AI Research Pulse

New Model Releases

• 🔥 MiniMax M3: MiniMax M3 is open-sourced on Hugging Face with proprietary blockwise sparse attention enabling efficient...

Two papers address transformer inference bottlenecks from complementary angles:

• Still trains a lightweight Perceiver for one-pass KV cache...

A four-layer generative-verifier pipeline—bad-case filtering, solution normalization, multi-judge parallel scoring, and pessimistic min...

MLLMs often fail on real-world degraded images from noise, compression, or weather. Robust-U1 adds explicit recovery via a three-stage pipeline on the...

EvoArena introduces a benchmark for evaluating how LLM agents track and evolve memory across changing environments, offering practical testing insights for teams building adaptive agents.

MiniMax's MaxProof advances automated theorem proving by pairing generative-verifier RL with population-level test-time scaling on the M3 series.

-...

CodeSpear reveals that grammar-constrained decoding—widely used for reliable JSON or code outputs—masks refusal tokens, preventing models from...

LARA aligns latent action representations in vision-language-action models, improving generalization and efficiency for embodied AI tasks.

Two fresh sparse attention designs target practical LLM deployment:

• MiniMax Sparse Attention delivers a minimal blockwise mechanism supporting full...

EvoArena provides a benchmark that tracks memory evolution to evaluate LLM agents in dynamic environments, moving beyond static setups where most current tests fall short.

This paper introduces the Internet of Agentic AI (IoAI) vision: an open ecosystem where heterogeneous agents discover one another and negotiate...

Frontier LLMs from Google, OpenAI, and Anthropic outperformed specialized tools like OpenEvidence and UpToDate in medical information retrieval, as...

MaxProof introduces a generative-verifier RL approach via MiniMax-M3, which jointly trains generation, error-finding verification, and refinement to...

Neural Thickets shows random weight perturbations (RandOpt) can match gradient-based methods like GRPO around pretrained LLM weights. This reveals dense task experts in the geometry of neural nets, offering lightweight adaptation without complex RL.

Three developments highlight practical paths to more reliable LLMs:

• Root-cause analysis distinguishes misinformation from misalignment using game...

Is the AI safety field asking the right questions?

• Defines "AI safety research" narrowly as alignment, x-risk, and long-term safety agenda
• Title...

New research directly tests Random Forest classifier weaknesses in ML-based network security systems facing adversarial attacks. The findings flag practical risks for engineers deploying traditional models in hostile environments.

A Y Combinator video spotlights five key papers charting AI research directions, including a world model of protein biology via Evolutionary Scale Models.

TRL-Bench evaluates frozen tabular encoder representations at row, column, and table levels using lightweight probes rather than end-to-end...

Safety-aligned systems remain exposed to novel exploit classes that traditional testing misses.

• Prompt injection defeats current web agents:...