Skip to content

Weak-Link Optimization for Multi-Agent Reasoning and Collaboration

🕒 Published (v1): 2026-04-17 11:36 UTC · Source: Arxiv · link

Why this paper was selected

Weak-link optimization prevents error amplification in multi-agent collaborative reasoning

Ask a follow-up

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

💬 Ask ChatGPT✦ Ask Claude

TL;DR

WORC is a two-stage optimization framework that identifies the weakest agent in a multi-agent LLM reasoning pipeline (via swarm intelligence + meta-learning) and then compensates by allocating additional inference budget to that agent. On six reasoning benchmarks, WORC achieves 82.2% average accuracy while improving cross-architecture stability compared to unoptimized baselines.

Problem

Multi-agent reasoning chains are only as reliable as their weakest component: errors introduced by an underperforming agent propagate downstream, and consensus mechanisms like majority voting or debate do not identify or selectively compensate low-capability agents. Existing work focuses on strengthening high-performing agents or applying uniform budget allocation, leaving systematic weak-agent detection and targeted reinforcement underexplored.

Method

WORC operates in two stages applied on top of any multi-agent framework (demonstrated via the hand-built AgentChain, AC):

Stage 1 — Weak Agent Localization. - A weight knowledge base is built offline: Swarm Intelligence Algorithms (SIA — PSO, GWO, Hippopotamus Optimization) search for optimal per-agent weight vectors \(X^* = (x^*_1, \ldots, x^*_N)\) that maximize accuracy \(f(X)\) on few-shot samples from diverse task datasets, subject to \(\sum_j x_j = 1\), \(x_j \in [0.05, 1.0]\). - Each task is encoded as a task signature \(s_T = [\mathbf{e}_{\text{mean}}; \mathbf{f}_{\text{stat}}] \in \mathbb{R}^{d+k}\), concatenating a mean OpenAI embedding over \(M\) unlabeled samples with four structural statistics (length mean/variance, entity count, operator density). - A two-layer MLP meta-weight predictor \(\varphi: s_T \mapsto \hat{w}\) is trained with MSE loss \(\mathcal{L}(\varphi) = \sum_k \|\hat{w}_{T_k} - w^*_{T_k}\|_2^2\) to zero-shot predict agent weight vectors on unseen tasks. The agent with the lowest \(\hat{w}_i\) is the identified weak link.

Stage 2 — Weak-Link Optimization. - Additional inference budget \(B\) is allocated via a softmax-inverse weighting: $\(m_i = \left\lfloor B \cdot \frac{\exp(\tau(1 - \hat{w}_i))}{\sum_j \exp(\tau(1 - \hat{w}_j))} \right\rceil\)$ so lower-weight (weaker) agents receive more repeated sampling passes, each conditioned on prior outputs. - A VoteAgent scores and selects the best output from each agent's sample set before passing it downstream.

An EvalAgent provides real-time quality scoring to bind SIA fitness updates to actual output quality during knowledge base construction.

Key Contributions

  • Frames multi-agent reasoning degradation as a weak-link problem (Buckets Effect) and proposes targeted budget compensation rather than uniform or strength-focused allocation.
  • SIA-driven weight knowledge base: uses population-based global search (PSO, GWO, HO) to extract per-agent marginal contribution weights from few-shot task samples without explicit supervision.
  • Meta-learning weight predictor: zero-shot mapping from task signatures to agent weight vectors, enabling cross-task generalization without re-running SIAs on each new task.
  • Uncertainty-driven budget allocation: temperature-controlled exponential inverse weighting ensures weak agents receive proportionally more inference budget.
  • Plug-in framework: WORC is method-agnostic and demonstrated across multiple backbone architectures.

Results

  • 82.2% average accuracy across six benchmarks (MATH, GSM8K, BBH, MMLU-CF, HotpotQA, LongBench) using GPT-4o; outperforms the unoptimized AgentChain (AC) baseline on all six.
  • Improves framework stability (lower variance across runs) compared to AC and consensus baselines (majority voting, debate).
  • Demonstrates cross-architecture generalization: WORC weights derived from one backbone transfer to other multi-agent architectures.
  • Majority voting and debate baselines mitigate but do not eliminate weak-agent influence; WORC shows further gains over both (specific delta figures were in the truncated portion of the text).

Limitations

  • All experiments use GPT-4o as the sole backbone; performance on open-weight or heterogeneous-model architectures is not fully characterized.
  • Knowledge base construction requires running SIAs on few-shot samples from related task datasets; cold-start overhead exists for genuinely novel task domains.
  • The EvalAgent and VoteAgent rely on prompt-based scoring, introducing a dependency on the quality of evaluation prompts.
  • AgentChain is a linear chain — the framework's behavior on more complex topologies (tree, graph, mixture-of-experts) is not systematically evaluated.
  • The SIA variants tested (PSO, GWO, HO) are chosen as illustrative examples rather than optimized for the task; sensitivity to SIA choice is not analyzed.

Relevance to Harnesses / Meta-Harnesses

WORC is a meta-harness in the strict sense: it wraps an arbitrary multi-agent reasoning framework and modifies agent-level resource allocation without changing the underlying agents themselves. The offline SIA training + meta-learning predictor pipeline is itself a harness-of-harnesses — a meta-level controller that learns to configure the inner framework from task features. The pattern of building a knowledge base of optimal configurations via search and then retrieving/adapting them at inference time is directly analogous to meta-harness design patterns (cf. automated workflow generation, prompt orchestration frameworks). Researchers tracking harness construction should note the SIA-as-optimizer-for-agent-weights pattern as a reusable primitive for any system where agent contributions are heterogeneous and task-dependent.