PaperlayerView exact revision on arXiv

Source-linked AI summary

Large language models surpass human experts in predicting neuroscience results

Xiaoliang Luo, Akilles Rechardt, Guangzhi Sun, Kevin K. Nejad, Felipe Yáñez, Bati Yilmaz, Kangjoo Lee, Alexandra O. Cohen, Valentina Borghesani, Anton Pashkov, Daniele Marinazzo, Jonathan Nicholas, Alessandro Salatiello, Ilia Sucholutsky, Pasquale Minervini, Sepehr Razavi, Roberta Rocca, Elkhan Yusifov, Tereza Okalova, Nianlong Gu, Martin Ferianc, Mikail Khona, Kaustubh R. Patil, Pui-Shee Lee, Rui Mata, Nicholas E. Myers, Jennifer K Bizley, Sebastian Musslick, Isil Poyraz Bilgin, Guiomar Niso, Justin M. Ales, Michael Gaebler, N Apurva Ratan Murty, Leyla Loued-Khenissi, Anna Behler, Chloe M. Hall, Jessica Dafflon, Sherry Dongqi Bao, Bradley C. Love

arXiv:2403.03230v4q-bio.NCcs.AI

TL;DR

Synthesizing exponentially growing scientific literature challenges human information-processing capacity, motivating tests of whether LLMs can predict experimental outcomes. BrainBench evaluates that ability, and LLMs outperform experts across neuroscience subfields while confidence tracks correctness.

  • Problem

    Synthesizing exponentially increasing scientific literature challenges human information-processing capacity, motivating the question of whether LLMs can predict experimental outcomes.

    Source: Introduction · p. 3, Discussion · p. 15

  • Method

    BrainBench asks test-takers to choose which of two neuroscience abstract versions contains the study’s actual results.

    Source: Discussion · p. 15

  • Results

    LLMs outperform human experts by a considerable margin across neuroscience subfields, and confident predictions are more likely to be correct.

    Source: Discussion · pp. 15, 17

  • Takeaways & Limitations

    BrainBench provides a way to evaluate LLMs’ prediction of neuroscience results, with confidence offering an additional signal about likely correctness.

    Source: Discussion · pp. 15, 17

    One risk is that scientists may not pursue studies when their predictions run counter to LLM predictions.

    Source: Discussion · p. 18

Abstract

from arXiv · show

Scientific discoveries often hinge on synthesizing decades of research, a task that potentially outstrips human information processing capacities. Large language models (LLMs) offer a solution. LLMs trained on the vast scientific literature could potentially integrate noisy yet interrelated findings to forecast novel results better than human experts. To evaluate this possibility, we created BrainBench, a forward-looking benchmark for predicting neuroscience results. We find that LLMs surpass experts in predicting experimental outcomes. BrainGPT, an LLM we tuned on the neuroscience literature, performed better yet. Like human experts, when LLMs were confident in their predictions, they were more likely to be correct, which presages a future where humans and LLMs team together to make discoveries. Our approach is not neuroscience-specific and is transferable to other knowledge-intensive endeavors.

Introduction

Synthesizing the expanding scientific literature challenges human information-processing capacity, motivating a test of whether LLMs can predict neuroscience results. BrainBench evaluates that ability, and LLMs surpassed human experts, with the margin increasing after additional neuroscience training.

Introduction

  • Neuroscience prediction is difficult because relevant research is extensive, individual studies can be noisy, and the field spans multiple levels and complex methods.Source: Introduction · p. 4
  • Existing scientific benchmarks mainly assess factual retrieval and reasoning, not the prediction of novel outcomes.Source: Introduction · p. 5
  • For forward-looking tasks, information mixing that may count as hallucination in backward-looking tasks can instead amount to generalization or prediction.Source: Introduction · p. 5
  • BrainBench tests outcome prediction by asking test-takers to choose the original result from a coherent altered abstract, using recent abstracts across five neuroscience domains.Source: Introduction · p. 5, Figure 2 · p. 7
  • LLMs surpassed human experts on BrainBench by a substantial margin, and the margin increased after additional neuroscience training.The neuroscience-trained model is called BrainGPT.Source: Introduction · p. 8

Results

LLMs outperformed human experts on BrainBench, and their judgments reflected information integrated across abstracts. Neuroscience fine-tuning further improved performance, while both LLM and human confidence tracked accuracy.

General-purpose LLMs best neuroscientists on BrainBench

  • 81.4% was the average LLM accuracy, versus 63.4% for human experts on BrainBench.Every LLM outperformed the experts.Source: General-purpose LLMs best… · p. 8, Figure 3 · p. 9
  • LLMs showed similar results when the test cases were generated by GPT-4 rather than neuroscience-expert coauthors.Source: General-purpose LLMs best… · p. 8
  • Smaller models performed comparably to larger ones, while base models outperformed chat- and instruction-optimized counterparts.The authors suggest conversational alignment may hinder scientific inference.Source: General-purpose LLMs best… · p. 8, Figure 3 · p. 9
  • LLMs outperformed human experts in every neuroscience subfield, although Behavioral/Cognitive cases were overrepresented in BrainBench and the source journal.The figure reports average performance across 15 LLMs and human experts.Source: Figure 3 · p. 9, General-purpose LLMs best… · p. 10
  • LLMs performed worse when given only altered results sentences, supporting the authors’ conclusion that their advantage draws on information across the abstract.The authors identify background and methods as part of that context.Source: General-purpose LLMs best… · pp. 10–11
  • The authors found no indication that LLMs had memorized BrainBench, and their checks indicated the benchmark items were novel to the models.The zlib-perplexity analysis detected memorization for the Gettysburg Address, used as a comparison.Source: General-purpose LLMs best… · pp. 11–12

LLMs and human experts are calibrated

  • Both human experts and all LLMs showed a positive relationship between confidence and accuracy, with higher confidence corresponding to greater likelihood of correctness.Individual-level analyses also found significant positive correlations for both groups.Source: LLMs and human experts are calibrated · p. 12, Figure 4 · p. 13

Augmenting LLMs with neuroscience knowledge to create BrainGPT

Discussion

The discussion argues that LLMs can predict neuroscience results by integrating information across abstracts, with performance exceeding human experts. It also considers benchmark construction, model adaptation, collaboration, and limits on where these findings may apply.

Discussion

  • LLMs outperformed human experts across neuroscience subfields, with superior performance attributed to integrating information across abstracts.The authors report no indication of BrainBench memorization and describe predictions on studies novel to the models.Source: Discussion · p. 15
  • The authors found no indication that LLMs memorized BrainBench; analyses instead suggested learned field patterns enabled predictions on studies novel to them.A relatively small model trained from scratch on neuroscience literature excluding preprints and BrainBench items also achieved superhuman performance.Source: Discussion · pp. 15–16
  • LoRA-based BrainGPT models reoriented general-purpose LLMs for neuroscience and improved BrainBench performance; retrieval-augmented generation was proposed for accessing current articles.The authors suggest that keeping models abreast of expanding literature matters.Source: Discussion · p. 16
  • 100 GPT-4-generated test cases were close in quality to human-created items, and 8 were word-for-word matches.The largely automated process addresses the labor and expertise involved in creating forward-looking benchmarks.Source: Discussion · p. 16
  • LLM confidence was calibrated, and items difficult for LLMs differed from those difficult for experts, supporting complementary human-machine judgments.The authors state that combining these judgments can outperform either alone.Source: Discussion · pp. 16–17
  • The analyses relied on access to LLM weights; natural-language prompting may yield less reliable judgments and degrade competency compared with probability scores or separate classifiers.The authors favor models whose weights and training sets are publicly available.Source: Discussion · p. 17
  • Scientists may avoid studies whose predictions conflict with an LLM, although such studies could reveal gaps or errors in the literature.The authors note that a high-confidence prediction may also make a study seem incremental.Source: Discussion · p. 18
  • The approach is not neuroscience-specific, but efficacy may depend on domain structure; the authors foresee experts contributing scientific explanations alongside predictions.They note that mathematics, which relies heavily on logical deduction, might benefit less than pattern-based fields.Source: Discussion · pp. 18–19

Data availability

Human participant data and intermediate simulation and analysis data are publicly available, alongside model weights and training data.

Data availability

  • Human participant data and intermediate data generated through simulations and analyses are publicly available on GitHub.Model weights and training data are available on Hugging Face.Source: Data availability · p. 19

Code availability

The computer code associated with the work is publicly available on GitHub.

Code availability

  • Code for model training, evaluation, data processing, and analyses is publicly available on GitHub.Source: Code availability · p. 19

Materials and Methods

The study reports ethics approval from University College London and confirms that participants gave informed consent.

Materials and Methods

Dataset Creation

BrainBench test cases were made by altering published neuroscience abstracts to change their empirical results while preserving their methods and background. The dataset combined cases created by human experts and GPT-4, with quality control and coverage across five neuroscience subfields.

Dataset Creation

  • Test cases came from Journal of Neuroscience abstracts and covered five subfields: Behavioral/Cognitive, Systems/Circuits, Neurobiology of Disease, Development/Plasticity/Repair, and Cellular/Molecular.Source: Dataset Creation · p. 21
  • BrainBench includes 200 human-crafted test cases and 100 generated by GPT-4, and all cases underwent extensive human-expert and GPT-4 quality control.Source: Dataset Creation · p. 21
  • Each altered abstract changed the study's results without changing its methods or background, using minimal edits that remained coherent with the abstract.Source: Dataset Creation · pp. 21–22
  • Edits were intended to require neuroscience insight to recover the original finding, rather than general reasoning alone.Source: Dataset Creation · p. 22
  • Case creators were cautioned against changing a result to one that could also have occurred but was omitted from the abstract.Source: Dataset Creation · pp. 22–23
  • Edits were designed not to be inferable from the rest of the abstract, and changes to findings had to remain consistent with its other statements.Source: Dataset Creation · p. 23

Evaluations

The evaluations tested whether human experts and language models could distinguish abstracts with actual results from altered versions, using model scoring and diagnostic analyses alongside a human testing procedure. They also describe human participant and trial exclusions and assess human–LLM agreement by comparing item-difficulty rankings.

Model Evaluation

  • LLMs compared each abstract version separately and selected the one with lower perplexity; accuracy measured correct identification of the original.The evaluation used perplexity as the preference indicator and assessed overall accuracy across BrainBench.Source: Model Evaluation · pp. 25–26
  • Model confidence was the absolute perplexity difference between versions, and calibration was assessed by comparing accuracy across 20 confidence-ranked bins.A linear regression related bin number to mean accuracy.Source: Model Evaluation · p. 26
  • Agreement between LLMs was evaluated by Spearman correlation of their test-case difficulty rankings, based on perplexity differences between incorrect and correct abstracts.A larger positive difference indicated an easier test case from the model’s perspective.Source: Model Evaluation · pp. 26–27
  • To test contextual integration, researchers evaluated individual sentences with result alternations, averaged accuracy across multiple alternations, and also swapped results between same-subfield abstracts while preserving result counts.They compared performance on full abstracts with individual sentences.Source: Model Evaluation · p. 27
  • To examine possible training-data memorization, researchers used zlib entropy and perplexity ratios, and tested whether publication date correlated with abstract difficulty.They reasoned that memorization would predict a negative date–difficulty correlation.Source: Model Evaluation · pp. 27–28

Human Evaluation

  • Of 202 recruited neuroscience experts, 171 remained after exclusions; each completed nine test trials—six human-created and three machine-created cases—and two catch trials.Participants selected one version of each trial abstract.Source: Human Evaluation · pp. 28–29
  • Participants completed 11 trials, indicated whether they had previously encountered each study, and were debriefed on correctness before reporting any cheating.Expertise ratings used a slider with a 1–100 scale.Source: Human Evaluation · p. 30
  • Participants failing both catch trials, never adjusting expertise or confidence sliders, recognizing abstract content, responding in under 5 seconds, or reporting cheating were excluded.Recognized-content and fast-response trials were omitted; participants reporting external-resource use or cheating were excluded from final analyses.Source: Human Evaluation · p. 30
  • Human–LLM agreement was assessed by correlating item difficulty, defined as human mean accuracy and the corresponding LLM measure, using Spearman correlation.The LLM item-difficulty procedure matched the one described earlier in the paper.Source: Human Evaluation · p. 30

Fine-tuning on neuroscience corpora

The authors collected neuroscience literature to fine-tune a base model with LoRA adapters, then evaluated the fine-tuned model on BrainBench using the same procedure as before.

Fine-tuning on neuroscience corpora

  • LoRA adapters added low-rank trainable parameters to fine-tune general-purpose models without training the entire model.The authors applied the adapters after self-attention blocks and fully connected layers; the trainable parameters were roughly 8% of the base model.Source: Fine-tuning on neuroscience corpora · p. 31
  • Using PubMed abstracts and PubMed Central Open Access Subset full-text articles published from 2002 to 2022, data collection yielded 332,807 abstracts and 123,085 articles, totaling 1.3 billion tokens.Figures and tables were excluded.Source: Fine-tuning on neuroscience corpora · p. 31
  • The fine-tuned model was tested on BrainBench using the prior procedure, with a paired t-test comparing correct-option perplexity before and after fine-tuning.Source: Fine-tuning on neuroscience corpora · p. 32

Relative perplexities evaluated at token and sentence levels

The authors tracked relative perplexity across altered and original abstracts at token and sentence levels, finding that early classifications often persisted but could change later.

Relative perplexities evaluated at token and sentence levels

On the relationship between performance, model size and training data

The authors compared BrainBench performance across models of different sizes and training backgrounds. Their results suggest that model size and training-set quality may work together, while leaving scaling behavior unresolved.

On the relationship between performance, model size and training data

  • Models starting at 7 billion parameters performed similarly, and larger models did not necessarily achieve higher accuracy.The authors also report that fine-tuned models performed worse.Source: On the relationship between… · p. 33
  • 63% accuracy on BrainBench brought GPT-2 (124M), trained from scratch on the study’s neuroscience data, close to human-level performance but below bigger models.Source: On the relationship between… · p. 33
  • GPT-2 large achieved 72% accuracy, reducing the gap to the larger tested models.Source: On the relationship between… · p. 33
  • TinyLlama achieved 70.5% (base) and 65.5% (chat), while Phi-3 achieved 82.5% accuracy on BrainBench.The authors attribute Phi-3’s performance to Microsoft’s curated high-quality training data.Source: On the relationship between… · pp. 33–34
  • The authors posit that superhuman BrainBench performance may reflect synergy between model size and training-set quality, and raise questions about scaling, diminishing returns, and resource trade-offs.Source: On the relationship between… · p. 34

Tables

The tables document model training and release timing, training-data sources, calibration analyses, LoRA fine-tuning journals, and author contributions.

Tables

  • Table S.1 overviews the training cutoff and initial release dates of the evaluated LLMs.The caption does not provide the individual dates.Source: Table S.1 · p. 35
  • Table S.2 catalogs which data sources were included in AI model training sets.The caption does not report source-specific inclusion values.Source: Table S.2 · p. 35
  • Table S.3 describes logistic fits relating model perplexity differences or human confidence to test-case correctness.The caption reports significant positive correlations, suggesting calibration for both models and humans.Source: Table S.3 · p. 36
  • Table S.4 lists journals used for LoRA fine-tuning, drawing abstracts and full articles published from 2002 to 2022.The caption does not name the journals.Source: Table S.4 · p. 37
  • Table S.5 breaks down author contributions by task, with authors earlier in the list contributing more.Source: Table S.5 · p. 38

Figures

The figures report BrainBench performance patterns across models, subfields, and context conditions, alongside analyses of calibration, difficulty, memorization, and fine-tuning.

Figures

  • Models with at least 7 billion parameters performed similarly, while performance also varied with model size and training-data quality and relevance.GPT-2 models trained from scratch on neuroscience literature improved with size; Phi3 was competitive, while TinyLlama lagged.Source: Figure S.2 · p. 40
  • Removing abstract context impaired LLM performance, and swapped context performed worse than coherent context, with further decline when context was absent.The figure captions attribute the advantage to integrating information across the abstract.Source: Figure S.3 · p. 41, Figure S.4 · p. 42
  • The memorization analysis found no distinct signature for pretraining data overall, although repeated items such as the Gettysburg Address showed signs of memorization.Low-frequency training data showed no sign of memorization; the analysis covered the three largest evaluated LLMs.Source: Figure S.7 · p. 45
  • LoRA fine-tuning fixes pretrained weights and adds trainable low-rank matrices, reducing computational burden while achieving performance comparable to training from scratch.The figure describes tuning a 7-billion-parameter LLaMa2 model on neuroscience abstracts.Source: Figure S.18 · p. 52
  • LLMs outperformed human experts across neuroscience subfields, and base models outperformed conversationally tuned chat and instruct versions.The subfield comparison reports mean performance for 15 LLMs and human experts using GPT-4-created test cases.Source: Figure S.20 · p. 53, Figure S.19 · p. 53
  • For GPT-4-created test cases, LLM difficulty rankings had a higher average Spearman correlation with human experts’ rankings than human experts’ rankings did with LLMs: 0.83 (±0.06) versus 0.27 (±0.05).Relative test difficulty was determined by the difference in perplexity between incorrect and correct abstracts.Source: Figure S.23 · p. 56
  • Higher confidence corresponded to higher accuracy for both human experts and LLMs, indicating calibrated BrainBench judgments.The figure caption notes that calibration is beneficial for human-machine teams.Source: Figure S.24 · p. 57
  • BrainBench presents original and altered abstract findings as alternatives, with participants selecting the original; the interface lets one click select all options of one color.Source: Figure S.1 · p. 39, Figure S.25 · p. 58

Related papers

Loading 2403.03230v4…