goes beyond simply making machines smarter. It is about teaching them to work together in groups, much like a team of humans trying to solve a complex problem where everyone has different opinions and goals. This field, known as multi-agent systems, is crucial for tasks ranging from coordinating autonomous vehicles to planning healthcare strategies. However, when these digital agents are powered by large language models—the same technology behind modern chatbots—they often struggle with two major issues. First, they tend to be overly wordy, generating long, repetitive explanations that waste computing power and money. Second, they often fail to balance the needs of different people involved in a decision, such as a business owner wanting profit and a community group wanting safety. Traditional methods for training these systems often rely on step-by-step learning, similar to how a student learns math by solving one problem after another. But for complex, creative strategy planning, this approach can be inefficient and rigid.

To address these hurdles, a team of researchers has developed a new framework called Lark, which takes its cues from the way living things evolve. Instead of forcing the artificial agents to learn through a strict sequence of steps, Lark treats the creation of strategies like a biological population. Imagine a group of digital agents proposing different solutions to a problem. In each round, the system evaluates these proposals not just on whether they work, but on how well they satisfy a diverse group of stakeholders. The system then selects the best ideas, makes small, targeted improvements to them, and occasionally copies the strongest ideas to create specialized new versions. This process repeats over several generations, with the system getting better at balancing competing needs and trimming away unnecessary words with every cycle. The result is a method that produces high-quality, concise strategies without the heavy computational cost usually associated with such complex reasoning.

The researchers tested this approach by creating thirty different simulated scenarios, ranging from policy proposals and product roadmaps to clinical decision-making. In each scenario, they asked Lark and fourteen other competing systems to generate solutions for problems involving multiple stakeholders with conflicting interests. To ensure fairness, they used an independent judging system to score the answers based on five criteria: how complete the solution was, whether it was realistic, if it included specific details, if it followed the rules, and how clearly it was written. The results showed that the full version of Lark consistently outperformed the other systems. On average, it ranked as the second-best system out of the fourteen, achieving a high score of nearly thirty out of fifty. It finished in the top three in eighty percent of the rounds. While other powerful systems came close, Lark managed to maintain this high level of performance while costing about the same as the leading commercial models, roughly sixteen cents per task.

What makes Lark work so well is a combination of four specific mechanisms inspired by biology. The first is a process called plasticity, which allows the system to make quick, context-sensitive adjustments to a strategy without starting over. If a solution is mostly good but has a small flaw, the system tweaks it to fix that specific issue. The second mechanism is duplication and maturation. When a strategy performs well, the system copies it and then specializes the copy to focus on a specific group of people or a particular goal, much like how a gene might duplicate in nature and evolve to perform a slightly different function. The third mechanism involves a voting system where the preferences of different stakeholders are combined. Instead of simply averaging their opinions, the system uses a method that weighs how much each person cares about the outcome, ensuring that the final decision respects the most critical concerns. The final mechanism is a penalty for being too wordy. The system is designed to reward brevity, automatically reducing the score of any solution that uses more words than necessary, which keeps the computing costs low and the reasoning sharp.

When the researchers removed any one of these four mechanisms to see how important they were, the system's performance dropped significantly. The biggest loss occurred when they stopped the duplication and specialization process, causing the quality of the solutions to fall by a noticeable margin. Removing the ability to make quick tweaks also hurt performance, followed by the voting system and the word-count penalty. This suggests that all four parts are essential for the system to function effectively. The study indicates that by mimicking these natural evolutionary processes, artificial intelligence can become much better at handling the messy, conflicting realities of human decision-making. The researchers emphasize that while these results are promising, they are based on controlled simulations. They plan to test the system in real-world situations, such as actual policy-making or hospital planning, to see if it can handle the unpredictability of the real world. For now, the work offers a proof of concept that a biologically inspired approach can make artificial agents more efficient, fair, and capable of finding common ground among diverse groups.

Technical Summary: Lark – Biologically Inspired Neuroevolution for Multi-Stakeholder LLM Agents

1. Problem Statement

Traditional Multi-Agent Systems (MAS) utilizing Large Language Models (LLMs) face three critical limitations in complex decision-making domains:

Verbosity and Cost: LLM-driven agents often produce redundant outputs, leading to high computational expenses.
Limited Exploration: Current systems struggle to explore diverse alternative strategies and potential future states effectively.
Multi-Stakeholder Misalignment: Existing frameworks lack robust mechanisms to handle multiple, often conflicting stakeholder objectives without relying on constrained optimization or Markov Decision Processes (MDPs) that require immediate reward signals.

The authors argue that multi-stakeholder strategy generation is better suited for discrete, population-based search rather than sequential state-action transitions, as strategies are evaluated holistically upon completion rather than incrementally.

2. Methodology: The Lark Framework

Lark is a biologically inspired neuroevolutionary framework that couples LLM-driven reasoning with a discrete-generation evolutionary search. Unlike Reinforcement Learning (RL) approaches that require MDP formulations, Lark operates on populations of candidate strategies, evolving them through selection, mutation, and specialization. The framework integrates four core mechanisms:

2.1 Plasticity (Context-Sensitive Refinement)

Inspired by Hebbian learning and neuromodulation, this mechanism allows agents to make concise, context-aware adjustments to candidate strategies within a single generation.

Implementation: An LLM-driven refinement operator ( $\Phi$ ) takes a candidate strategy and scenario context to produce an improved variant.
Process: The LLM identifies context-specific weaknesses, proposes targeted modifications, and refines implementation details while preserving the core structure to maintain diversity.
Control: Applied with a probability ( $p_{plast}$ ) controlled by a decaying temperature parameter to balance exploration and exploitation.

2.2 Duplication and Maturation (Modular Specialization)

Motivated by gene duplication and subfunctionalization in evolutionary biology, this mechanism promotes functional diversity by copying and specializing high-performing strategies.

Duplication: High-performing strategies are selected for duplication based on fitness-proportional selection (using a logistic function of the aggregated stakeholder score).
Maturation: Duplicated strategies undergo a stochastic specialization process where the LLM is prompted to:
- Identify a specific stakeholder subgroup or objective dimension.
- Modify the strategy to specialize for that subgroup.
- Introduce novel implementation details not present in the parent.
Goal: To model the divergence of duplicated genes into distinct but related functions, leveraging LLM semantic understanding rather than random mutations.

2.3 Stakeholder Ranked-Choice Voting

To aggregate heterogeneous preferences without assuming cardinal utility or requiring interpersonal comparisons, Lark uses an influence-weighted Borda count.

Mechanism: Each stakeholder provides an ordinal ranking of candidate strategies.
Scoring: The consensus score $B(x_i)$ is calculated as the sum of influence weights multiplied by the inverse rank position.
Robustness: This method is robust to strategic manipulation and does not require stakeholders to reveal cardinal utilities. A Coefficient of Variation (CV) is computed to quantify the strength of consensus.

2.4 Compute-Efficiency Penalties

To enforce resource awareness and reward brevity, a token-based penalty is applied to the fitness evaluation.

Formula: The adjusted fitness $R(x_i)$ reduces the Borda score if the token count $T(x_i)$ exceeds a target budget $T_{target}$ .
Metric: The system tracks a running efficiency metric $E(t)$ (average quality-per-token) to monitor improvements over generations.

2.5 The Evolutionary Loop

The system iterates through generations:

Initialization: A population of strategies is sampled via LLM.
Plasticity: Strategies are refined with probability $p_{plast}$ .
Evaluation: Stakeholders rank all strategies.
Scoring: Borda scores are computed and adjusted for token penalties.
Selection & Reproduction: Strategies are selected for duplication based on fitness; duplicates undergo maturation.
Next Generation: The top- $k$ strategies from the combined pool (original + duplicated) form the next generation.

3. Key Contributions

Novel Framework: Introduction of Lark, an LLM-driven MAS that combines plasticity, duplication/maturation, ranked-choice voting, and compute-aware penalties in a single neuroevolutionary loop, a combination the authors note has not been previously achieved in this specific context.
Paradigm Shift: A shift from MDP-based constrained optimization to a discrete, compute-aware evolutionary search that evaluates strategies holistically.
Mechanism Integration: Demonstration that biological principles (synaptic adaptation, gene duplication) can be effectively translated into LLM agent behaviors for multi-stakeholder alignment.
Transparency: The framework makes trade-offs transparent through per-step efficiency metrics and final analyses, avoiding the "black box" nature of some RL approaches.

4. Experimental Results

The authors conducted a controlled evaluation over 30 synthetic rounds comparing Lark Full against 13 other systems (including proprietary models like GPT-4o, GPT-o3, and Qwen3-Next-80B) and four ablation variants.

4.1 Performance Metrics

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]).
It finished in the Top-3 in 80% of rounds.
Cost Efficiency: Lark Full was cost-competitive with leading commercial models, averaging $0.016 per task, comparable to GPT-o3 ($0.0164).

4.2 Ablation Studies

Paired Wilcoxon signed-rank tests confirmed that all four mechanisms contribute significantly. The performance deficits when removing mechanisms were:

Duplication/Maturation: Largest deficit ( $\Delta$ Score = 3.5, $d_z$ = 2.53, $p < 0.001$ ).
Plasticity: Significant deficit ( $\Delta$ Score = 3.4, $d_z$ = 1.86).
Ranked-Choice Voting: Significant deficit ( $\Delta$ Score = 2.4, $d_z$ = 1.20).
Token Penalties: Significant deficit ( $\Delta$ Score = 2.2, $d_z$ = 1.63).

4.3 Baseline Comparisons

Lark Full showed statistically significant improvements over GPT-4.1 and GPT-4o (large effect sizes).
Performance was statistically comparable (parity) to GPT-o3 and Qwen3-Next-80B, though Lark Full maintained a slight edge in mean rank.

5. Significance and Claims

The paper positions Lark as a proof-of-concept rather than a finalized production system. Its primary significance lies in:

Demonstrating Feasibility: Showing that biologically inspired neuroevolution can effectively scale stakeholder-aligned strategy generation in LLM agents.
Practicality: Offering a compute-aware alternative to formal constrained optimization that is transparent and adaptable.
Community Invitation: The authors explicitly invite community feedback and plan to expand toward real-world validation studies in policy, healthcare, and organizational settings.

The authors remain modest about their claims, acknowledging limitations such as the use of synthetic benchmarks, the need for broader model coverage, and the absence of direct wall-clock time or energy measurements in the current study. They emphasize that while the results are promising, real-world ecological validity remains to be established.