When Value-Aware KV Eviction Boosts Cache Compression
Explore how value-aware KV eviction improves cache compression and task accuracy in long-context language models with a fixed-contract diagnostic.
TRAM: Low-Power Approximate Multipliers for AI Accelerators
Discover TRAM, a method for training approximate multipliers that reduces AI accelerator power by 25%+ while maintaining model accuracy.
Provenance-Aware Pipeline for Historical Tables to Knowledge Graphs
Transform handwritten historical tables into accurate knowledge graphs with a modular, provenance-aware AI pipeline ensuring transparency and traceability.
NoisyCoconut: Boost LLM Reliability with Latent Space Noise
Enhance large language model accuracy using NoisyCoconut's inference-time latent space noise for reliable AI without retraining or data changes.
Bangla-WhisperDiar: Enhanced ASR & Speaker Diarization
Discover Bangla-WhisperDiar, a fine-tuned model improving Bangla long-form speech recognition and speaker diarization with advanced training techniques.