La Nuova AI “Telepatica” che Sta Shockando i Ricercatori

Overview

The video explains a new “telepathic” multi-agent AI approach from the research paper Recursive Multi-agent System (name as stated in the subtitles). In this method, agents communicate without passing text tokens. Instead, they exchange latent vectors that represent internal reasoning/concepts directly through intermediate representations inside the models.

---

Core Idea / Intuition (Why “Telepathy”)

• In typical multi-agent setups, agents communicate by sending text back and forth (tokens).
• The creator argues this is inefficient because an LLM already contains latent vectors that encode internal meaning and reasoning.
• The researchers propose that agents should communicate by passing these latent vectors directly, avoiding the costly and information-losing conversion step:
  • concepts → text tokens → other model

---

How the System Works (Conceptual Mechanism)

Per-agent behavior

Each agent:

• receives input tokens from the user (context/instructions/questions),
• then instead of outputting text for the next agent, it extracts latent vectors from “inner” layers.

Communication via an outer connector/module

Subtitles reference an “Outer Link” and Sim(s):

• the sending agent produces a vector,
• the connector transforms/aligns it,
• the receiving agent ingests it as part of its input representation.

Recursive multi-agent loop

The system can run an iterative recursive loop among multiple agents (e.g., critic/solve/other roles), continuing until:

• a max recursion depth is reached, or
• one agent decides to verbalize the final answer as text.

---

Reported Results / Metrics

Compared to a similar recursive approach using text/token-based communication:

Token efficiency (token savings)

• 1 recursion round: 34.6% fewer tokens
• 2 rounds: 65.5% fewer tokens
• 3 rounds: 75.6% fewer tokens

Speed (inference time)

• up to 2.4× faster (3 rounds)
• 1.9× (2 rounds)
• 1.2× (1 round)

Performance (benchmarks)

• about +8.1% average improvement on benchmarks (as stated in subtitles) versus other setups (including the token-recursive approach and other comparisons like mixed/single-agent approaches).
• Claim: exchanging richer latent information (instead of text) improves accuracy.

---

Training Cost / Practicality

• No need to train the full LLM(s).
• Only trains connector parameters (“Sim”/links) between agents.
• Subtitles claim connector training is very cheap (numbers mentioned: $4.27 and/or ~13 million parameters).
• Claimed benefit: cheaper than full supervision or fine-tuning of large models.

---

Model Flexibility

Supports heterogeneous models

• Different model architectures and sizes can be connected.

Examples mentioned (as named in subtitles)

• Gemma 3 variants with Llama 3.2
• Ken-series models with Minimax/GLM/Kimi (as named in subtitles)

Collaboration patterns (described in the video)

• sequential
• mixture/distillation
• deliberation-style

---

Code / Model Availability

Subtitles claim the code and models used are public/open, implying the approach can be reused.

---

Tutorial: How Connectors Are Trained (Step-by-Step)

The video outlines alignment for two models (example given: Llama and Ken):

1. Prepare a dataset

   • tasks (e.g., math problems) with inputs and output answers

2. Choose internal layers for attachment

   • identify layers around embeddings/softmax regions
   • determine input/output embedding points where vectors are exchanged

3. Collect vectors

   • run the input through model A and save a vector from its chosen layer
   • run the same input through model B and save the corresponding target vector

4. Train the connector module (small network)

   • map model A’s vector space to model B’s vector space
   • alignment objective uses a distance metric (subtitles mention Euclidean distance, i.e., making vectors “as close as possible”)

5. Train the reverse direction connector

   • align from B → A as well

6. Preliminary inner-loop training

   • freeze the main LLM parameters
   • train only connector parameters first

7. Recursive outer-loop training

   • connect agents together
   • since vectors are differentiable, use backpropagation mainly through connector/link parameters (and possibly LoRA parameters, per subtitles)

8. Stopping conditions

   • stop when loss drops below a threshold, or
   • stop after a fixed number of recursive rounds (forcing verbalization if needed)

---

Main Speakers / Sources (As Stated)

• Video narrator/speaker
  • the creator (who says they “read it all” and made the video to explain the intuition)
• Research teams/affiliated labs credited in subtitles
  • Stanford, NVIDIA, MIT
• Referenced paper
  • “Recursive Multi-agent System” (spelled/auto-captioned ambiguously in subtitles)