inference-time

awesome-LLM-controlled-constrained-generation

Prompt level, inference time, AI frameworks to improve safety and alignment.

RLG: Inference-Time Alignment Control for Diffusion Models with Reinforcement Learning Guidance

[ICML 2026] Context-Robust Remasking for Diffusion Language Models

EAGer: Entropy-Aware GEneRation for Adaptive Inference-Time Scaling

Hands-on tutorials on training-free alignment of language models at inference time.

Official implementation of SEER: a self-grounded evidence interface for controlled spatial relation classification.

A comparative study of Faster R-CNN and YOLOv5 on Pascal VOC 2012, analyzing mAP, speed, and detection quality to understand the trade-offs between accuracy and real-time performance in object detection models.

Official EMNLP 2026 implementation of GGSS: inference-time demographic debiasing for generative VLMs/MLLMs via norm-preserving, token-level geodesic activation steering. No retraining.

LAteNT v2 — A 9-agent neuro-symbolic manifold for zero-shot abstraction. This system replaces hardcoded DSLs with a 64-dimensional Latent Transformation Space, implementing autonomous Bayesian Meta-Learning and online dictionary learning to discover causal laws purely from observation. Pure Inductive Intelligence.

BALM: Bias-Aware Language Model with inference-time bias detection and correction.

TACT: signed, label-free confidence weighting for self-consistency voting — with the thin-window boundary (2.5–7.5% of items) that explains why six other designs died

How much of an LLM agent’s apparent behavior comes from the decoder rather than the model alone? SamplerScope holds the model’s logits fixed, changes only the decoding method, and measures how this alters outcomes in small decision environments.

Experimental local assistant runtime for GGUF models that steers token generation with activation perturbations and verbal control loops for self-correction, continuity, and future memory-driven support.