🤖 AI

Learning to Collaborate: An Orchestrated-Decentralized Framework for Peer-to-Peer LLM Federation

Original authors: Inderjeet Singh, Eleonore Vissol-Gaudin, Andikan Otung, Motoyoshi Sekiya

Published 2026-01-27

📖 4 min read☕ Coffee break read

Original authors: Inderjeet Singh, Eleonore Vissol-Gaudin, Andikan Otung, Motoyoshi Sekiya

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

Imagine a group of brilliant but isolated experts, each working in their own private office. They are all trying to solve a very difficult puzzle: writing perfect computer code.

The Problem: Each expert has a unique set of notes (data) that they cannot share with anyone else because of strict privacy rules. If they try to combine their notes, they break the rules. If they try to send their notes to a central boss to mix them, they risk the boss stealing the notes or the boss getting hacked.
The Old Way (Centralized): Usually, everyone sends their work to a central "Boss Server." The Boss mixes everything together and sends the result back. But this is risky (the Boss is a single point of failure) and often fails because the experts are so different from each other that mixing their notes creates a mess.
The Random Way (Decentralized): To avoid the Boss, the experts start talking directly to each other. But they do it randomly, like people at a party bumping into whoever is standing next to them. Sometimes they talk to someone who helps them; other times, they talk to someone who confuses them and makes their work worse.

Enter KNEXA-FL: The "Smart Matchmaker"

The authors of this paper introduce a new system called KNEXA-FL. Think of it as a Smart Matchmaker who runs a very special, secure matchmaking service.

Here is how it works in simple terms:

1. The "Profile" (No Secrets Shared)

Instead of sending their actual notes or code, each expert sends a tiny, anonymous "profile" to the Matchmaker.

What's in the profile? It says things like: "I am an expert in Python," "I am good at solving math problems," or "I am currently struggling with loops."
What's NOT in the profile? No actual code, no private data, and no secret formulas. It's just a summary of their skills and current state.

2. The "Smart Matchmaker" (The CPM)

The Matchmaker (called the Central Profiler/Matchmaker or CPM) doesn't just guess who should talk to whom. It uses a clever learning trick (called a Contextual Bandit, which is like a smart gambler who learns from every win and loss).

The Matchmaker looks at the profiles and asks: "If Expert A talks to Expert B, will they both get better?"
It learns over time which pairings work best. It avoids pairing two experts who will confuse each other and seeks out pairs where one can teach the other something new.

3. The "Secret Handshake" (Secure Exchange)

Once the Matchmaker decides who should talk, the two experts meet directly.

They don't swap their private notes.
Instead, they play a game of "Guess the Answer." One expert (the Teacher) solves a problem and says, "I think the answer is this." The other expert (the Student) listens and tries to learn how the Teacher thought about it, without ever seeing the Teacher's private notes.
This is called Knowledge Distillation. It's like learning a recipe by tasting the soup, rather than stealing the chef's notebook.

4. The Result: A Learning Loop

After they talk, the experts tell the Matchmaker: "That conversation was helpful!" or "That conversation was a waste of time."

The Matchmaker remembers this feedback.
Next time, it gets even better at pairing the right people together.

Why is this a big deal?

The paper tested this on a very hard task: generating code. They compared their system to three other methods:

Working Alone: Experts tried to learn by themselves. (Result: They got stuck).
Random Chat: Experts talked to whoever was available. (Result: They got better, but slowly and inconsistently).
The "Boss" Method: Experts sent everything to a central server. (Result: Disaster. The system collapsed and got worse because the different experts' styles clashed too much).

The Winner:
The KNEXA-FL system (Smart Matchmaker + Secure Chat) was the clear winner.

It improved the experts' performance by about 50% compared to random chatting.
It was stable, meaning it kept getting better without crashing, unlike the "Boss" method which failed completely.

The Bottom Line

This paper shows that you don't need a central boss to share knowledge safely, and you don't need to guess who to talk to. By using a smart, learning-based matchmaker that only sees "profiles" and not "secrets," a group of isolated AI models can learn from each other effectively, securely, and without ever breaking privacy rules. It turns a chaotic crowd of strangers into a highly efficient, collaborative team.

Technical Summary: KNEXA-FL

Problem Statement

The fine-tuning of Large Language Models (LLMs) for specialized domains faces a fundamental conflict: the necessity for diverse, cross-organizational data versus the strict requirements of data privacy, sovereignty, and security. While Federated Learning (FL) offers a paradigm for collaborative training without raw data exchange, its canonical centralized form introduces a trusted central aggregator. This entity acts as a single point of failure and a privileged attack surface for model inversion attacks, which can reconstruct sensitive training data.

Decentralized FL (DFL) mitigates these risks by removing the central server in favor of direct peer-to-peer (P2P) communication. However, existing DFL architectures typically rely on inefficient, random, or static peer pairing strategies. This "random-walk" approach is statistically inefficient, blind to agent heterogeneity, and risks "negative transfer," where poorly matched peers degrade each other's performance. The prevailing dichotomy between a vulnerable centralized aggregator and inefficient random P2P interaction limits the potential of collaborative AI.

Methodology: The KNEXA-FL Framework

The paper introduces KNEXA-FL, a novel framework for "orchestrated decentralization" that resolves the trade-off between security and efficiency. The system operates as a decentralized knowledge data space where agents remain sovereign, and knowledge is exchanged via secure distillation rather than raw data or model parameters.

System Architecture

The framework comprises three logical components:

LLM Agents: Autonomous entities holding a frozen base model ( $W_0$ ) and a trainable Parameter-Efficient Fine-Tuning (PEFT) module (primarily LoRA). Each agent operates a Local Gateway with a Guardrail Filter to prevent sensitive data disclosure.
Central Profiler/Matchmaker (CPM): A trusted, non-aggregating entity. Its sole function is to orchestrate P2P interactions. It receives abstract, privacy-preserving agent profiles but never accesses raw data or the models themselves.
Secure P2P Channels: Direct, ephemeral, and encrypted connections established between matched agents for knowledge exchange.

Core Mechanisms

Adaptive Knowledge Distillation (AKD): To handle architectural and tokenizer heterogeneity, KNEXA-FL uses text-based distillation. A "teacher" agent generates text predictions on a shared, privacy-vetted transfer set. The "student" agent re-encodes these predictions using its own tokenizer to create soft target sequences, minimizing token-level cross-entropy loss. This ensures compatibility regardless of model differences.
Contextual Bandit Matchmaking: The CPM formulates the P2P pairing problem as a contextual combinatorial bandit problem.
- Profiles: Agents submit abstract profile vectors ( $p_i$ ) containing static features (model family, PEFT config), dynamic features (task performance, perplexity), and historical trust features.
- Algorithm: The CPM employs the LinUCB algorithm to learn an optimal matchmaking policy. It models the expected reward of a pairing based on context vectors and selects disjoint pairs that maximize an Upper Confidence Bound (UCB) score, balancing exploitation of known good pairs with exploration of new potential synergies.
- Reward Signal: The reward is calculated based on the local loss reduction achieved by the student agent, penalized by communication costs. This allows the CPM to learn to down-weight low-utility or malicious peers over time.

Key Contributions

Orchestrated Decentralization: The paper proposes KNEXA-FL, a hybrid architecture that resolves the matchmaking inefficiency of DFL without reintroducing the security vulnerabilities of a central model aggregator.
Contextual Bandit Formulation: It is the first work to formulate P2P LLM collaboration as a contextual bandit problem. The CPM uses LinUCB to dynamically optimize the P2P knowledge graph based on interaction rewards, rather than relying on static heuristics or random selection.
Empirical Validation: The framework is validated on a challenging, heterogeneous code generation task, demonstrating substantial performance gains and stable convergence where other methods fail.

Experimental Results

Experiments were conducted on a federation of 6 heterogeneous clients (using models like Qwen, Cerebras-GPT, BLOOM, and Pythia) with non-IID data distributions derived from HumanEval and MBPP benchmarks.

Performance Gains: KNEXA-FL achieved a Pass@1 of 13.33%, representing a ~50% relative improvement over Random-P2P collaboration (8.89%) and a 100% improvement over a heuristic-based P2P baseline (6.67%).
Stability vs. Centralized Collapse: A strong centralized knowledge distillation baseline (Central-KD) suffered from catastrophic performance collapse, dropping from a peak of 18.33% to a final 2.00% Pass@1 due to destructive parameter interference and catastrophic forgetting. In contrast, KNEXA-FL demonstrated stable convergence.
Matchmaking Efficacy: The CPM's learned policy enabled a student to achieve 86.70% Pass@1 on a transfer set during collaboration, a 2.6x improvement over the best pairing found by random chance. This confirms the CPM actively discovers synergistic knowledge transfers.
Scalability: Synthetic experiments with up to 64 clients showed that the LinUCB-enhanced CPM scales robustly, with performance gains peaking at ~48.5% in high-heterogeneity settings.

Significance and Claims

The paper argues that adaptive, learning-based orchestration is a foundational principle for building robust decentralized AI ecosystems. By treating peer selection as a learnable problem rather than a static or random process, KNEXA-FL successfully navigates the trade-offs between diversity and compatibility.

The authors claim that their work establishes a new paradigm where:

Governance is emergent: Trust and utility are learned properties of the system, allowing the CPM to naturally deprioritize low-quality peers without static rule-based policies.
Privacy is preserved: The non-aggregating CPM and text-based distillation ensure that neither raw data nor model parameters are exposed, while still enabling effective knowledge transfer.
Heterogeneity is an asset: Unlike centralized approaches that struggle with statistical heterogeneity, the orchestrated P2P approach leverages diversity to achieve superior collective performance.

The work positions profiler-guided P2P learning as a robust alternative to both insecure centralized FL and inefficient random P2P systems, offering a path toward effective, privacy-preserving collaborative intelligence.