# OraclePhys: A Systematic Framework for LLM Fine-Tuning on Structural Mechanics

arXiv: https://arxiv.org/abs/2608.17162 · HTML: https://mlsanity.com/paper/2608.17162/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 24 of its 34 references that are in the map.
- First posted on arXiv 17 Aug 2026.

## Builds on (most cited first, 24 of 24)

- [Let's Verify Step by Step](https://mlsanity.com/paper/2305.20050/) (2023)
- [Rethinking the Role of Demonstrations: What Makes In-Context Learning Work?](https://mlsanity.com/paper/2202.12837/) (2022)
- [Understanding R1-Zero-Like Training: A Critical Perspective](https://mlsanity.com/paper/2503.20783/) (2025)
- [LIMA: Less Is More for Alignment](https://mlsanity.com/paper/2305.11206/) (2023)
- [Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model?](https://mlsanity.com/paper/2504.13837/) (2025)
- [Solving math word problems with process- and outcome-based feedback](https://mlsanity.com/paper/2211.14275/) (2022)
- [RAFT: Reward rAnked FineTuning for Generative Foundation Model Alignment](https://mlsanity.com/paper/2304.06767/) (2023)
- [SFT Memorizes, RL Generalizes: A Comparative Study of Foundation Model Post-training](https://mlsanity.com/paper/2501.17161/) (2025)
- [Designing and Interpreting Probes with Control Tasks](https://mlsanity.com/paper/1909.03368/) (2019)
- [Reinforced Self-Training (ReST) for Language Modeling](https://mlsanity.com/paper/2308.08998/) (2023)
- [Emergent World Representations: Exploring a Sequence Model Trained on a Synthetic Task](https://mlsanity.com/paper/2210.13382/) (2022)
- [Beyond Human Data: Scaling Self-Training for Problem-Solving with Language Models](https://mlsanity.com/paper/2312.06585/) (2023)
- [Do Prompt-Based Models Really Understand the Meaning of their Prompts?](https://mlsanity.com/paper/2109.01247/) (2021)
- [Emergent Linear Representations in World Models of Self-Supervised Sequence Models](https://mlsanity.com/paper/2309.00941/) (2023)
- [Physics of Language Models: Part 3.1, Knowledge Storage and Extraction](https://mlsanity.com/paper/2309.14316/) (2023)
- [Many-Shot In-Context Learning](https://mlsanity.com/paper/2404.11018/) (2024)
- [Fine-Tuning Enhances Existing Mechanisms: A Case Study on Entity Tracking](https://mlsanity.com/paper/2402.14811/) (2024)
- [A Minimalist Approach to LLM Reasoning: from Rejection Sampling to Reinforce](https://mlsanity.com/paper/2504.11343/) (2025)
- [Evaluating the World Model Implicit in a Generative Model](https://mlsanity.com/paper/2406.03689/) (2024)
- [Can Language Models Serve as Text-Based World Simulators?](https://mlsanity.com/paper/2406.06485/) (2024)
- [Do Models Really Learn to Follow Instructions? An Empirical Study of Instruction Tuning](https://mlsanity.com/paper/2305.11383/) (2023)
- [What Has a Foundation Model Found? Using Inductive Bias to Probe for World Models](https://mlsanity.com/paper/2507.06952/) (2025)
- [Tailored Primitive Initialization is the Secret Key to Reinforcement Learning](https://mlsanity.com/paper/2511.12429/) (2025)
- [BeamPERL: Parameter-Efficient RL with Verifiable Rewards Specializes Compact LLMs for Structured Beam Mechanics Reasoning](https://mlsanity.com/paper/2603.04124/) (2026)

## Abstract (arXiv)

What a language model internalizes from fine-tuning is usually diagnosed after the fact. We make it an experimental variable. OraclePhys is a systematic fine-tuning framework with three components: OraclePhys-Bench, an exactly-graded structural-mechanics benchmark whose finite-element oracle scores every answer and counterfactual edit -- no human labels, no LLM judging; OraclePhys-30K, a supervision dataset of seven answer forms over byte-identical structure descriptions; and a controlled training study across the seven forms and three verifier roles. The study yields two findings. First, the label's answer form -- not its bit count -- causally determines what fine-tuning teaches: a ranking objective installs an out-of-distribution forward model where the untrained base sits at the guessing prior, a scalar objective at best a partial one, a boolean nothing detectable; the vector-scalar gulf survives a second physics domain, a second model family, and a paraphrased evaluation surface. Second, written or score-filtered answers install this capability, while advantage-weighted scores (GRPO) raise reward yet leave the model statistically equivalent to its start on held-out physics -- within the recipes and budgets tested -- sufficing only for routing. The trained 8B -- the first LLM on spatial structural response -- reaches the task's data-precision frontier: above a frontier LLM at zero- and 32-shot, at a specialist's level. What the label spells out about the target computation is what fine-tuning teaches; what you train on is what you route.