AI Engineer World's Fair 2024

LLM Scientific Reasoning: How to Make AI Capable of Nobel Prize Discoveries

Loading slides…

Read the talk

LLM Scientific Reasoning: How to Make AI Capable of Nobel Prize Discoveries

A petunia experiment that produced the opposite of its intended result exposes a harder task for RAG: connecting scattered observations into a grounded scientific hypothesis.

From a talk by Hubert Misztela

Before you start: Basic familiarity with embeddings and retrieval-augmented generation is helpful; no biology background is required.

Why did adding a gene remove the color?

Could an LLM connect unexpected biological results before scientists recognize their common cause? Hubert Misztela approaches that question from generative molecular design at Novartis, where he works with small molecular graphs. His scientific-discovery project with medicinal chemist Derek Lowe starts with a particularly useful puzzle: researchers trying to intensify petunia flower color obtained the opposite result.

In the early 1990s, scientists overexpressed a gene expecting stronger pigmentation. Instead, the flowers lost purple pigment. Similar surprises appeared in other biological subdomains, but the phenomena had different names. Researchers could see that something was interfering with gene activity without knowing the shared explanation.

Unintuitive Phenomena slide with purple, variegated, and mostly white flowers and text describing the unexpected loss of pigment.Open full source frame
Gene overexpression unexpectedly produced flowers with less purple pigment.

In Misztela’s account, connecting and resolving the three phenomena took roughly eight years and led to a Nobel Prize in medicine. That delay creates a concrete discovery task: find a plausible common cause from the literature available before the answer was known. Perhaps the necessary clues were already present, scattered across disciplines. Asking a modern model for the answer is easy; determining whether it can assemble those clues is harder.

0:180:34
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

0:18 · section reference included

Retrieval inherits the difficulty of the question

Ordinary retrieval-augmented generation starts by processing documents and representing them as embeddings. At query time, it embeds the question in the same latent space, retrieves nearby chunks, and supplies them to a response generator. This works naturally when the question points toward text that already contains an answer.

Different RAG improvements repair different parts of that pipeline. HyDE, ranking, and document indexing are not interchangeable: improving the match between a question and answer-bearing text is a different problem from selecting the best documents or chunks. But those improvements still depend on understanding what evidence the question requires. A convoluted question can defeat that understanding before retrieval begins. Misztela describes this as “trying to run before we crawl.”

Question and data processing paths converge on retrieval and response, with augmentation references and a paper title at upper right.Open full source frame
The naive RAG pipeline alongside “Easy Problems That LLMs Get Wrong.”
3:193:30
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

3:19 · section reference included

Reason over the question and the representation

Reasoning can happen before retrieval on either side of the search. Routing interprets a question before deciding where to send it; the query cannot simply be thrown at an embedding index and expected to produce the right evidence. On the document side, GraphRAG and GraphReader represent knowledge using graphs. The shared architectural move is to process meaning before selecting material, whether by understanding the request or by changing how document knowledge is represented.

After retrieval, an LLM usually performs the final reasoning, but that is a design choice rather than a requirement. The more semantically complex the question, the more work may be needed before retrieval as well. A useful pipeline therefore becomes reasoning → embedding → retrieval → reasoning, with each reasoning stage chosen for a specific job.

This resembles the molecular-design workflow Misztela already uses. An encoder–decoder model, such as a variational autoencoder, represents molecular graphs in a latent space. The system then navigates that space to find points useful for a desired objective. RAG also selects points from an embedding space, but usually through similarity to a question.

WorkflowRepresentationSelection objective
Standard RAGDocument embeddingsSimilarity to the question
Molecular designMolecular latent spaceOptimize a desired property

The analogy broadens what counts as reasoning: goal-directed optimization over a latent space can perform part of the job. The system need not express every operation as an LLM-generated chain of thought.

4:465:04
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

4:46 · section reference included

How many concepts must the evidence connect?

A needle-in-a-haystack test asks a system to locate a particular piece of information inside a large context. Misztela uses the example of finding the ingredient for a perfect pizza. A multi-needle test extends the task to several pieces of information. But counting answer fragments captures only part of the difficulty: the relationship between the question and those fragments matters too.

The database analogy makes that relationship explicit:

Question conceptsEvidence locationsRelationship
OneOneOne-to-one
OneSeveralOne-to-many
SeveralOneMany-to-one
SeveralSeveralMany-to-many

A question can combine several concepts that must jointly identify one relevant passage, or it can require concepts scattered across multiple chunks. In either case, searching for the whole question as one undifferentiated vector may hide the structure that determines what to retrieve.

Multiple concepts are a signal to consider reasoning before retrieval. First determine what the question asks the system to connect; then decide which embeddings to retrieve and how to assemble their evidence. The petunia problem is difficult precisely because the desired answer must connect observations that originated in separate biological subdomains.

8:058:17
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

8:05 · section reference included

Choose the operation, then choose its executor

Reasoning here means processing information in a logical way, not merely asking for chain of thought. That includes several distinct operations:

  • Aggregation and arithmetic: combine observations or calculate over data.
  • Logical and causal reasoning: deduce relationships, or extract entities and causal links from which further hypotheses can be proposed.
  • Algorithmic reasoning: carry out a defined computational procedure.
  • Probabilistic reasoning: update beliefs under uncertainty; Misztela describes attempts to make LLMs reason in a Bayesian fashion.
  • Structured reasoning: use compositional relationships within a structure, which need not be causal.
  • Geometric reasoning: infer transformations and relationships in visual arrangements, as in ARC.

The historical ARC Prize launch offered a $1 million-plus prize pool for open solutions to ARC-AGI. Its geometric tasks illustrate how far reasoning can extend beyond familiar text questions.

Nine teal hexagons labeled Simple Aggregation, Algebraic, Logical, Causal, Algorithmic, Probabilistic, Structured, Geometric, and Others, beside an arcprize.org link.Open full source frame
Reasoning types extend from aggregation and algebra to causal, probabilistic, and geometric reasoning.

Expecting a single LLM to perform all these operations well is an unnecessary architectural constraint. A causal-inference library can handle a causal computation. An agent can generate Python and send it to a REPL for algorithmic work. The LLM may interpret the task and select a tool while the specialist tool executes the operation. That division applies to reasoning after retrieval just as much as to reasoning before it.

10:0210:15
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

10:02 · section reference included

Build for one question—and a historical evidence boundary

Return to the biological problem: find the common cause of three phenomena. This requires causal reasoning and a retrieval strategy that can connect multiple observations. It also changes the economics of the system. The goal is not to answer many different questions over a fixed dataset. It is to answer one persistent scientific question, processing the dataset repeatedly to extract every relevant theme.

The answer is already part of modern scientific knowledge: RNA interference, whose discovery Misztela dates to 1998. A model trained on sources such as Wikipedia may know it before receiving any retrieved papers. Groundedness therefore matters as much as retrieval quality. A relevance classifier that examines the entire dataset can be attractive when exhaustive processing is affordable, even if it costs more than embedding search. The architecture depends on the required reasoning, the question–answer relationship, and the tradeoff between computation and grounding.

The proposed historical experiment follows a short sequence:

  1. State the scientific question and classify the reasoning and evidence relationships it requires.
  2. Restrict the retrieval corpus to papers published before the discovery.
  3. Reconstruct the pre-discovery setting and test whether the system can assemble a useful explanation from that evidence.

Misztela calls the restriction a knowledge cutoff. It limits the retrieved evidence, but it does not remove the answer from the model’s pretrained knowledge. This distinction becomes consequential in the experiments, where the first output does draw on later knowledge.

Working repeatedly on this question resembles training an agent on a particular scientific problem, in the sense of developing and testing its approach against that setting; no weight-training procedure is specified. The larger ambition is to reuse the same discovery scheme across multiple scientific problems.

12:1612:33
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

12:16 · section reference included

Recover facts, connect them, then explain them

Success has several levels. The first is to recover what the corpus already knows: relevant facts, existing hypotheses, and relationships among them. This is broader than extracting a short answer to a narrow question. The useful output is a connected account of the evidence surrounding the problem.

Next come less obvious links, especially between the three biological subdomains. Ideally, the system would recover all relevant facts rather than only the most conspicuous ones. Beyond that lies new hypothesis generation: taking the observed relationships one step further to propose a connection that the corpus does not explicitly state. The highest ambition is to explain how the phenomenon happens, not merely identify related observations. Misztela’s joke about meeting in Sweden makes clear that this final level remains an aspiration.

Levels of Success slide showing Level 3 Advanced Hypothesis Generation and Level 4 Predictive Hypothesis Generation, with text and node-link diagrams.Open full source frame
Advanced and predictive hypothesis generation define the higher levels of discovery success.

The distinction between discovering a phenomenon and explaining its mechanism needs care. The human discovery identified an effective trigger before the complete biochemical machinery was understood. A system can therefore make meaningful progress by identifying what causes an effect without yet providing a complete account of how that effect is produced.

14:5915:14
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

14:59 · section reference included

Select a score cluster instead of an arbitrary chunk count

The experiments start with a practical retrieval question: how many chunks should be returned? Rather than choosing a fixed top k, Misztela proposes Jenks natural breaks. Group the scalar query-to-chunk similarity or distance scores into classes, considering variation within and between groups, then retrieve the best-scoring class. These are groups of scores, not clusters fitted directly to the full embedding vectors.

The highest-similarity group—or lowest-distance group—provides an adaptive chunk count. Misztela’s intuition is that a coherent top group may contain useful evidence. Low score variance alone, however, does not establish scientific informativeness, and the number of score classes still has to be chosen.

The talk mentions a Python snippet. A compact implementation of the score-selection idea can use Jenkspy. Here, higher values mean greater similarity; the chunk IDs and scores are illustrative inputs:

python

import jenkspy

chunks = [
    {"id": "paper-a:1", "similarity": 0.91},
    {"id": "paper-b:2", "similarity": 0.89},
    {"id": "paper-c:1", "similarity": 0.52},
    {"id": "paper-d:3", "similarity": 0.49},
    {"id": "paper-e:2", "similarity": 0.15},
    {"id": "paper-f:1", "similarity": 0.12},
]

scores = [chunk["similarity"] for chunk in chunks]
breaks = jenkspy.jenks_breaks(scores, n_classes=3)

# The final class is open on the left and closed on the right.
selected = [
    chunk for chunk in chunks
    if breaks[-2] < chunk["similarity"] <= breaks[-1]
]

print([chunk["id"] for chunk in selected])

This selects the members of the strongest score class rather than specifying their count in advance. For a distance measure, selection must instead target the lowest-valued class.

16:3916:58
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

16:39 · section reference included

Ask whether evidence advances a hypothesis

The initial RAG output used knowledge from after the discovery. Stronger instructions—strict prompting—reportedly improved grounding. The next experiment passed every dataset chunk through an LLM relevance classifier, asking whether it was relevant to the scientific problem. That basic classifier added little: its judgments were largely redundant with embedding distance.

The more useful change was to judge each paper in the context of hypothesis advancement. Instead of asking only whether a paper concerned the topic, the prompt asked the model to approach it more like a scientist evaluating evidence. Misztela reports that this produced the literature’s DNA-related hypotheses together with an RNA-related hypothesis, which he identifies as the correct direction.

ExperimentSelection or grounding changeReported observation
Initial RAGRetrieve supporting chunksUsed post-discovery knowledge
Strict promptingStrengthen evidence restrictionsGrounding improved
Basic relevance classifierJudge every chunk for relevanceMostly duplicated embedding distance
Hypothesis-aware assessmentJudge papers for hypothesis advancementSurfaced an RNA-related hypothesis

The important shift is the criterion for relevance. A passage can resemble the question without helping distinguish explanations; evidence that advances a hypothesis has a more specific role. Misztela reports movement toward the known result after reasoning over both the question and the dataset before retrieval. These are qualitative observations, without a numerical performance evaluation or a demonstration that pretrained answer leakage was eliminated.

Scientific discovery therefore asks more of the architecture than ordinary Q&A. Classifying the problem’s reasoning demands and question–evidence relationships helps determine where additional processing belongs; the generalized needle-in-a-haystack view is one proposed design aid, not a claim to the best architecture.

There is also a practical alternative to making embedding search ever more elaborate: when the workload permits it, test exhaustive LLM assessment of the corpus. For one persistent scientific question, spending computation on every chunk may be reasonable. Compare that approach with distance-based selection and check whether the LLM’s judgments actually improve the evidence available for discovery.

17:3417:53
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

17:34 · section reference included

Resources

From the talk

  • Python implementation of Fisher-Jenks natural breaks, with examples for calculating numeric class boundaries and groups.

Transcript download
Full transcript

Loading transcript…