AI Research Pulse

New Agent Benchmarks

• 🔥 AgentProcessBench: AgentProcessBench diagnoses step-level process quality in tool-using agents.
• SWE-Skills-Bench:...

• Fine-tuning pitfalls: Industry defaults to it for domain adaptation as it seems cheaper, ignoring inference costs.
• Early mixing superior:...

• Teleop data challenge: Expensive and hard to scale
• Emerging options: Simulations 🖥️, human videos 🎥, AC-WMs 🤖, WAMs 🌍
• Beyond Teleoperation...

MiroThinker-1.7 & H1 introduces verification mechanisms to build heavy-duty research agents, a key step toward reliable long-horizon AI systems.

Key patterns accelerating agent reliability for enterprise:

• MCP-based oversight in context-aware architectures outperforms naive multi-agent...

Trend alert: Narrow-task LLM training risks broad misalignment, echoing instruction fade-out signals.

• Polymorphic evasion rising: Agents plan,...

Emerging specialized benchmarks reveal gaps in LLM agent reliability for high-stakes tasks:

• FinToolBench evaluates real-world financial tool use
  -...

• Core issue: Deeper LLM layers suffer signal degradation, diluting informative features.
• MoDA solution: Attention heads access KV pairs from...

A new paper introduces Latent Entropy-Aware Decoding for mitigating hallucinations in MLRMs by thinking in uncertainty, advancing safer multimodal reasoning.

SocialOmni benchmark evaluates audio-visual social interactivity in omni models, pushing multimodal world models on social dynamics. Join the discussion.

Enterprise AI agents demand new standards for secure scaling.

• NIST's wake-up call: Initiative sets governance and security baselines for CISOs to...

New paper outlines a cognitive framework for tracking AGI advancement, sparking interest with 58 HN points—pushing beyond narrow tasks toward holistic benchmarks.

HorizonMath launches with 100+ unsolved math problems, enabling genuine AI progress tracking.

• Automated verification: Open-source framework checks...

Querywise prompt routing treats prompt choice as a per-query decision problem for LLMs, using a learned offline proxy reward to score query-prompt pairs. Key for serving optimization.

Transformer Architecture Advances

• 🔥 Attention Residuals: Moonshot AI proposes replacing transformer residual connections with a depth-wise...

Key trend in multimodal foundations:

• MoE architectures resolve vision-language compute asymmetry, enabling efficient unified pretraining from...

Key patterns in agent advancements:

• Competitive benchmarks like PokeAgent push long-horizon decision-making in multi-agent battles.
• High-stakes...

RAG solves the core issue where frozen LLMs like GPT-4 hallucinate on recent data—new SKUs, policy changes, or fresh tickets—by enabling real-time reasoning over operational data.

Attention Residuals paper page live: Join the discussion.

• MR-Search uses meta-RL and self-reflection to improve LLM agents on complex search tasks, learning from past episodes.
• Introduces multi-turn RL...