adaptive-computation

Mixture-of-Recursions: Learning Dynamic Recursive Depths for Adaptive Token-Level Computation (NeurIPS 2025)

[ICLR 2025] Dynamic Mixture of Experts: An Auto-Tuning Approach for Efficient Transformer Models

A curated reading list of research in Adaptive Computation, Inference-Time Computation & Mixture of Experts (MoE).

Exploration into the proposed "Self Reasoning Tokens" by Felipe Bonetto

Yet another random morning idea to be quickly tried and architecture shared if it works; to allow the transformer to pause for any amount of time on any token

Dynamic weight generation for recursive transformers via input-conditioned LoRA modulation

Lev: Thinking Fast and Deep with Looped Decision Models. One checkpoint, multiple depths: System 1.5 decision models that think in loops, not tokens.

(NeurIPS-2019 MicroNet Challenge - 3rd Winner) Open source code for "SIPA: A simple framework for efficient networks"

LLM reasoning: use draft agreement to route queries and allocate adaptive thinking budgets without training.

Temporalmesh-transformer. It is the first architecture to simultaneously fuse dynamic graph topology, token-level adaptive compute, and temporal semantic decay into a single unified model. No prior work does all three together.

The ARL Hierarchical MultiScale Framework (ARL-HMS) is a software library for development of multiscale models on heterogeneous high-performance computing systems.

[NeurIPS 2026] To Think or Not to Think: Pre-Decisional Reasoning Budgets for Referring Audio-Visual Segmentation

Jev for MoR and MoE: context-complete Canopy-Jev computation plans by Ruoxi Qiu.

Frozen KV Context for Mixture-of-Recursions on a Modernized BERT

Research prototype: hierarchical module/expert routing and adaptive depth in a small byte-level language model (preliminary results, under audit)

Recursive Convergent Inference — dynamic MoE with convergence-gated stopping. Unexpected finding: model-relative complexity diverges from human difficulty labels

Curated papers on dynamic neural networks, pruning, growing architectures, sparse training, and adaptive computation.

Token-level early-exit Transformer with Gumbel-Softmax routing and a latency-aware joint loss — skips redundant layers for "easy" tokens while going deep on hard ones.

Lightweight PyTorch implementation of Mixture-of-Recursions with Expert-Choice & Token-Choice routing | Runs on your laptop!

A Universal Transformer that solves 4x4 Sudoku by repeatedly applying one shared block and learning when to stop