Here’s this week’s Ritual Research Digest, a newsletter covering the latest in the world of LLMs and the intersection of Crypto x AI.
With hundreds of papers published weekly, staying current with the latest is impossible. We do the reading so you don’t have to.
DeepSeekMath-V2: Towards Self-Verifiable Mathematical Reasoning
This paper introduces DSMath-V2, a model trained on deepseek-3.2-exp for natural language proving in mathematics. The generation-verification gap is a major hurdle for informal proving.
They first train a verifier for the model using expert annotations to assess both the correctness of the answers and the analysis.
This verifier is used to train the final prover model, which both writes proofs and analyzes their correctness.
They achieve gold in IMO 2025.
Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond)
The paper introduces INFINITY-CHAT, a dataset of 26K real-world queries that accept multiple answers. Using this, they study intra- & inter-model mode collapse in 70+ LMs.
They uncover an Artificial Hivemind effect with intra-model repetition, where a model repeatedly generates similar outputs, & inter-model homogeneity, where different models converge on similar ideas with minor phrasing changes. This raises questions about model diversity.
Latent Collaboration in Multi-Agent Systems
The work introduces Latent MAS, an end-to-end collaborative framework that operates in continuous latent space. The design integrates both latent thought generation and cross-agent latent memory transfer.
LatentMAS is based on reasoning expressiveness, communication fidelity and collaboration complexity. Across both sequential and hierarchical MAS settings, Qwen 3(4B, 8B, and 14B), LatentMAS outperforms text-based MAS baselines improving accuracy, & reducing output token usage.
ToolOrchestra: Elevating Intelligence via Efficient Model and Tool Orchestration
The authors propose an orchestration paradigm where intelligence emerges from a composite system. An orchestrator model invokes the right tools in right order for a task.
Using ToolOrchestra, an 8B model is trained with RL to decide when and how to invoke other LMs and tools. The rewards balance correctness, efficiency and alignment with user preferences.
On HLE, Orchestrator outperforms prior methods with far lower computational cost.
Follow us @ritualdigest for more on all things crypto x AI research, and
@ritualnet to learn more about what Ritual is building.
471
3
Innholdet på denne siden er levert av tredjeparter. Med mindre annet er oppgitt, er ikke OKX forfatteren av de siterte artikkelen(e) og krever ingen opphavsrett til materialet. Innholdet er kun gitt for informasjonsformål og representerer ikke synspunktene til OKX. Det er ikke ment å være en anbefaling av noe slag og bør ikke betraktes som investeringsråd eller en oppfordring om å kjøpe eller selge digitale aktiva. I den grad generativ AI brukes til å gi sammendrag eller annen informasjon, kan slikt AI-generert innhold være unøyaktig eller inkonsekvent. Vennligst les den koblede artikkelen for mer detaljer og informasjon. OKX er ikke ansvarlig for innhold som er vert på tredjeparts nettsteder. Beholdning av digitale aktiva, inkludert stablecoins og NFT-er, innebærer en høy grad av risiko og kan svinge mye. Du bør nøye vurdere om handel eller innehav av digitale aktiva passer for deg i lys av din økonomiske tilstand.