KPSA Offers Amortized Input-Specific Parameter Attribution
The proposed KPSA method targets the trade-off between costly input-specific gradient attribution and cheap but input-independent importance scores.
Research Digest
Preprints and papers, rewritten in plain English.
RAMP Cuts Recurrent-Depth LM Inference Overhead with Adaptive Precision
A new method applies adaptive mixed precision to the prefill phase of recurrent-depth language models, targeting the extra cost of recurrent computation.
Trajectory-Aware RL Method Addresses State Distribution Shift in LLMs
A new policy-optimization method targets the compounding state-distribution mismatch that undermines long-horizon reinforcement learning for LLMs.
CORAL-KV shrinks video diffusion KV caches via cross-layer prediction
A new compression scheme reduces the memory footprint of attention key-value caches, potentially enabling longer generated videos.
Benchmark Gains Don't Equal Real-World Deployment, arXiv Paper Argues
A new arXiv paper contends that AI's strong benchmark results have not led to meaningful professional deployment, and identifies a cause that the abstract leaves incomplete.
Baseline Assessment Proposed for Agentic Web Stack
A new arXiv paper argues for a baseline assessment to help developers navigate the fast-growing landscape of LLM agent protocols and frameworks.
DP-MIND: A Differential Privacy Approach for Long-Context LLM Inference
A new arXiv paper proposes a differentially private memory-guided method to reduce privacy risks in LLM-as-a-Service.
Half-Precision PPO Gets Structure-Aware Stabilization
A new arXiv paper targets numerical stability and efficiency for half-precision proximal policy optimization on resource-constrained platforms.
Adapting Multi-Agent Collaboration to LLM Competence
New research proposes tailoring collaboration among LLM agents to the backbone model's actual capabilities rather than assuming uniform performance.
Adaptive DHiSparse eases KV memory pressure for long-context models
The proposed method adjusts the KV cache budget per load, addressing a key limitation of prior HiSparse.
Semantic Query Engines Vary Run-to-Run Even at Temperature Zero
A new arXiv study measures how much semantic query engine outputs change between executions, questioning the reliability of current benchmarks.
PRIME: A Pre-Retrieval Index Model for Structured Document Search
A new arXiv preprint proposes improving LLM agent retrieval by enhancing the index side of structured document search, rather than only tuning query-side retrievers.
Branching On-Policy Distillation to Improve Teacher Supervision Coverage
The paper identifies a coverage limitation in on-policy distillation and proposes branching to address it, though the abstract excerpt cuts off before presenting results.
Model Looks Ahead and Back to Generate Road Trajectories
A new generative model for road-network trajectories anticipates both future and past context to improve realism in urban simulation.
Medium-Difficulty Samples Key in GRPO Curriculum Learning
A new arXiv paper examines how curriculum difficulty progression affects GRPO training, highlighting the value of medium-difficulty samples for group-relative learning signals.
Knowledge Graph-Guided Pre-training for Legal Reasoning
A new multi-task pre-training approach targets structural reasoning in semi-structured legal documents.
CASCADE: Agentic Framework Predicts Gene Perturbation Effects
A new arXiv paper introduces CASCADE, an agentic framework that uses precomputed ARACNe regulatory networks to predict downstream transcriptional effects of gene perturbations.
Does Learning Recurrent Networks Matter in Partially Observable RL?
A new arXiv paper questions the necessity of training recurrent modules jointly with policies in partially observable reinforcement learning.