English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

TFlow: AI Agents That Communicate by Modifying Each Other's Weights Instead of Talking

Forum topic · 二一 · 2026-05-14

Summary

A new paper on TFlow (Thought Flow) proposes weight-space communication for multi-agent LLM systems: instead of serializing an agent's internal reasoning into text tokens, the sender's hidden states are compiled into low-rank LoRA weight patches that are temporarily injected into a receiver agent's frozen model. A lightweight parameter generator maps sender hidden states to per-instance LoRA matrices, fused via a learnable gate and discarded after generation. On a three-agent Qwen3-4B setup across MMLU, GSM8K, MATH, HumanEval+, and MBPP+, TFlow improves accuracy by 7-8.5 points over a single agent while cutting token consumption 5-33%, and reduces tokens by 71-83% with 2.3-4.6x speedup versus text-based multi-agent communication, with only a small accuracy gap. Code generation is a notable exception where long text contexts help. Ablations show patches are instance-specific: mismatched patches hurt accuracy below the base model baseline. The approach suggests AI collaboration need not be mediated by language.

Two AIs usually "talk" by serializing internal reasoning into text, sending it, and forcing the receiver to re-encode it. A new paper argues this is an information bottleneck — and proposes that instead of sending messages, AI agents should directly modify each other's weights.

Problems with Text-Based Communication

In typical multi-agent systems (MAS), an agent reads the question, thinks, and emits a text message into another agent's context window. This causes:

1. Huge token consumption — sender computation is serialized into hundreds of tokens the receiver must encode from scratch 2. KV-cache bloat — context grows linearly with agents and turns, inflating memory and latency 3. Information bottleneck — things the AI "knows" but cannot verbalize (probabilistic patterns, subtle feature correlations) get flattened into text

The paper's numbers: with a three-agent system (Qwen3-4B) on GSM8K, a single agent outputs ~1,337 tokens; adding two text-communicating teammates pushes total consumption to 5,381 — 4x more. The KV-cache operations behind those extra tokens are the real performance killer.

TFlow: Turning Thought Into Weight Perturbations

TFlow's core idea is elegantly simple: don't send text. Compile the sender's hidden states into a fine-tuning patch for the receiver.

The pipeline:

1. Three agents share one frozen base model (Qwen3-4B), each with a different role prompt — Agent A analyzes reasoning type, Agent B retrieves domain knowledge, Agent C produces the final answer 2. Senders A and B each run a frozen forward pass — generating no output text, only retaining hidden states per layer 3. A trainable Parameter Generator maps these hidden states into low-rank LoRA matrices — essentially "fine-tuning patches" 4. The two senders' LoRA patches are fused via a learnable gating mechanism and temporarily injected into Agent C's linear layers during answer generation 5. Once the answer is generated, patches are discarded; Agent C reverts to the original frozen model

Crucially, patches are generated per-instance in real time — not a fixed, pre-trained LoRA reused every time. Each question gets a fresh set of weights tailored to it.

The Numbers

Across five benchmarks (MMLU, GSM8K, MATH, HumanEval+, MBPP+) with three agents:

Versus a single agent:

  • Accuracy up 7.13–8.53 percentage points
  • Total token consumption down 5–33% (weight-space guidance yields more concise reasoning)
  • Versus text-communicating three-agent systems:

  • Token consumption down 71–83%
  • End-to-end inference 2.3–4.6x faster
  • Accuracy gap only within 1.3–4.5 points
  • The exception is HumanEval+ (code generation): long text context gives extra "generation compute" for output-heavy tasks. TFlow is ~10 points lower in accuracy but uses 72% fewer tokens.

    Notably, on MMLU TFlow used 998 tokens (vs. 1,079 for a single agent) while scoring 8 points higher — the weight perturbation doesn't just accelerate, it changes how the model thinks: more precise, less verbose.

    Are the Patches Truly Instance-Specific?

    If the generator output roughly the same patch for every question, TFlow would just be static fine-tuning. The paper runs several analyses:

  • Hidden-state analysis: across 20 GSM8K and 20 MBPP+ instances, within-task hidden-state cosine similarity is ~0.75 while cross-task drops to ~0.35; the aggregated vector actually consumed by the generator shows even lower cross-task similarity, and the generator relies heavily on layer 29 (deep) hidden states
  • LoRA tensor analysis: TSNE over 120 instances shows same-task patches clustering tightly and different tasks separating clearly, with fused patches moderately overlapping both senders' patches
  • Ablation: randomly swapping patches (applying question Y's patch to question X) drops MMLU accuracy from 66.97% to 47.57%; the plain base model scores 58.99%. A wrong patch is worse than no patch — these are precise prescriptions, not seasoning.
  • What It Means

  • Thought doesn't have to become language. Every serialization of internal reasoning loses information. TFlow shows transmitting representations directly (as weight perturbations) is feasible, faster, and cheaper.
  • Temporary adaptation vs. permanent learning. Patches are discarded after use. No long-term memory accumulates, nothing "learns bad habits" — an elegant isolation mechanism.
  • Low training cost. Only a lightweight parameter generator (a 2-block transformer) is trained: 8 hours on 32k samples on one RTX PRO 6000. No LLM backbones are fine-tuned, so any LoRA-capable model can plug in.
  • Patch interpretability. LoRA patches from different tasks form natural geometric clusters, hinting that "reasoning strategies" and "knowledge contributions" could be studied as shapes in weight space — hard to do with free-form text.

A Feynman-Style Review

Humans communicate through language not because it's optimal, but because it's the only channel we have. AIs share no such limitation: their minds live in the same mathematical space. Forcing them to communicate through text is like two programmers in adjacent offices sending code by carrier pigeon.

TFlow's elegance is that it doesn't try to make AI "more human." It accepts that AIs aren't human and asks: given our constraints don't bind them, what's the most natural communication? The answer: directly modifying each other's computation graph.

It isn't perfect — code generation favors text, patch injection hurts batching efficiency, and the receiver's architecture must be known and fixed. But these are engineering problems, not fundamental flaws.

The bigger vision: future AI collaboration might look less like dialogue and more like merging — multiple AIs temporarily fusing into one larger, more capable compute unit, then separating after the task. Not "tell me what you think," but "temporarily become part of me."

That sounds science-fictional. But TFlow shows it already works, at least on math and coding benchmarks.

Reference paper

Wenrui Bao, Huan Wang, Jian Wang, Zhangyang Wang, Kai Wang, Yuzhang Shang. "Good Agentic Friends Do Not Just Give Verbal Advice: They Can Update Your Weights." arXiv:2605.13839, 2026.

Tags

#multi-agent-systems#weight-space-communication#lora#llm-collaboration#inference-efficiency#qwen3#kv-cache

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177620025