PaperlayerView exact revision on arXiv

Source-linked AI summary

Mind's Eye: Grounded Language Model Reasoning through Simulation

Ruibo Liu, Jason Wei, Shixiang Shane Gu, Te-Yen Wu, Soroush Vosoughi, Claire Cui, Denny Zhou, Andrew M. Dai

arXiv:2210.05359v1cs.CLcs.AI

TL;DR

Text-trained language models lack firsthand experience of the physical situations described in language, and can misrepresent physical knowledge when reasoning about them. Mind’s Eye simulates a physical question and feeds the outcome into the LM prompt. The authors report strong zero-shot and few-shot benchmark performance, including results from small augmented models comparable to vanilla models 100× larger.

  • Problem

    Text-trained LMs lack firsthand experience of described situations, and the paper investigates how well they understand and reason about basic physical rules.

    Source: §1 · pp. 1–2

  • Method

    Mind’s Eye translates a physical reasoning question into rendering code, simulates it with a physics engine, and adds the result to the LM prompt.

    Source: §1 · p. 2

  • Results

    Mind’s Eye outperforms other methods by a large margin in zero-shot and few-shot settings, while small augmented LMs can match or outperform vanilla LMs 100× larger.

    Source: §1 · p. 3

  • Takeaways & Limitations

    The authors conclude that simulation-grounded reasoning can boost small-scale LMs without handcrafted prompts or costly fine-tuning.

    Source: §6 · p. 9

    Performance can suffer if the simulator is inaccurate, and the experiments are in English, so generalization across languages is not claimed.

    Source: ETHICS AND REPRODUCIBILITY STATEMENT · p. 10

Abstract

from arXiv · show

Successful and effective communication between humans and AI relies on a shared experience of the world. By training solely on written text, current language models (LMs) miss the grounded experience of humans in the real-world -- their failure to relate language to the physical world causes knowledge to be misrepresented and obvious mistakes in their reasoning. We present Mind's Eye, a paradigm to ground language model reasoning in the physical world. Given a physical reasoning question, we use a computational physics engine (DeepMind's MuJoCo) to simulate the possible outcomes, and then use the simulation results as part of the input, which enables language models to perform reasoning. Experiments on 39 tasks in a physics alignment benchmark demonstrate that Mind's Eye can improve reasoning ability by a large margin (27.9% zero-shot, and 46.0% few-shot absolute accuracy improvement on average). Smaller language models armed with Mind's Eye can obtain similar performance to models that are 100x larger. Finally, we confirm the robustness of Mind's Eye through ablation studies.

1 INTRODUCTION

The introduction argues that text-trained language models lack firsthand physical experience, which can lead them to misrepresent physical knowledge and answer simple questions incorrectly. Mind’s Eye adds physics simulation to the reasoning pipeline, and the authors introduce UTOPIA to evaluate physical reasoning across 39 subtasks.

1 INTRODUCTION

  • Text-trained LMs lack firsthand experience of described situations and can answer physical questions incorrectly, as GPT-3 does for free fall.The introduction also links this lack of grounded reasoning to truthfulness and factuality concerns.Source: §1 · p. 1
  • The authors frame grounded physical understanding as important for human-level reasoning and general-purpose embodied intelligence.They investigate how well current LMs understand basic physical rules and how simulation can ground their reasoning.Source: §1 · p. 2
  • UTOPIA covers 39 physics subtasks, and GPT-3 175B reaches 24% zero-shot accuracy and 38.2% few-shot accuracy on the benchmark.The benchmark is designed to assess understanding and reasoning over basic laws of physics.Source: §1 · p. 2
  • Mind’s Eye transforms a text question into rendering code, runs a physics simulation, and appends the results to the LM prompt during inference.The authors describe the framework as plug-and-play, requiring neither handcrafted prompts nor costly fine-tuning.Source: §1 · pp. 2–3
  • Mind’s Eye outperforms other methods by a large margin in zero-shot and few-shot settings, and small augmented LMs can match or outperform vanilla LMs 100× larger.The authors evaluate popular LMs of different sizes on UTOPIA before and after augmentation.Source: §1 · p. 3

2 UTOPIA BENCHMARKING

UTOPIA benchmarks language models’ understanding of basic physical concepts through questions spanning six representative scenes. Its 39 subtasks use relative property comparisons and physics-engine-generated answers, with data designed to be scalable beyond human-annotated datasets.

2 UTOPIA BENCHMARKING

  • UTOPIA contains 39 subtasks across six representative scenes, with questions based on compositions of observed and queried physical concepts.The questions use textbook-like language and include various hand-designed examples within each subtask.Source: §2 · p. 3, Table 1 · p. 3
  • Physics-engine-generated ground-truth answers support reliable labels and make UTOPIA scalable to larger sizes.The present LM benchmark uses text-form data; simulation videos were recorded for future multimodal research.Source: Table 1 · p. 3, §2 · p. 4
  • UTOPIA describes properties through relative comparisons rather than absolute numbers, separating reasoning evaluation from gains attributable to numeracy shortcuts.The authors contrast this design with mathematical tasks where models may solve by symbolic manipulation rather than physical reasoning.Source: §2 · pp. 3–4
  • Unlike vision-centered physics datasets and human-annotated benchmarks, UTOPIA uses engine-generated data and can in principle provide unlimited samples.The cited MMLU physics subset contains 206 samples, combining college- and high-school-level questions.Source: §2 · p. 4

3 MIND’S EYE

Mind’s Eye combines a text-to-code model, a physics engine, and a foundation model for general reasoning. It translates questions into simulations, turns outcomes into prompt injections, and uses those outcomes to ground LM reasoning in physical rules.

3 MIND’S EYE

  • Mind’s Eye uses a text-to-code frontend, MuJoCo as the simulation backend, and a foundation model for general reasoning.The inference formulation compares how different methods ground a response to a question.Source: §3 · p. 4, Table 2 · p. 5
  • A text-to-code model generates MCJF rendering code from questions, using task templates and training on 200,000 text-code pairs.The authors train decoder-only models with 0.3B and 1.5B parameters, and also pre-train on C4 to improve natural-language understanding.Source: §3 · p. 4
  • MuJoCo runs the rendering code to produce a ground-truth outcome, which the system parses into text-form hints and appends to the question for the foundation model.The injected outcome lets the LM ground its reasoning in the simulated physical world.Source: §3 · p. 4
  • Mind’s Eye imitates experiment-based reasoning with quick, inexpensive simulation as an alternative to physical-world experiments.The authors describe mental visualization as a possible cognitive analog for the approach.Source: §3 · p. 4
  • Simulation supplies physical-world outcomes, while zero-shot prompts and handcrafted reasoning steps depend on LM knowledge or human annotations that may be incorrect, outdated, or unscalable.The pipeline is automated and uses the simulator’s encoded world knowledge to ground the prompt.Source: §3 · pp. 4–5
  • Delegating physics knowledge to MuJoCo lets smaller augmented LMs achieve reasoning performance similar to vanilla models 100× larger.The prompting-based approach also avoids the instability issues of training mixture-of-experts models.Source: §3 · p. 5

4 EXPERIMENTS

Experiments evaluate Mind’s Eye on UTOPIA using language models of different sizes, then test whether its gains depend on correct simulations, additional demonstrations, or the text-to-code model. Mind’s Eye improves reasoning across settings, while the ablations identify simulation correctness as important.

4.1 EXPERIMENTS SETTINGS

  • Evaluation uses about 3,900 UTOPIA samples and GPT-3 and PaLM models ranging from 340M to 540B; MuJoCo simulations run at 171 fps on one A6000 GPU.Generating 100 simulations of a two-second collision scene takes 0.67 seconds.Source: §4.1 · p. 5

4.2 BENCHMARK RESULTS ON UTOPIA

  • 34.2% in few-shot and 18.4% in zero-shot are Mind’s Eye’s average absolute accuracy gains over vanilla LMs across model sizes.Figure 2 also reports that smaller augmented LMs can match or outperform larger vanilla LMs.Source: Figure 2 · p. 6, §4.2 · p. 6
  • Mind’s Eye outperforms other reasoning-enhancement methods on UTOPIA in zero-shot and few-shot settings, including with GPT-3 Babbage 1.3B.Table 3 compares Mind’s Eye with prompting, decoding, retrieval, and fine-tuning approaches.Source: §4.2 · p. 6, Table 3 · p. 7
  • 46.0% vs. 10.3% average grounding gain for GPT-3 175B vs. 1.3B shows larger LMs leverage Mind’s Eye more, especially in few-shot settings.When scaling from 1.3B to 175B, few-shot accuracy rises 1.8% with vanilla LMs (36.4% to 38.2%) and 37.5% with Mind’s Eye (46.7% to 84.2%).Source: §4.2 · p. 7
  • 68.6% to 99.1%: Instruct-GPT augmented with Mind’s Eye achieves nearly perfect physics-alignment performance in few-shot settings.The authors attribute this result to Instruct-GPT’s instruction-following optimization and proper reasoning rationale.Source: §4.2 · p. 7

4.3 ABLATION STUDY

  • As shot count increases, vanilla few-shot and Semi-Mind’s Eye do not improve as much as Mind’s Eye; CoT and semi-augmented results vary with demonstration similarity.The authors connect Mind’s Eye’s effectiveness to LMs grounding reasoning on recent context, supplied by the question’s simulation results.Source: Figure 3 · p. 8, §4.3 · pp. 8–9
  • 4.7% vs. 3.6% accuracy decreases with a smaller text-to-code LM occur for 1.3B vs. 175B foundations in zero-shot; for 175B, the decrease is 3.6% zero-shot and 2.1% few-shot.The authors say the smaller text-to-code model’s conversion errors are mitigated by few-shot demonstrations.Source: Table 5 · p. 8, §4.3 · p. 9
  • Mismatched simulation outcomes substantially reduce performance, and incorrect outcomes worsen it further; removing “Hints:” has marginal influence.The ablation tests altered or incorrect simulation results and removal of the prompt trigger words.Source: §4.3 · pp. 7–8, Table 4 · p. 8

5 RELATED WORK

Mind’s Eye is positioned against earlier approaches that rely on hand-authored rules or prompt additions, grounding LM reasoning instead through automated simulation. The related work connects this approach to augmented inputs and world-model research while distinguishing its focus on natural-language reasoning.

5 RELATED WORK

  • Mind’s Eye grounds LM reasoning through automated simulation rather than manually created rules or human-crafted rationales.Earlier work used manually created linguistic and physical rules, or added reasoning steps, task descriptions, and target formats to prompts.Source: §5 · p. 9
  • Mind’s Eye adds simulation results as extra sensory input, in contrast to interactive API augmentation or using an LM to help robots understand instructions.The paper describes its prompt-based approach as less costly than TALM’s interactive fine-tuning.Source: §5 · p. 9
  • Mind’s Eye plans experiments with a text-to-code LM and reasons over their simulated results, focusing on natural-language reasoning rather than robotics or computer vision.The related work situates simulated environments in robotics and computer vision research as well.Source: §5 · p. 9

6 CONCLUSION AND FUTURE WORK

The paper concludes that Mind’s Eye grounds LM reasoning with simulations and boosts small-scale LM performance without handcrafted prompts or costly fine-tuning. It suggests extending simulation-based grounding to other domains, especially where simulation engines already exist.

6 CONCLUSION AND FUTURE WORK

  • Mind’s Eye grounds LM reasoning with simulations and boosts small-scale LM performance without handcrafted prompts or costly fine-tuning.The paper characterizes the paradigm as effective, scalable, and efficient.Source: §6 · p. 9
  • The authors suggest extending simulation-based reasoning to domains such as economic changes or thermodynamics, especially where simulation engines already exist.They mention possible applications in policy making or engineering.Source: §6 · p. 9
  • The authors state that Mind’s Eye’s dynamic generation of grounding evidence unlocks the models’ scaling potential.Source: §6 · p. 9

ETHICS AND REPRODUCIBILITY STATEMENT

The paper identifies simulation accuracy and English-only experiments as scope boundaries. For reproducibility, it uses mainly publicly available LMs and open-source baselines and includes UTOPIA and model outputs in supplementary materials.

ETHICS AND REPRODUCIBILITY STATEMENT

A.1 DETAILS ABOUT MIND’S EYE

The appendix describes a pipeline that converts a physics question into rendering code, runs a simulation, and parses selected results into prompt-ready rationale. The generated code identifies the scene and properties of interest, which guide the manager’s extraction and conclusion templates.

A.1 DETAILS ABOUT MIND’S EYE

  • A text question is converted into rendering code for simulation, and a manager parses the results to create simulation-based prompt injection.Figure A.1 presents this procedure.Source: §A.1 · p. 17, Figure A1 · p. 17
  • The rendering code marks the scene name and physics attributes of interest with #%#% metadata, which the text-to-code LM is trained to generate.The LM is trained on text-code pairs prepared for this purpose.Source: §A.1 · p. 17
  • The simulation manager records sensory data and extracts only the requested physical property, such as acceleration.It uses MuJoCo APIs to record quantities such as velocity, acceleration, and energy.Source: §A.1 · p. 17
  • The manager fills conclusion templates with extracted observations, adds prompt cues, and appends the result as grounded rationale for the LM.The example concludes that two objects have the same acceleration when acceleration is the requested property.Source: §A.1 · p. 17

A.2 SCENE GENERALIZATION WITH MIND’S EYE

Mind’s Eye is presented as a way to generalize across scenes without fine-tuning on each new context. In the comparison, fine-tuned PaLM 62B performs well in-domain but poorly on an unseen scene, while Mind’s Eye maintains consistent zero-shot alignment across six scenes.

A.2 SCENE GENERALIZATION WITH MIND’S EYE

  • Simulation-based prompts let Mind’s Eye avoid fine-tuning the language model each time it encounters a new context.Source: §A.2 · p. 17
  • Fine-tuning performed well on in-domain data but was limited on scenes the model had not seen during fine-tuning.The Motion-to-Projection comparison illustrates this out-of-domain drop.Source: §A.2 · p. 18, Figure A2 · p. 18
  • 70% in-domain alignment accuracy after Motion-scene fine-tuning fell to 24% on Projection, while Mind’s Eye enabled consistent zero-shot alignment across six scenes.The comparison uses fine-tuned PaLM 62B and vanilla PaLM 62B armed with Mind’s Eye.Source: Figure A2 · p. 18, §A.2 · p. 18

A.3 MIND’S EYE ERROR ANALYSIS

The error analysis names three failure types and reports that they are less common with larger language models. The authors connect their greater prevalence in small models to limited reasoning over the supplied evidence.

A.3 MIND’S EYE ERROR ANALYSIS

  • The analysis identifies three errors: ignoring the grounded rationale, extracting the answer from nearby words, and generating nonsensical output.The labels are ignorance error, recency bias, and “I don´t know” error, respectively.Source: §A.3 · p. 18
  • Recency bias can lead the model to extract its final answer from nearby words, such as “the object Y.”Source: §A.3 · p. 18
  • These errors are much less common with larger language models; the authors attribute their prevalence in small models to limited reasoning over the evidence.Source: §A.3 · p. 18, Figure A3 · p. 18

Related papers

Loading 2210.05359v1…