ML Research Pulse

27B model outperforms a 397B model and MiniMax-M2.5 on the SWE-Bench coding benchmark, sparking debate: real efficiency breakthrough or benchmaxxed?

Different language models independently learn similar internal representations for numbers, revealing convergent evolution in their architectures.

ReImagine rethinks controllable high-quality human video generation via image-first synthesis.

SAVOIR proposes Shapley-based reward attribution to learn social savoir-faire in agents.

Trend alert: Google and Sakana AI Labs push specialized agents automating scientific workflows—from reports to evaluations.

• Google's Deep...

Sakana AI solves the deceptively deep challenge of LLMs performing fair internal coin tosses using prompts alone. Their paper "SSoT: Prompting LLMs for Distribution-Faithful and Diverse Generation" was accepted to #ICLR2026.

• Claude Code is 4x bigger than Codex
• Codex is 2x bigger than Cursor
• Antigravity nearly matches Cursor, likely due to Googler usage

One HF Hub data point – tagging for more agents soon.

Emerging trend in self-evolving LLM techniques:

• Memory extraction across heterogeneous tasks
• Efficient terminal agents via observational context compression
  Watch for advances in agent memory management.

A 1.7B parameter model beats GLM-5 (744B) on Schema Guided Dialogue—even when training data is corrupted—a 437x size difference showcasing data-efficient small models' dominance.

New paper unpacks reward hacking mechanisms, emergent misalignment, and challenges in the era of large models. Join the discussion on this key LLM alignment topic.

LLaDA2.0-Uni introduces a Diffusion Large Language Model that unifies multimodal understanding and generation. Breakthrough for seamless multimodal LLMs from core AI research.

GPT-5.5 and internal model names spotted on Codex—a classic sign OpenAI is gearing up for a new release. Eyes on the horizon for their next major leap.

New Models from Labs

• 🔥 OpenAI Privacy Filter: OpenAI released Privacy Filter, an open-weight 1.5 billion parameter model using...

CoInteract introduces physically-consistent human-object interaction video synthesis via spatially-structured co-generation, pushing realistic video models in human-object dynamics.

PlayCoder bridges LLM code generation to interactive, playable GUIs, turning generated code into functional interfaces. Join the paper discussion for deeper insights.

• TEMPO scales test-time training for large reasoning models, optimizing at inference.
• Neural memory breakthrough: LMs learn end-to-end to forget...

OpenAI ditches Sora for enterprise-focused image gen, prioritizing text-heavy designs like infographics, magazines, and posters.

• Real-world...

• DR-Venus advance: Targets frontier edge-scale deep research agents using only 10K open data, enabling efficient low-data training.
• Meta's...

• WavAlign enhances intelligence and expressiveness in spoken dialogue models
• Leverages adaptive hybrid post-training for key gains in speech AI
• New paper page invites discussion on this breakthrough

New Nature Machine Intelligence paper uncovers two competing biases explaining LLMs' over- and under-confidence:

• Choice-supportive bias: LLMs...