adaptive-inference

MSDNet

Spatially Adaptive Inference with Stochastic Feature Sampling and Interpolation, ECCV 2020 Oral

[NeurIPS'24] Training-Free Adaptive Diffusion with Bounded Difference Approximation Strategy

[ICML 2026] MoSE: Mixture of Slimmable Experts for Efficient and Adaptive Language Models

This repository features RARE-UNet a resolution-aware 3D U-Net for adaptive medical segmentation. It uses multi-scale entry blocks and resolution-based routing to dynamically adjust the inference path to input resolution. Combined with consistency-based training, RARE-UNet delivers accurate, efficient segmentation across resolutions.

Experimental HRM fork exploring more intuitive task-aware reasoning with meta-learning, latent context modeling, dynamic module weighting, and a cleaned training/evaluation workflow.

Scalable framework for adaptive LLM serving: classify prompt complexity → select quantized drafts → verify with FP16 target, no model retraining required.

This repository contains the code for comparing the performance of various UCB algorithms on ADNNs

Adaptive LLM inference control with safe live policies and evidence-grade evaluation

Selective verification for budget-aware LLM reasoning, with reusable routing, gate training, and policy evaluation.

Reproducible PyTorch study of learned normal, width, and depth routing for conditional compute

DoT-inspired Qwen/DeepSeek routing with dependency DAGs, alpha-tree labels, and AdapterMLP.

Research prototype: category-aware layer skipping for small LLMs on Apple Silicon. Honest write-up of what worked (1.9x batch throughput on Llama 3.2 3B) and what didn't (MMLU collapse). Outperformed by LayerSkip/MoD/vLLM.

PyTorch implementation of "Train for the Worst, Plan for the Best." Investigating adaptive token ordering in Masked Diffusion Models (MDMs) to sidestep hard subproblems and elicit reasoning in discrete domains.

面向3D空间理解的自适应推理MLLM

A no-signal baseline for visual escalation: reproducible tests of routing signals, visual budgets, and real per-query costs across VLMs.

Evidence-first audiovisual event inference with adaptive local-VLM escalation and grounded answers

Adaptive inference control plane for AI apps and agents

This is a small, working extraction of the two ideas from the "SCCS" document that actually hold up, rebuilt from scratch with real math and real measured numbers (not the invented architecture, file names, or formulas from that document).