Research

research

Credibly Neutral AI Oracles

A credibly neutral AI report layer allows dispute mechanisms to use a higher voting threshold without sacrificing accurate resolution. The result is a system that is harder to manipulate and easier to verify.

research

The Role of Feedback Alignment in Self-Distillation

We propose step-aligned feedback for self-distillation: feedback that follows the solver's reasoning trace step by step. By anchoring training on reasoning rather than stylistic tokens, it resolves a central bottleneck of the method and outperforms the misaligned context types used in prior work.

research

Markets That Explain Themselves

We introduce evidence markets, a mechanism generalizing prediction markets by paying for both accurate forecasts AND the evidence that explains them. Evidence markets also generalize resolution by allowing the evidence gathered within a markets lifecycle to settle the market itself.

research

Look Beyond One-Size-Fits-All LLMs with IR3DE

We propose IR3DE, a lightweight router that directs prompts to the most suitable domain experts based on token embeddings. IR3DE enables the addition or removal of experts with minimal updates, avoiding the need to rebuild the router. As a result, it is highly suitable for decentralised settings.

research

DEI: Diversity in Evolutionary Inference for Quality-Diversity Search

Introducing DEI (Diversity in Evolutionary Inference), a distributed Quality-Diversity search framework that uses heterogeneous LLMs as complementary mutation operators. This allows nodes running different LLMs to collaborate and reach a solution which outperforms solutions by an individual LLMs.

research

F-TIS: Heterogeneous GRPO Without Homogeneous Assumptions

Our method, F-TIS, enables diverse models to collaborate in decentralised GRPO by combining truncated importance sampling with filtering. Across model-size, expertise, and PEFT heterogeneity, F-TIS matches on-policy convergence and, in some cases, improves out-of-distribution reasoning.

research

Backdoor in the Middle: Attacking Pipeline Parallelism

Most works on adversarial robustness in distributed training focus on data parallelism, e.g., poisoned gradients, malicious clients, or aggregation attacks. Pipeline parallelism has a different attack surface. Here, the model itself is partitioned across nodes.

research

Stop Multitask Training. Just DUME.

Dynamic Upcycling MoE (DUME) cleverly reuses dense experts trained on different domains to create a unified MoE multidomain expert model. DUME retains the knowledge of the original dense experts without any additional training, offering a cost-effective, scalable solution with no compromises.

article

Prediction Markets are Learning Algorithms

In this piece we’ll unpack this similarity and reveal that, in many cases, they are formally equivalent in a strong sense. We’ll discuss which classes of prediction markets are mathematically identical to standard online learning...