Federation over Text: Insight Sharing for Multi-Agent Reasoning

This paper introduces "Federation over Text" (FoT), a gradient-free framework where multiple LLM agents iteratively share and aggregate their reasoning traces into a central metacognitive insight library, significantly improving reasoning accuracy and efficiency across diverse tasks without requiring gradient optimization or supervision.

The Gist

Imagine a world where every time a brilliant student solves a difficult math problem, they write down their solution, throw it in a trash can, and then the next student has to figure out the exact same problem from scratch, completely forgetting the first student's clever tricks.

That is essentially how most current AI "agents" (smart computer programs) work today. They are incredibly smart, but they suffer from amnesia. Every time they face a new challenge, they start from zero, wasting time and energy reinventing the wheel. They also tend to work in isolation, never sharing their "aha!" moments with their peers.

The paper you shared, "Federation over Text" (FoT), proposes a brilliant solution to this problem. It's like creating a global, shared "Wisdom Library" where AI agents can swap their best thinking strategies without ever revealing their private secrets.

Here is the breakdown using simple analogies:

1. The Problem: The "Reinventing the Wheel" Syndrome

Currently, if you have 100 AI agents trying to solve 100 different math problems, they all work alone.

• Agent A solves a hard algebra problem by using a specific shortcut.
• Agent B tries to solve a similar problem but doesn't know the shortcut, so they struggle for hours.
• The Result: Agent A's hard-earned wisdom is lost. Agent B wastes time. The system is inefficient.

2. The Solution: The "Shared Wisdom Library" (FoT)

The authors propose a framework called Federation over Text. Think of it as a centralized "Recipe Book" that gets updated every day.

Here is how the process works, step-by-step:

• Step 1: The Local Chef (The Agent)
  Imagine an AI agent is a chef in a restaurant. They are given a specific dish to cook (a task). They cook it using their own skills.
• Step 2: The "Meta-Commentary" (The Insight)
  Instead of sending the raw ingredients or the messy kitchen to the boss, the chef writes a short, clever note on a card.
  • Bad Note: "I cooked a steak." (Too vague)
  • Good Note (The Insight): "When the pan is too hot, lower the flame immediately and add butter at the end to prevent burning. This works for any searing task."
    This note is called a Reasoning Trace. It's a distilled lesson, not the raw data.
• Step 3: The Librarian (The Server)
  All the chefs send their cards to a central Librarian. The Librarian doesn't just stack them; they read them, find patterns, and combine similar notes into Master Recipes.
  • Example: If Chef A says "Lower heat for steak" and Chef B says "Lower heat for fish," the Librarian creates a new rule: "Rule: Always lower heat for delicate proteins."
• Step 4: The Update
  The Librarian sends this updated "Master Recipe Book" back to all the chefs. Now, when Chef B goes to cook a fish tomorrow, they don't have to guess; they just look up the rule in the book and cook perfectly.

3. Why is this special? (The "Text" vs. "Gradients" Magic)

In traditional AI training, computers share "weights" (mathematical numbers that define the brain). This is like trying to share a brain by sending someone a pile of sand; it's messy, requires huge computing power, and you can't read it.

FoT is different. It shares Text (ideas and language).

• Analogy: Instead of trying to merge two people's brains by mixing their DNA (gradients), FoT is like two people sitting down to have a conversation and agreeing on a set of best practices.
• Privacy: Because they only share the lessons (the "how-to" notes) and not the problems (the raw data), the agents don't need to reveal their private information. It's like sharing a tip on how to fix a leaky faucet without showing the plumber your entire house blueprint.

4. The Results: Superpowers for Everyone

The paper tested this on three tough challenges:

1. Math: Agents solved harder math problems faster and with fewer mistakes.
2. Cross-Domain: An agent learned a math trick and successfully applied it to a chemistry problem (e.g., using a logic puzzle technique to solve a molecule structure).
3. Research: Agents read old scientific papers and created a "cheat sheet" that predicted the core ideas of future papers with over 90% accuracy.

The Bottom Line:
Federation over Text turns a group of isolated geniuses into a collective super-intelligence. It allows AI to learn from its own past mistakes and successes, share those lessons instantly, and get better at everything—math, coding, and science—without needing to be retrained from scratch.

It's the difference between a student who studies alone in a library and a student who has access to a living, breathing textbook written by thousands of other students, updated every single day.

Technical Summary

1. Problem Statement

Large Language Model (LLM) agents currently face two critical limitations when solving complex, multi-step tasks:

• Inefficiency: Agents often reason "from scratch" for every new problem instance, leading to redundant computation and slow performance, even when similar problems have been solved previously.
• Isolation: In multi-agent systems, reasoning processes are typically isolated. While agents may exchange intermediate outputs or final solutions, their internal reasoning traces (the "how" and "why" of a solution) are not stored, shared, or distilled into reusable knowledge. Consequently, hard-won insights are lost and cannot be transferred across agents or domains.

Existing solutions like Federated Learning (FL) aggregate model gradients or weights, which is computationally prohibitive for large generative models, impossible for closed-source LLMs, and lacks interpretability (parameter aggregation does not equate to sharing high-level reasoning strategies).

2. Methodology: Federation over Text (FoT)

The authors propose Federation over Text (FoT), a framework that enables multiple agents to collaboratively build a shared, evolving library of metacognitive insights without sharing raw problem instances or performing gradient optimization.

Core Workflow

The process operates in iterative rounds, analogous to Federated Learning but at the semantic/text level:

1. Local Reasoning & Reflection: Each agent solves a specific task using a base LLM. After solving, the agent reflects on its process to extract reusable techniques, strategies, or "behaviors" (reasoning traces).
2. Trace Upload: Agents upload these concise, abstracted reasoning traces to a central server. Crucially, raw problem data and full reasoning histories remain local; only the distilled insights are shared.
3. Server Aggregation: The server acts as an aggregator. It clusters similar reasoning traces, identifies connections between different approaches to the same problem, and distills them into a unified Insight Library. This library contains high-level, cross-domain principles (e.g., "Constraint Satisfaction Propagation" or "Dynamic Programming State Decomposition").
4. Library Distribution: The updated Insight Library is broadcast back to all agents. In the next round, agents use this library as an external tool to guide their local reasoning, improving efficiency and accuracy.

Key Technical Distinctions

• Semantic vs. Parameter Level: Unlike FL, FoT operates entirely in the text space. It requires no fine-tuning, reinforcement learning, or gradient computation.
• Decoupling: The model used to generate the insight library can differ from the models used by agents to solve tasks (heterogeneous models).
• Privacy: Only abstracted metacognitive summaries are shared, protecting raw data and proprietary problem instances.

3. Key Contributions

• Novel Framework: Introduction of FoT, the first federated-like framework for multi-agent reasoning that aggregates reasoning traces into a shared insight library rather than model weights.
• Cross-Domain Generalization: Demonstration that insights derived from one domain (e.g., mathematics) can effectively guide reasoning in unrelated domains (e.g., chemistry or coding), facilitating true cross-domain collaboration.
• Weak-to-Strong Generalization: Evidence that insight libraries generated by weaker models can significantly improve the performance of stronger models, suggesting that high-level reasoning patterns can be distilled and transferred regardless of model capacity.
• Research Insight Discovery: A novel application where FoT is used to analyze academic papers (ICLR) to generate a library of research insights that can predict and cover the core contributions of future papers.

4. Experimental Results

The authors evaluated FoT across three distinct applications:

A. Mathematical Problem Solving (LiveMathBench)

• Performance: FoT improved average accuracy by 24% across various benchmarks (AIME, AMC, etc.) compared to isolated agents.
• Efficiency: It reduced reasoning token consumption by 28% and decreased the number of reasoning loops (redundant steps).
• Cross-Model Transfer: An insight library generated by a weaker model (DeepSeek-R1-Distill-Qwen-7B) improved the performance of a stronger model (Gemini-3.0-Pro-Preview) on hard tasks, achieving a 94.42% accuracy on GPQA-Diamond, surpassing the standalone performance of Gemini-3.1-Pro-Preview (94.1%).

B. Multi-Domain Collaboration

• Agents solved tasks across Math, Science (GPQA), and Coding.
• Case Study: An insight derived from a math problem ("Constraint Satisfaction Propagation") was successfully applied to a chemistry problem to correctly identify molecular structures, demonstrating the transfer of abstract logical strategies.
• True-Thinking Score: FoT increased the "true-thinking score" (proportion of reasoning steps essential to the answer), indicating that agents skipped "decorative" or redundant reasoning steps.

C. Research Insight Discovery

• Setup: Agents analyzed ICLR papers from one year to generate an insight library.
• Result: The generated library covered the core technical contributions of >90% of accepted papers in the subsequent year.
• Comparison: FoT significantly outperformed baselines using standalone LLMs or Retrieval-Augmented Generation (RAG) in guiding future research directions.

5. Significance and Implications

• Scalable Knowledge Sharing: FoT provides a scalable mechanism for LLM agents to learn from collective experience without the computational cost of retraining models.
• Privacy-Preserving: By sharing only abstract insights rather than raw data, FoT addresses privacy and intellectual property concerns in collaborative AI research.
• Interpretability: The resulting insight library is human-readable and interpretable, unlike the opaque weight updates in traditional federated learning.
• Future Directions: The framework opens new avenues for designing personalized agent systems, handling distribution shifts, and optimizing communication efficiency in multi-agent ecosystems. It bridges the gap between distributed optimization and multi-agent reasoning, suggesting that "learning to reason" can be a collaborative, federated process.

In conclusion, Federation over Text represents a paradigm shift from isolated agent reasoning to a collaborative ecosystem where agents collectively distill and share high-level cognitive strategies, leading to more efficient, accurate, and generalizable problem-solving capabilities.