AI4AI Hits 71.21% as AskChem Indexes 2.4M Claims

Featured

01 Architecture: World Models Need Physical Language

Predicting pixels answers what the next frame looks like. Expressing physical change requires an account of how the world changes. PhiZero separates those jobs.

The model learns a compact, discrete “physical language” from real-world video through self-supervision. It first reasons about state changes, then renders those changes as video. Physical rules no longer need to stay buried inside a high-dimensional pixel predictor.

The paper reports physically consistent generation and understanding. It also supports interactive world modeling, fine-grained action-conditioned simulation, and zero-shot motion transfer. Full results must show whether this language is interpretable, reusable across scenes, and better than direct pixel prediction.

Key takeaways:

• Evaluate whether a world model can explicitly represent and track state changes, not only produce attractive frames.
• Reasoning before rendering could make simulations easier to debug, control, and reuse.
• Interpretability and cross-scene generalization will determine whether physical language works beyond demonstrations.

Source: PhiZero: A World Model Built Around Physical Language

02 AI for Science: Search Returns Verifiable Claims

A useful answer to a chemistry question needs a conclusion, a source, and traceable evidence. AskChem changes the retrieval unit from papers to atomic, typed claims.

Each claim links to a DOI and either a direct quotation or a precise evidence location. A taxonomy and evidence graph support further synthesis. The system indexes 2.4 million claims from 147,000 papers and offers web, REST, SDK, and MCP interfaces.

Agents can work directly with evidence chains instead of selecting papers from search results. The abstract reports that AskChem raised resolvable DOIs in GPT-5.5 outputs from 88.3% without retrieval to 100%. Claim extraction and evidence links still need separate validation.

Key takeaways:

• Research search products should return evidence that traces directly to the source text.
• Structured claims may serve research agents better than entire papers as context units.
• Better DOI resolution improves citation usability, but does not prove claim or evidence accuracy.

Source: AskChem: Claim-Centered Infrastructure for Chemistry Literature Synthesis

03 Agent: Memory Moves Inside the Model

Agent memory usually lives in external components such as vector databases and retrievers. Metis instead treats memory as a native foundation-model capability.

The architecture adds a persistent, mutable memory state to the model backbone. “Memory attention” reads compressed historical information, allowing storage and retrieval to train end to end. During deployment, model weights stay frozen. Updating memory requires no gradients and only one forward pass.

The abstract demonstrates native memory but provides too few numbers to isolate the source of the gains. Architecture, training data, and extra compute could all contribute. Teams should compare native memory with external modules on task quality, cost, control, and auditability.

Key takeaways:

• Native memory can bring storage and retrieval into end-to-end optimization.
• Frozen weights and forward-only memory updates could reduce online maintenance complexity.
• Choose memory architecture through task-level comparisons, not the “native” label.

Source: Metis: Memory Foundation Model

04 Code Intelligence: Two Gains Under One Budget

A base model scores 39.39% on MLE-Bench Lite under strict constraints: one RTX 4090 and 12 hours per task. A 35B model with a search framework reaches 60.61%. Experience priors and asynchronous search lift it to 71.21%.

OpenMLE combines task environments, execution feedback, operator training, and long-horizon search in an open stack. Actual execution results guide program improvement and composition. The staged scores suggest model training, experience priors, and search all affect outcomes under fixed resources.

Held-out tests also show separate gains from model training and the search framework. These results come from the Lite benchmark and a specific budget. Ablations must isolate each component, and broader machine-learning engineering tasks may behave differently.

Key takeaways:

• Measure model-training and search-framework gains separately under fixed time and compute budgets.
• The jump from 60.61% to 71.21% makes experience priors and asynchronous search worth independent evaluation.
• Strong Lite results still require confirmation on more tasks and the full benchmark.

Source: Frontis-MA1: Training an AI4AI Model towards Recursive Self-Improvement in Machine Learning Engineering