Telos papers
Preliminary reports on masked diffusion, scaling, and code prediction.
For all Wing It projects, visit Wing It Research ↗
télos (τέλος): Preliminary Report 2 — Capability Divergence, Positional Encoding, and the Limits of Masked Diffusion Scaling
Completes the scaling sweeps across 12.5M, 25M, and 50M parameter tiers, identifying a fundamental limitation in MDLM capacity allocation (CE-Rank Divergence) and evaluating the impact of Rotary Positional Embeddings (RoPE).
Read paper →Preliminary Report #1 · August 2026 · Ivan Samueltélos (τέλος): Preliminary Evidence on Token-Budget Scaling in Masked Diffusion Language Models for Code Autocomplete
An empirical investigation into the optimal training data volume for Masked Diffusion Language Models (MDLMs) across three parameter scales (12.5M, 25M, and 50M). Provides preliminary evidence that optimal token budgets grow with parameter count in discrete diffusion.
Read paper →