Skip to content

Context Learning for Multi-Agent Discussion

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

Why this paper was selected

In-context learning for multi-agent discussion prevents cascading incoherence

Ask a follow-up

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

💬 Ask ChatGPT✦ Ask Claude

TL;DR

M2CL trains a per-agent context generator that dynamically rewrites each LLM's instruction prompt every discussion round, resolving the context misalignment that causes multi-agent debates to stall in incoherent "majority noise." Evaluated across 9 benchmarks spanning reasoning, embodied navigation, and mobile GUI control, M2CL outperforms the best existing MAD baselines by 20–50% with ≤10% runtime overhead.

Problem

Multi-Agent Discussion (MAD) frameworks assign each LLM a static role/instruction, but these pre-assigned contexts lack mechanisms to integrate intermediate conclusions from other agents. The result is discussion inconsistency: agents re-derive already-settled steps, contradict each other, or converge on wrong majorities. Existing solutions either require manual per-round prompt engineering or optimize workflows in a one-shot, coarse-grained manner.

Method

M2CL comprises two stages grounded in a formal theorem (Theorem 4.1) bounding the summed activation distance to the correct answer:

Context Initialization. A lightweight projector \(F\) (trained to match \(f(a([I^b_i; P]))\), the activation of initial context plus problem) selects \(N\) initial instructions from a pool by solving: $\(I^b = \arg\min_{I^b} \min_\omega \left\| \sum_i \omega_i F([I^b_i; P]) - v_P \right\|\)$ where \(v_P\) is the sentence vector of the problem. This encourages near-orthogonal activations, giving agents diverse, non-redundant starting perspectives.

Context Evolution. A generator \(G_{\theta_i}\) produces per-round instruction \(I^t_i = G_{\theta_i}(P, I^b_i, \bar{X}^{t-1}_i)\). The training objective (per LLM) is: $\(L(\theta_i) = \|a(G_{\theta_i}(\cdot)) - a(X^{t-1}_i)\| + \alpha \|C^t_i - C^b_i\|\)$ The coherence weight \(\alpha\) is updated via a dual gradient descent on a constrained formulation, creating a self-adaptive mechanism: early in discussion when answers diverge, \(\alpha\) drops (allowing large context shifts toward convergence); as consensus forms, \(\alpha\) stabilizes (preventing premature collapse of diversity).

Key Contributions

  • M2CL algorithm: per-agent learned context generators that evolve instructions each round, replacing static role prompts.
  • Context initialization: activation-orthogonality criterion for selecting diverse initial prompts, distilled into a lightweight projector \(F\).
  • Round-wise utility: novel per-round criterion decoupling inter-LLM dependencies by replacing cross-agent activation alignment with a local temporal consistency term (proven as an upper bound).
  • Self-adaptive \(\alpha\): dual-variable update that automatically balances convergence speed vs. perspective diversity without manual tuning.
  • Comprehensive evaluation across 9 benchmarks with transferability to unseen LLM architectures.

Results

All results use 4 LLMs; best-of-N (BoN) is the primary baseline ceiling.

  • MMLU (Qwen-7B): M2CL 92.5% vs. BoN 74.2%, Debate 71.1%, GPTSwarm 76.3%
  • MATH (Qwen-7B): M2CL 47.8% vs. BoN 24.9%, DyLAN 26.7%
  • GPQA (Qwen-7B): M2CL 66.1% vs. BoN 36.4%, GPTSwarm 35.6%
  • Code (Qwen-7B): M2CL 80.3% vs. BoN 62.5%, DyLAN 63.4%
  • ALFWorld embodied (Qwen-7B): M2CL 39.9% vs. BoN 31.5%, MacNet 33.6%
  • SciWorld embodied (Qwen-7B): M2CL 45.3% vs. BoN 35.3%, MacNet 37.0%
  • GAIA (Qwen-7B): M2CL 33.6% vs. BoN 21.1%
  • PDDL planning (Qwen-7B): M2CL 34.7% vs. BoN 26.3%, MacNet 29.5%
  • Gains replicate on Qwen-14B (e.g., MATH 51.7% vs. BoN 27.8%)
  • Runtime overhead ≤10% over baseline MAD while achieving >20% absolute gains
  • Context generators transfer to different LLM architectures with consistent improvement

Limitations

  • Requires training the context generator and projector on task-specific data, introducing a supervised pre-training phase absent in prompt-only MAD methods.
  • The context pool \(\{I^b_i\}^M_{i=1}\) must be predefined; pool quality and diversity bound initialization effectiveness.
  • Majority vote aggregation in the final round may still fail when agent counts are small or errors are correlated.
  • Theoretical bound (Theorem 4.1) assumes \(L_a\)-smoothness of attention activations, which may not hold uniformly across model families.
  • Transferability experiments are described qualitatively but full cross-architecture ablations appear limited in the main text.

Relevance to Agentic AI / LLM Agents

This work directly targets a core failure mode in multi-agent LLM systems: static role assignments that cannot incorporate emerging inter-agent information. By formalizing context as a learnable, dynamically evolving object and providing a principled training signal (activation alignment + diversity), M2CL advances the state of structured multi-agent reasoning beyond topology-optimization or one-shot prompt search. The self-adaptive convergence mechanism addresses the "sycophancy to majority noise" problem that plagues debate-style agents, and the transferability finding suggests context generators could serve as plug-in modules in heterogeneous agent pipelines—directly relevant to building robust agentic systems for embodied AI, GUI automation, and complex reasoning.