Discovering Novel LLM Experts via Task-Capability Coevolution

🕒 Published (v1): 2026-01-01 · Source: ICLR · Venue: ICLR 2026 · link

Why this paper was selected

Task-capability coevolution to discover novel expert LLMs for emergent skills

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

AC/DC (Assessment Coevolving with Diverse Capabilities) is an open-ended framework that jointly evolves a population of LLMs (via model merging) and a population of synthetic tasks (via LLM-generated data), discovering diverse specialist models without any explicit benchmark optimization. The resulting "task forces" of small models collectively cover more benchmark problems than larger single models or hand-curated expert ensembles, achieving GPT-4o-level collective Coverage with a fraction of the parameters.

Problem

Standard LLM development produces one model per training run on static datasets or reward functions, leaving no mechanism for autonomous, continual discovery of complementary capabilities. Individual models suffer from entangled representations and cannot simultaneously master all domains; scaling up parameters is prohibitively expensive. There is no principled way to grow a diverse collective of specialized models in a single automated run.

Method

AC/DC maintains two co-evolving archives: a model archive \(A_M\) and a synthetic task archive \(A_Q\).

Model evolution applies crossover (weighted linear interpolation of task vectors \(\tau_{p_i} = \theta_{\text{parent}_i} - \theta_{\text{base}}\) from two randomly selected parent LLMs) and mutation (singular value decomposition perturbation: \(W = U\Sigma V^T\), noising the top-\(k\) singular values of \(\Sigma\)). Offspring are evaluated on \(A_Q\) to produce binary skill vectors indicating per-task success. A gibberish filter (LLM coherence judge) discards degenerate offspring. The archive is updated via Dominated Novelty Search (DNS), a quality-diversity algorithm that rewards behavioral distance from higher-performing neighbors: $\(\tilde{f}_i = \frac{1}{k}\sum_{j \in K_i} d_{i,j}\)$

Task evolution runs every \(G_{\text{task}}\) generations: a "scientist" LLM proposes new tasks (question + Python scoring function) by transforming parent tasks toward higher/lower difficulty or novel variants. Candidates pass novelty filtering (cosine similarity against a global task embedding database), self-reflection/validation (scientist solves its own task; compilation errors trigger correction), and a minimal-criterion filter that discards tasks no model can solve.

After \(G\) generations, a fixed-size task force of \(N_{\text{tf}}\) models is selected from the historical archive to maximize collective Coverage — the fraction of problems solved by at least one model: $\(\text{Coverage} = \frac{1}{Q}\sum_{q=1}^{Q}\bigvee_{i=1}^{N}(x_{q,i} = y_q)\)$

Key Contributions

• AC/DC: a unified coevolution framework combining evolutionary model merging and synthetic task generation for autonomous LLM population discovery.
• Demonstration that small evolved collectives (e.g., 3× Qwen2.5-7B) exceed Coverage of single models 10× larger (72B) on out-of-distribution benchmarks.
• Evidence of open-ended improvement: Coverage increases across generations without any benchmark-specific optimization signal.
• Best-of-N performance improvement over control and expert baselines (+0.99–1.34% avg), narrowing the gap to large proprietary models.
• Qualitative evidence of emergent domain specialization (biology, chemistry, CS) across evolved models despite no explicit niche assignment.

Results

• Coverage vs. Experts (N=3): AC/DC improves average Coverage by +3.99% across four model families (Qwen2 7B, Qwen2.5 7B, Qwen3 14B, DeepSeek V1 7B).
• Coverage vs. Big Model (N=8): +10.19% average improvement; Qwen2.5-7B N=8 task force surpasses a 72B model using only ~29% of parameters.
• Coverage vs. GPT-4o (N=8): +2.99% average; Qwen2.5-7B N=3 (3× 7B) outperforms GPT-4o.
• Best-of-N vs. Experts (N=3): +1.34% average; DeepSeek V1 7B shows +11.73%.
• Best-of-N vs. Big Model (N=3/N=8): DeepSeek 7B reaches within 1.27% of DeepSeek 67B at N=3, surpasses it by 4.94% at N=8 using 16% fewer parameters.
• vs. prior QD methods (Tab. 3): AC/DC Coverage 69.00% at N=8 vs. DNS 66.48%, CQD 65.42% — without any benchmark-specific optimization.
• Ablations: Removing QD selection (DNS) costs −2.39%/−0.88% (N=3/N=8); removing gibberish filter costs −2.46%/−1.18%; removing all evolutionary components costs −2.36%/−7.02%.
• Coevolution vs. static tasks: coevolving tasks improves Coverage by 3.62% at N=8 over static synthetic datasets.

Limitations

• Best-of-N selection does not consistently outperform baselines vs. large/frontier models; closing this gap is acknowledged as an open problem.
• Qwen3 14B N=3 underperforms the expert baseline (−0.21%), indicating architecture/scale sensitivity.
• GPT-4o comparison at N=3 shows −4.52% average Coverage, meaning small task forces do not uniformly exceed frontier models.
• The scientist LLM and judge LLM introduce external compute dependencies and potential quality ceilings on task complexity.
• Experiments are restricted to model families derived from the same base architecture; cross-family merging is not evaluated.
• No wall-clock or compute-cost comparison against the full training baseline is provided, making efficiency claims partially informal.

Relevance to Agentic AI / LLM Agents

AC/DC directly addresses multi-agent system design by automating the discovery of diverse specialist LLMs that function as complementary agents in a collective, which is a key bottleneck in building robust multi-agent pipelines. The Coverage metric and Best-of-N evaluation mirror real agentic deployment patterns where multiple models vote, route, or select answers collaboratively. The open-ended coevolution paradigm — where models and tasks mutually drive each other toward novelty — provides a principled alternative to static expert assignment in LLM routing and mixture-of-experts systems. For researchers tracking LLM agent ensembles, this work offers both an algorithmic template (QD + model merging + synthetic task generation) and empirical evidence that collective intelligence from small models can match frontier models at reduced inference cost.