English

Artificial Intelligence research

What AI tools are, how they work, where your data goes, and how to use and build them responsibly at work.

Our articles on Artificial Intelligence →

Filters
Format
Peer review
Publisher
Added

688 entries

The summaries are open to everyone. Sign in, free, to read the full entry.

  • Preprint
  • Not peer reviewed

A preprint describes a distributed MoE redesign cutting cross-GPU communication

Original title: Federation of Experts: Communication Efficient Distributed Inference for Large Language Models

Researchers report an architecture that groups experts and KV heads to shrink communication overhead in distributed LLM inference.

Muhammad Shahir Abdurrahman et al. · Added

  • Preprint
  • Not peer reviewed
  • From Workato

Researchers report adaptive data scheduling boosts LLM reasoning RL training

Original title: Learning at the Right Pace: Adaptive Data Scheduling Improves LLM Reinforcement Learning

A preprint reports that scheduling training data by semantic clusters and policy difficulty improves reinforcement learning for LLM reasoning over uniform sampling.

Added

  • Preprint
  • Not peer reviewed
  • From Google DeepMind

Study links fine-tuning behaviour spread to a model’s ‘default persona’

Original title: The Persona Hierarchy Model: Understanding Contextual Generalization in Fine-Tuning LLMs

A preprint proposes that fine-tuned behaviours spread widely when training nudges a model's shared default persona rather than a context-specific one.

Added

  • Preprint
  • Not peer reviewed

A preprint reports 2-bit KV cache compression for looped transformers

Original title: ResidualQuant: KV Cache Quantization for Looped Transformers with 2-Bit Residuals

Researchers report a quantisation method that shrinks KV cache memory in looped transformers while keeping accuracy close to full precision.

Heejun Kim et al. · Added

  • Preprint
  • Not peer reviewed

Researchers trace LLM forgetting to output embeddings of rare tokens

Original title: A Deafening Silence: Catastrophic Forgetting Lives in the Output Embeddings of Tokens the Data Never Speaks

A preprint reports that catastrophic forgetting during fine-tuning concentrates in output embeddings of under-exposed tokens, and a simple optimizer tweak reduces much of it.

Jonghyun Han et al. · Added

  • Preprint
  • Not peer reviewed
  • From Kuaishou Technology

Researchers report a reward model that rivals test-based code RL training

Original title: CodeScaler: Scaling Code LLM Training and Test-Time Inference via Reward Models

A preprint describes CodeScaler, a reward model that matches or beats execution-based reinforcement learning for code generation while cutting inference latency.

Added

  • Preprint
  • Not peer reviewed

Preprint reports strategy-guided RL framework boosts long-horizon agent performance

Original title: StraTA: Incentivizing Agentic Reinforcement Learning with Strategic Trajectory Abstraction

A preprint introduces StraTA, which adds an explicit trajectory-level strategy to agentic reinforcement learning, reportedly improving efficiency and task success.

Xiangyuan Xue et al. · Added

  • Preprint
  • Not peer reviewed

Researchers report new bioengineering benchmark exposes gaps in top AI models

Original title: BioEVAL: A global, multi-institutional benchmark of large language and multimodal models for bioengineering

A multi-institutional preprint finds leading LLMs score well on bioengineering recall but struggle more with experimental reasoning and visual data.

Shun Ye et al. · Added

  • Preprint
  • Not peer reviewed

Preprint proposes dynamic budget allocation for testing when LLMs jailbreak or fail

Original title: How Many Iterations to Jailbreak? Dynamic Budget Allocation for Multi-Turn LLM Evaluation

Researchers report a method that adaptively spends a fixed interaction budget to more reliably bound how many turns it takes for an LLM to jailbreak or complete a task.

Shai Feldman et al. · Added