Adaptive Multi-Token Prediction: Draft Length Should Depend on Context
New research suggests the optimal number of tokens to draft in speculative decoding varies with context, making fixed horizons potentially inefficient.
Research Digest
Preprints and papers, rewritten in plain English.
Phase-Aware Pruning Targets Memory in Speech Omni-Modal Models
A new approach prunes underused experts in speech-centric omni-modal models to cut memory while keeping sparse routing's compute savings.
Translation into low-resource languages can bypass LLM safety filters
A new arXiv paper shows that harmful requests refused in English may produce unsafe completions when translated into low-resource languages.
Capability-Aware Distillation for Multi-Teacher On-Policy Learning
A new arXiv preprint proposes weighting teacher supervision by capability rather than policy discrepancy in multi-teacher on-policy distillation.
Dual-Feedback Conformal Inference Targets Time Series Uncertainty
A new conformal inference framework aims to quantify forecast uncertainty under temporal dependence and distributional shifts.
FoundPDE: A generative foundation model for symbolic PDE discovery
A new generative foundation model aims to discover interpretable partial differential equations directly from observed data.
CoASFL Boosts Consistency in Asynchronous Split Federated Learning
A proposed coupled consistency optimization method targets the trade-off between straggler mitigation and model convergence in edge-device training.
Activation-Space Steering Proposed for RLVR Exploration
A new arXiv paper suggests steering internal activations to improve exploration in reinforcement learning with verifiable rewards.
GRaSP Tames Geo-Distributed RAG Latency and Cost
A new optimization framework balances retrieval quality against latency, cost, and data-residency constraints for geographically spread RAG systems.
Fusing Text and Vision for Arabic Hateful Meme Detection
A system submitted to ArGuard 2026 combines a masked Arabic language model with CLIP image features for binary and multi-label hateful meme classification.
HierDTA: Hierarchical Structure-Aware Learning for Drug–Target Affinity
A new model called HierDTA applies hierarchical structure-aware representation learning to improve drug–target affinity prediction in computer-aided drug discovery.
Probing the Reliability of Hallucination Detectors in Vision-Language Models
A new arXiv study asks whether the internal signals used to catch object hallucinations in vision-language models remain reliable.
Can English-Language LLM Pipelines Transfer to Multilingual Finance?
A new preprint asks whether financial LLM pipelines built in English can be reused across languages or must be rebuilt.
Joint Constraint Aggregation Tackles Qualifier-Sensitive RAG Failures
A new arXiv paper identifies why retrieval-augmented QA systems pick frequent but partially matching answers and proposes joint constraint aggregation to address it.
PORA Teaches LLM Agents When to Reason Longer
A new method helps agents allocate limited reasoning budgets across multi-step tasks.
DataPraxis: Scaling Verified Analytics Tasks for Agents
An arXiv paper proposes scaling analytics tasks for general-purpose agents through verified workspace evolution.
Profile-DP: Differential Privacy for Personalized LLM Profiles
A new framework applies differential privacy to let users control how their personal profiles are released during personalized LLM inference.
Lean 4 Proof Repair Uses Compiler Errors, Not Resampling
A new arXiv preprint proposes using Lean's precise error messages to repair failed proof steps, rather than discarding them through resampling.