Source-linked AI summary
Guided scenarios with simulated expert personae: a remarkable strategy to perform cognitive work
David Van Buren
TL;DR
The paper explores whether corpus-derived persona behaviors can help LLMs perform cognitive work, including checking factual claims and exploring a quantum-optics idea. It uses guided scenarios and simulated expert personae; in the VITAL example, the dialogue notes that cited evidence does not mention support from the named funders.
Problem
Vetted knowledge sources can make truthfulness assessment easier, but cover only a limited share of the claims LLMs can produce.
Source: §2 · p. 3Method
The paper uses guided scenarios, participants, and goals to elicit simulated persona behaviors for cognitive tasks rather than prescribing a specialist reasoning process.
Source: §2 · p. 3Results
In the VITAL example, Watson notes that the cited evidence attributes development to NASA’s Jet Propulsion Laboratory and does not mention support from NSF, the Gates Foundation, or AHA.
Source: §8.1 · p. 10Takeaways & Limitations
The paper proposes that corpus-represented persona behaviors could be assembled into teams of simulated expert assistants for a broad range of intellectual work.
Source: §1 · p. 2The quantum-optics simulation makes a mathematical error by using the wrong form for the square of a complex function, which access to a computational engine such as Wolfram Alpha could have avoided.
Source: §4 · p. 5
Abstract
from arXiv · showhide
Large language models (LLMs) trained on a substantial corpus of human knowledge and literature productively work with a large array of facts from that corpus. Surprisingly, they are also able to re-create the behaviors of personae that are captured within the corpus. By forming teams of simulated personae, supplying contexts that set the stage, and providing gentle prompts, one can move through scenarios that elicit expert behavior to perform meaningful cognitive work. The power of this strategy is demonstrated with two examples, one attacking factuality of LLM responses and the other reproducing a very recently published result in quantum optics.
1 Introduction
LLMs may draw on learned persona behaviors, not only stored facts, to perform cognitive work. The paper proposes using simulated-expert teams and staged scenarios to elicit that behavior for complex tasks.
1 Introduction
- Expert-role prompts can elicit specialist responses, and quantitative demonstrations of their advantage over simple prompts have recently been reported [3] [4].Source: §1 · p. 1
- Dialog prompts are proposed as a way to elicit believable character conversation; debate-format dialog has improved performance across benchmarks [5].The passage suggests LLMs may handle dialog well because dialog is represented in their training corpus [6].Source: §1 · p. 1
- LLMs learn both facts and behaviors associated with real and fictitious personae, alongside natural and other processes.Source: §1 · p. 1
- The paper hypothesizes that human-like behaviors learned by LLMs can be recruited as a cognitive resource for complex real-world tasks.It proposes assembling teams of simulated personae whose words are represented in the training corpus to perform intellectual work.Source: §1 · pp. 1–2
- Figure 1 contrasts a low-quality simple-prompt response with a guided simulated-expert dialog that produces a full quantum-optics solution.The guided output includes specialist text, mathematics, and Python visualization code for a topic beyond the model’s training horizon.Source: Figure 1 · p. 2
2 Simulated personae dialog
The paper applies simulated-persona dialog to checking claims and removing unsupported assertions from generated responses. It presents this exploratory work as evidence that learned behavior can support cognitive work beyond retrieving static information.
2 Simulated personae dialog
- Deconfabulation segments a response into assertions, assesses their truth, removes claims that cannot be determined true, and reassembles the remainder in the original style.Validated assertions are also placed in memory for retrieval in later context; the proposed architecture includes a semantic search mechanism and web access.Source: §2 · p. 3
- Truth evaluation is a challenging inference task, and vetted knowledge sources have limited coverage of the claims LLMs can produce [9].Source: §2 · p. 3
- The heuristic validation process searches for high-quality websites, retrieves relevant snippets for each atomic claim, and presents them to GPT-3.5-turbo with persona-focused instructions.Source: §2 · p. 3
- The Sherlock Holmes–Watson scenario elicits a purposeful detective-style dialog to assess whether evidence supports or refutes a claim.The paper favors scenario prompts over prescriptive analysis rules because the task calls for the characters’ behaviors and skills; dialog supplies a one-shot train of thought without injecting specialist knowledge.Source: §2 · p. 3
- Detection success varied by claim, and the paper describes detecting 90% of confabulations generated 90% of the time as a good-performance target.It refers readers to appendix transcripts and trial statistics rather than reporting a single overall success rate here.Source: §2 · p. 3
- The exploratory work supports the hypothesis that learned behavior is a cognitive resource distinct from static information and can be used for cognitive work.Source: §2 · p. 3
3 Guided scenarios strategy
For a topic emerging beyond the model’s training horizon, the paper describes selecting known expert personae and building a scenario to draw out their behavior. Stage directions, props, and further prompts shape and continue the response.
3 Guided scenarios strategy
- The method targets a recent topic unlikely to be represented in the training set, then uses a preliminary prompt to select known personae with relevant expertise.The selection method is described as reproducible, though it includes trained-in biases.Source: §3 · pp. 3–4
- The scenario uses minimal post-training-horizon information, stage directions, and props that provide semantic support for desired behaviors.Source: §3 · p. 4
- Further stage directions or additional simulated personae can continue the narrative until a suitable stopping point.Source: §3 · p. 4
4 Application - Exploring a new scientific idea
The authors use simulated Noether and Feynman personae to explore a recent time-domain double-slit result from its title alone. Their generated code reproduces the published result, and the authors report further predictions matching the real-world team's despite being absent from the model's training corpus and chat context.
4 Application - Exploring a new scientific idea
- The experiment explores Tirole et al.’s recent time-domain double slit from its title alone, a topic outside the LLMs’ training horizon.Source: §4 · p. 4
- The scenario sets Noether and Feynman in a physics-department lounge, gives them a whiteboard, and suggests an advanced-undergraduate starting point.Source: §4 · pp. 4–5
- The dialogue sums the two time-separated wavefunctions, makes an initial complex-square error, then returns to compute the probability correctly.The account identifies computational support such as Wolfram Alpha as a way the error could have been avoided.Source: §4 · p. 5
- For finite-duration openings, the described interference depends on opening shape and duration and on the delays between openings, with patterns varying across position and frequency.Source: §4 · p. 6
- The authors describe scenario writing and gentle guidance as requiring less effort and specialized knowledge than operating a Mathematica session.Source: §4 · p. 7
- The generated plot resembles Tirole et al.’s published result, although the simulation used different parameter values because it lacked the paper’s specific assumptions.The caption notes that the simulation selected parameters that make the fringes easily discernable.Source: §4 · p. 7, Figure 2 · p. 8
- The authors report that follow-up predictions match the real-world team's predictions despite being absent from the LLM's training corpus and chat context.Source: §4 · pp. 7–8
5 Implications
The paper argues that guided scenarios can elicit expert-like behavior from simulated personae, and proposes using persona teams and simulated tools for complex cognitive work. It projects that this capability could support broad creative and scientific work.
5 Implications
- More advanced models are expected to perform cognitive work at expert level when guided scenarios prompt simulated personae.The experiments reported here used GPT-4.Source: §5 · p. 8
- Figure 3 shows Tirole et al.’s temporal diffraction pattern, a result published after the models’ training horizon, without a derivation in this paper.The figure was modified to match Figure 2’s axes.Source: Figure 3 · p. 8
- The paper proposes assembling historical or fictional expert personae into teams and simulating tools or processes to enrich their working context.The proposed teams would perform complex cognitive tasks when given information and instructions.Source: §5 · p. 8
- The paper forecasts that this capability could contribute to creativity, problem-solving, scientific and technical advances, medicine, and the arts.These are projected possibilities, not reported outcomes.Source: §5 · pp. 8–9
8 Appendices
The appendices show a Sherlock Holmes and Watson dialogue examining a generated VITAL funding claim, alongside a small trial result. They also include ChatGPT-4-generated Python code for plotting a time-frequency interference pattern.
8.1 Simulated personae Sherlock Holmes and Watson detect a confabulation within the LLM training horizon
- Across the reviewed project and FDA evidence, Holmes and Watson find no direct support for GPT-3.5’s VITAL funding claim.The appendix describes the generated claim as a confabulation.Source: §8.1 · p. 10
- In a small test, effectiveness was over 90% when identical prompts were run on pristine GPT-3.5 chatbot instances, with responses classified manually.Source: §8.1 · p. 10
8.2 Simulated persona Feynman’s python code
- ChatGPT-4 generated Python code that combines sinc-based terms into a probability distribution and plots it across time delay and frequency.The plot labels the axes as time delay and frequency and identifies the output as a probability distribution.Source: §8.2 · p. 11