FUSION: a skill-based research agent for publicly obtainable nuclear-physics codes

summary

The gist

A skill-based research agent, FUSION, addresses the difficulty of running and interpreting unfamiliar nuclear-physics codes by embedding code-specific skills that ensure correct input conventions and

In short

FUSION is a skill-based research agent designed to run and interpret complex nuclear-physics codes. It embeds code-specific skills that ensure correct input conventions, validate outputs against benchmarks, and record failure modes. This system prevents general agents from producing physically incorrect results when dealing with specialized computational physics software.

Key concepts

Skill Mechanism: A skill is a self-contained set of instructions for a specific code. It includes steps for installation, using verified inputs to avoid errors, execution wrappers, output parsing rules, and documented failure modes. Crucially, each skill includes a benchmark to reproduce known results within specified tolerances.
Verification Process: Skills undergo rigorous testing before release using mechanical checks and adversarial review loops. This involves building the code on different operating systems (like macOS ARM and Linux x86-64) and ensuring that guard conditions are set up so they fail exactly as expected, confirming the skill's reliability.
Benchmark Certification: A benchmark confirms that a compiled code build reproduces a known reference result. The level of certification varies; Tier 1 requires byte-for-byte matches or agreement to six significant figures for some codes, while Tier 2 involves reproducing invariants or analytic solutions.
Separation Layers: FUSION uses three distinct layers: an engine (the core agent), a nuclear-physics layer (holding the code skills and literature), and a private user layer. This architecture ensures that public software remains isolated from private user data, preventing exposure during cloning or operation.

Terminology used across episodes

Reinforcement Learning (RL) — explained in 25 episodes
Retrieval-Augmented Generation (RAG) — explained in 16 episodes
Federated Learning — explained in 15 episodes
Interpretability — explained in 15 episodes
Sample Complexity — explained in 15 episodes
Generalization — explained in 13 episodes

This episode discusses

The paper

FUSION: a skill-based research agent for publicly obtainable nuclear-physics codes · Read on arXiv

School of Physics Science and Engineering, Tongji University · Southern Center for Nuclear-Science Theory (SCNT), Institute of Modern Physics, Chinese Academy of Sciences

Running an unfamiliar nuclear-physics code is rarely difficult because of the physics alone. One must find and build the program, learn its input conventions, and decide whether a plausible output is actually correct. A general-purpose coding agent helps with the first two tasks but may make the last one harder: it can write an input file that runs with the wrong physical convention. FUSION addresses this problem with code-specific skills. A skill obtains the code from its public source, starts from a verified input, runs and parses the calculation, records known failure modes, and must reproduce a stated benchmark to a stated tolerance before reporting a result. The current release covers twenty codes, spanning optical models and reactions, nuclear structure, fission and statistical models, astrophysics and R-matrix analysis, and heavy-ion transport. It also includes an offline, searchable collection of 61 167 pages derived from the nucl-th literature. User notes and credentials remain outside the public repository. FUSION is available under the MIT license at https://github.com/jinleiphys/FUSION; documentation is at https://vibeinscience.com. Here I describe the design, the checks behind the current release, and one complete calculation from input to comparison with measured data.

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "FUSION: a skill-based research agent for publicly obtainable nuclear-physics codes".

Jane: A skill-based research agent, FUSION, addresses the difficulty of running and interpreting unfamiliar nuclear-physics codes by embedding code-specific skills that ensure correct input conventions and validate outputs against benchmarks.

Tom: First, who's behind it and why it matters.

Title and authors: Tom: So, we’re talking about this paper called "FUSION: a skill-based research agent for publicly obtainable nuclear-physics codes." It sounds like they’ve built something specifically to handle the headache of running these complex physics programs.

Jane: It seems like the core idea is that just having the right code isn't enough; you still have to know how to set it up correctly and interpret what it spits out, which this agent aims to solve.

Lu: The focus on "skill-based" suggests they aren't just trying to be a general coding assistant; they are building specialized tools that understand the specific conventions of nuclear physics software.

Meng: I wonder how deep this specialization goes; is it just about following instructions, or does it actually understand the underlying physics constraints?

Lalam: This paper introduces a layered architecture, which suggests a very thoughtful approach to keeping sensitive user information separate from the public code execution environment.

The paper's summary: Tom: What’s really interesting about this is how FUSION tackles the problem of getting accurate results when using unfamiliar nuclear physics codes by embedding code-specific skills that handle everything from setup to validation.

Jane: They emphasize that a general assistant can write syntactically correct input, but it often fails because it uses the wrong physical conventions, leading to results that look plausible but are fundamentally wrong.

Lu: The paper describes how a single skill handles fetching the code from its public source, starting with verified inputs, running the calculation in a clean space, parsing the output according to specific rules, and crucially, checking that result against a benchmark before reporting anything.

Meng: That systematic approach of starting from known good inputs instead of writing everything from scratch sounds like a massive practical help for researchers who are just trying to get their experiments moving quickly.

Lalam: They detail the three layers: an engine, the nuclear-physics layer with skills and knowledge bases, and a private layer for user materials, which really shows they’re thinking about security from the start.

The paper's improvements: Tom: One of the key improvements FUSION proposes is this rigorous verification process built into every skill. They mandate that each skill must reproduce a stated benchmark to a specific tolerance before it can report a result, which is much stronger than just running the code.

Jane: They also focus heavily on documentation for known failure modes, which means the agent doesn't just fail silently; it’s supposed to record practical issues like exit codes or locale problems so you know what might go wrong.

Lu: The verification loop they suggest is quite thorough, involving building the code on different operating systems and having a second language model check for defects written into the skill itself.

Meng: From an engineering standpoint, this kind of self-checking mechanism addresses the exact issue of silent errors that plague complex software execution environments.

Lalam: This emphasis on reproducing results with stated tolerances provides a very high bar for soundness, moving beyond simple syntactic correctness to actual physical accuracy verification.

Conclusion: Tom: So, to wrap things up on "FUSION: a skill-based research agent for publicly obtainable nuclear-physics codes," the big idea is creating these highly specialized tools that ensure accuracy through rigorous skill execution and benchmark checks.

Jane: It really boils down to moving away from general agents that might make convention errors and toward systems that are explicitly designed to handle the unique idiosyncrasies of specific physics software.

Lu: The architecture, separating the engine from the nuclear-physics layer and having dedicated skills for twenty different codes, shows a very deliberate plan for comprehensive coverage within a structured framework.

Meng: I think what this paper offers is a tool that actually gets researchers past the initial setup hurdle and into meaningful calculation much faster because it handles those tedious checks automatically.

Lalam: This work sets a precedent for how specialized AI agents can be built to handle highly technical, domain-specific tasks while maintaining strict separation for user data, which is really important for future research tools.

More episodes

← Home