PreliminaryDataset construction and evaluation in progress

NLL Physics Research 1

Physics Research Grade problems are research-level physics and mathematics derivations, each reconstructed from a recent arXiv preprint and scored only on whether a model reaches the exact terminal result — a direct read on reasoning ability rather than style.

Pass@1

Strict — whole problem solved

Rubric score

Weighted partial credit

Each problem is broken into weighted subquestions, and the rubric score is the fraction of that credit an answer earns. It gives a model credit for correct structure and intermediate results even when it misses the final answer.

Cost vs performance

Avg $ per task · pass@1

Each mark is one model, placed by the selected x-metric on a log scale against strict pass@1. Cheaper or leaner systems sit to the left, and stronger systems sit higher.

Nature Loops Labs

Preliminary research note

Work in progress · July 2026

NLL Physics Research 1

Research-grade environments for theoretical and mathematical physics.

NLL-bench/2607.001

natureloops.ai

Why this work exists

Physics research is a chain of decisions.

Physics research combines formulation, derivation, tool use, verification and recovery from failed routes. Many benchmarks compress that work into a short answer.

NLL Physics Research 1 examines the longer chain. Each problem is reconstructed from a recent research paper and rewritten so it can stand on its own. The paper, worked solution and proof-bearing files stay outside the agent’s environment.

The agent must build the argument from the problem statement, use mathematical or computational tools where needed, and reach a result that can be checked.

A useful benchmark records where the reasoning failed as well as whether the final result was correct.

Capabilities under test

The environment tests the work research requires.

The aim is to make difficult reasoning observable and trainable.

Formulate

Make the problem precise

Translate prose, definitions and conventions into a complete mathematical representation.

Derive

Carry a long argument

Connect intermediate claims without a source paper or worked route to imitate.

Use tools

Compute when reasoning requires it

Write code, manipulate symbolic expressions and run numerical checks inside a controlled workspace.

Verify

Try to break the answer

Check dimensions, limits, symmetries, special cases and independent physical constraints.

Recover

Recognise a failed route and change course

Preserve useful partial work, identify the point of failure and attempt a different derivation.

From paper to environment

Every task starts with a result that can be checked.

Select

Choose a research paper with a deep, exact result and enough structure to support a self-contained problem.

Reconstruct

Work backwards from that result, defining every convention needed for a unique answer without leaking the route.

Verify

Write the golden answer and complete solution, then check the mathematics symbolically, numerically or against formal constraints.

Grade

Build a strict check for the final result and a weighted rubric for meaningful intermediate progress.

Isolate

Give the agent only the task and its tools. Keep the source, solution and private grader outside the runtime.

Audit

Review the task, answer and grader independently before freezing the evaluation set.

From benchmark to training programme

Cloud compute lets us test whether these environments improve a model.

Baseline

Freeze the evaluation protocol

Run selected models under matched conditions and record trajectories, costs, failures and verifier behaviour.

Training

Build a difficulty curriculum

Train on progressively harder environments while preserving unseen task families for evaluation.

Measurement

Test transfer and reliability

Compare frozen checkpoints, probe reward hacking and measure whether gains survive new problems and physical regimes.

The longer-term direction is a ladder of environments that moves from exact local calculations toward longer mathematical-physics arguments, including capabilities relevant to gauge theory and Yang–Mills.

The six-month pilot does not claim to address the mass-gap problem. It tests whether controlled training improves the component skills that such research demands.