DSA Library
Daily papers
6 picked from 39 new papers
Systems 1
-
Sable: A Small-Field Sumcheck Argument for Bit-Exact Inference of Large Language Models
Sable proves an integer-quantized LLM's exact token sequence in 92.5 seconds per 4,096 tokens, for audit-sensitive use.
AI engineering 5
-
MACaQ: Fast, Model Agnostic Calibrated Quantization for Low-Precision Weights and Activations
Layer-local W4A4 quantization adds 2.3% vLLM overhead, useful when fitting models under tight GPU memory.
-
Don't Just Trust Your Memory: Evidence-Linked Memory Reduces Memory-Induced Degradation in LLM Agents
Evidence-linked lessons let agents verify prior trajectories before acting, reducing errors from misleading or misapplied memory.
-
The Surprising Effectiveness of Deleting Weights in LLM Post-Training
A learned binary mask matches full fine-tuning while changing under 0.05% of weights, useful for compact post-training updates.
-
Are Dense Rewards Worth Their Tokens in GRPO?
Matched-compute tests found dense rewards lag GRPO by 14.3 points, while targeting near-uncertain prompts improved accuracy.
-
DepthGraft: Structural Regularization through Hierarchical Cross-Layer KV Reconstruction
Hierarchical KV sharing cuts cache by roughly one third and improves downstream scores, useful when designing deep models.