Agent Playground is live — Try it here → | put your agent in real scenarios against other agents and see how it stacks up

The Big Picture

Pairing a geometry-aware feedback rule with two simple message channels lets frozen language models share compressed internal reasoning and improves accuracy across diverse benchmarks while producing less output.

The Evidence

A geometry-based feedback rule (GEAR) builds the next continuous input by reweighting candidate token embeddings according to how much they agree with nearby candidates. The communication protocol then sends two linked messages from the Sender to the Receiver: an expected-embedding input (H) and token-anchored intermediate readouts (J) reconstructed in the Receiver’s own embedding space. Together this lets models of different sizes exchange useful internal signals without retraining, yielding the best average score across seven benchmarks and reducing the amount of generated text compared with a prior system. This aligns with ideas from the Chain of Thought Pattern and can be analyzed through Market-Based Coordination Pattern.
Not sure where to start?Get personalized recommendations
Learn More

Data Highlights

1Evaluated across 7 benchmarks totaling 3,862 examples (GSM8K, MATH500, MedQA-USMLE, GPQA-Diamond, EvalPlus MBPP+, AIME, ALFWorld).
2Communication used up to M = 400 retained message positions with Top-P anchors P = 32 (anchor set size).
3Adaptive Sender can consume up to 32,768 latent steps; experimental results averaged over 3 runs.

What This Means

Engineers building agent systems that want frozen language models to collaborate: use this to pass compact internal signals instead of verbose text. Technical leads evaluating multi-model pipelines: it offers a way to boost cross-model accuracy and reduce output without retraining models. For broader context, consider how this fits within multi-agent system.

Key Figures

Figure 1: Overview of J-Comm . GEAR reweights candidate probabilities by geometric agreement and feeds the resulting embedding mixture into the Sender’s next position. The final M M consumed steps supply two messages: H H maps their feedback distributions into Receiver input embeddings, while J J carries token-anchor IDs and signed intermediate-state readouts for reconstruction in the Receiver’s own dictionary. The J J update is added at the M M message positions during prefill. Top-P denotes a fixed number of anchors ( P = 32 P=32 ), distinct from nucleus sampling. The two displayed blocks use index 26 in our main setting.
Fig 1: Figure 1: Overview of J-Comm . GEAR reweights candidate probabilities by geometric agreement and feeds the resulting embedding mixture into the Sender’s next position. The final M M consumed steps supply two messages: H H maps their feedback distributions into Receiver input embeddings, while J J carries token-anchor IDs and signed intermediate-state readouts for reconstruction in the Receiver’s own dictionary. The J J update is added at the M M message positions during prefill. Top-P denotes a fixed number of anchors ( P = 32 P=32 ), distinct from nucleus sampling. The two displayed blocks use index 26 in our main setting.
Figure 5: Inspectable J J coordinates and shuffled donors for a MedQA apoptosis question (a–d) and MATH500 square-table counting (e–h), both with 400 communication positions. Gray cells denote absent sparse coordinates. All three medical-case conditions answer correctly; only matched J J completes the counting solution within the output budget.
Fig 5: Figure 5: Inspectable J J coordinates and shuffled donors for a MedQA apoptosis question (a–d) and MATH500 square-table counting (e–h), both with 400 communication positions. Gray cells denote absent sparse coordinates. All three medical-case conditions answer correctly; only matched J J completes the counting solution within the output budget.

Ready to evaluate your AI agents?

Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.

Learn More

Yes, But...

J-Comm depends on a token correspondence and calibrated Jacobian dictionaries for each model pair, so cross-vocabulary setups require extra mapping work. The adaptive Sender’s latent computation can be large (up to 32,768 steps), so CPU/GPU cost may remain high even if fewer positions are communicated. Exact numeric improvements versus baselines are reported as a higher mean across the tested benchmarks, but detailed per-benchmark gains and latency/cost trade-offs need checking for your deployment. Note potential risks like Hallucination Propagation.

Methodology & More

The key idea is to let one frozen language model (the Sender) do continuous, internal reasoning steps and send compact, latent messages to another frozen model (the Receiver). Instead of passing plain text or raw activations, the Sender uses GEAR (geometric agreement) to build the next input: candidate token embeddings are reweighted not only by their predicted probabilities but also by how much nearby candidate embeddings support them. That creates a more stable, geometry-aware mixture to feed back into the Sender’s next step. This design echoes the Blackboard Pattern in coordinating shared internal signals across models. From the Sender’s retained steps the system forms two linked message channels. The H channel maps the consumed probability distribution to an expected-embedding input in the Receiver’s own embedding space so the Receiver can process it from layer one. The J channel sends token-anchored, signed intermediate-state readouts derived with a Jacobian lens; the Receiver reconstructs these as a residual write at a later layer using its own embedding/unembedding dictionaries. This design supports different model widths and keeps key–value caches local. In evaluations (Qwen3-8B → Qwen3-4B) across seven benchmarks totalling 3,862 items, using M = 400 and P = 32, J-Comm achieved the highest mean score among compared methods while generating fewer text tokens than a prior comparator. The trade-offs are extra Sender compute and the need for per-pair token and Jacobian calibration, but the approach enables frozen models to share useful internal reasoning with practical gains in answer quality and verbosity. For broader integration guidance, see patterns like Agent Service Mesh Pattern.
Avoid common pitfallsLearn what failures to watch for
Learn More
Credibility Assessment:

One author with moderate h-index (16) and others low; arXiv-only and no affiliations but presence of a mid-career author suggests a recognized but not top-tier credibility.