̽��ֱ�� of Cambridge - Katherine Collins

New open-source platform allows users to evaluate performance of AI-powered chatbots

sc604 — Tue, 04 Jun 2024 10:34:36 +0000

A team of computer scientists, engineers, mathematicians and cognitive scientists, led by the ̽��ֱ�� of Cambridge, developed an open-source evaluation platform called CheckMate, which allows human users to interact with and evaluate the performance of large language models (LLMs).

̽��ֱ��researchers tested CheckMate in an experiment where human participants used three LLMs – InstructGPT, ChatGPT and GPT-4 – as assistants for solving undergraduate-level mathematics problems.

̽��ֱ��team studied how well LLMs can assist participants in solving problems. Despite a generally positive correlation between a chatbot’s correctness and perceived helpfulness, the researchers also found instances where the LLMs were incorrect, but still useful for the participants. However, certain incorrect LLM outputs were thought to be correct by participants. This was most notable in LLMs optimised for chat.

̽��ֱ��researchers suggest models that communicate uncertainty, respond well to user corrections, and can provide a concise rationale for their recommendations, make better assistants. Human users of LLMs should verify their outputs carefully, given their current shortcomings.

̽��ֱ��results, reported in the Proceedings of the National Academy of Sciences (PNAS), could be useful in both informing AI literacy training, and help developers improve LLMs for a wider range of uses.

While LLMs are becoming increasingly powerful, they can also make mistakes and provide incorrect information, which could have negative consequences as these systems become more integrated into our everyday lives.

“LLMs have become wildly popular, and evaluating their performance in a quantitative way is important, but we also need to evaluate how well these systems work with and can support people,” said co-first author Albert Jiang, from Cambridge’s Department of Computer Science and Technology. “We don’t yet have comprehensive ways of evaluating an LLM’s performance when interacting with humans.”

̽��ֱ��standard way to evaluate LLMs relies on static pairs of inputs and outputs, which disregards the interactive nature of chatbots, and how that changes their usefulness in different scenarios. ̽��ֱ��researchers developed CheckMate to help answer these questions, designed for but not limited to applications in mathematics.

“When talking to mathematicians about LLMs, many of them fall into one of two main camps: either they think that LLMs can produce complex mathematical proofs on their own, or that LLMs are incapable of simple arithmetic,” said co-first author Katie Collins from the Department of Engineering. “Of course, the truth is probably somewhere in between, but we wanted to find a way of evaluating which tasks LLMs are suitable for and which they aren’t.”

̽��ֱ��researchers recruited 25 mathematicians, from undergraduate students to senior professors, to interact with three different LLMs (InstructGPT, ChatGPT, and GPT-4) and evaluate their performance using CheckMate. Participants worked through undergraduate-level mathematical theorems with the assistance of an LLM and were asked to rate each individual LLM response for correctness and helpfulness. Participants did not know which LLM they were interacting with.

̽��ֱ��researchers recorded the sorts of questions asked by participants, how participants reacted when they were presented with a fully or partially incorrect answer, whether and how they attempted to correct the LLM, or if they asked for clarification. Participants had varying levels of experience with writing effective prompts for LLMs, and this often affected the quality of responses that the LLMs provided.

An example of an effective prompt is “what is the definition of X” (X being a concept in the problem) as chatbots can be very good at retrieving concepts they know of and explaining it to the user.

“One of the things we found is the surprising fallibility of these models,” said Collins. “Sometimes, these LLMs will be really good at higher-level mathematics, and then they’ll fail at something far simpler. It shows that it’s vital to think carefully about how to use LLMs effectively and appropriately.”

However, like the LLMs, the human participants also made mistakes. ̽��ֱ��researchers asked participants to rate how confident they were in their own ability to solve the problem they were using the LLM for. In cases where the participant was less confident in their own abilities, they were more likely to rate incorrect generations by LLM as correct.

“This kind of gets to a big challenge of evaluating LLMs, because they’re getting so good at generating nice, seemingly correct natural language, that it’s easy to be fooled by their responses,” said Jiang. “It also shows that while human evaluation is useful and important, it’s nuanced, and sometimes it’s wrong. Anyone using an LLM, for any application, should always pay attention to the output and verify it themselves.”

Based on the results from CheckMate, the researchers say that newer generations of LLMs are increasingly able to collaborate helpfully and correctly with human users on undergraduate-level maths problems, as long as the user can assess the correctness of LLM-generated responses. Even if the answers may be memorised and can be found somewhere on the internet, LLMs have the advantage of being flexible in their inputs and outputs over traditional search engines (though should not replace search engines in their current form).

While CheckMate was tested on mathematical problems, the researchers say their platform could be adapted to a wide range of fields. In the future, this type of feedback could be incorporated into the LLMs themselves, although none of the CheckMate feedback from the current study has been fed back into the models.

“These kinds of tools can help the research community to have a better understanding of the strengths and weaknesses of these models,” said Collins. “We wouldn’t use them as tools to solve complex mathematical problems on their own, but they can be useful assistants if the users know how to take advantage of them.”

̽��ֱ��research was supported in part by the Marshall Commission, the Cambridge Trust, Peterhouse, Cambridge, ̽��ֱ��Alan Turing Institute, the European Research Council, and the Engineering and Physical Sciences Research Council (EPSRC), part of UK Research and Innovation (UKRI).

Reference:
Katherine M Collins, Albert Q Jiang, et al. ‘Evaluating Language Models for Mathematics through Interactions.’ Proceedings of the National Academy of Sciences (2024). DOI: 10.1073/pnas.2318124121

Researchers have developed a platform for the interactive evaluation of AI-powered chatbots such as ChatGPT.

Anyone using an LLM, for any application, should always pay attention to the output and verify it themselves

Albert Jiang

da-kuk via Getty Images

Chatbot

̽��ֱ��text in this work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License. Images, including our videos, are Copyright © ̽��ֱ�� of Cambridge and licensors/contributors as identified. All rights reserved. We make our image and video content available in a number of ways – on our main website under its Terms and conditions, and on a range of channels including social media that permit your use and sharing of our content under their respective Terms.

Yes

How sure is sure? Incorporating human error into machine learning

sc604 — Wed, 09 Aug 2023 23:40:58 +0000

Human error and uncertainty are concepts that many artificial intelligence systems fail to grasp, particularly in systems where a human provides feedback to a machine learning model. Many of these systems are programmed to assume that humans are always certain and correct, but real-world decision-making includes occasional mistakes and uncertainty.

Researchers from the ̽��ֱ�� of Cambridge, along with ̽��ֱ��Alan Turing Institute, Princeton, and Google DeepMind, have been attempting to bridge the gap between human behaviour and machine learning, so that uncertainty can be more fully accounted for in AI applications where humans and machines are working together. This could help reduce risk and improve trust and reliability of these applications, especially where safety is critical, such as medical diagnosis.

̽��ֱ��team adapted a well-known image classification dataset so that humans could provide feedback and indicate their level of uncertainty when labelling a particular image. ̽��ֱ��researchers found that training with uncertain labels can improve these systems’ performance in handling uncertain feedback, although humans also cause the overall performance of these hybrid systems to drop. Their results will be reported at the AAAI/ACM Conference on Artificial Intelligence, Ethics and Society (AIES 2023) in Montréal.

‘Human-in-the-loop’ machine learning systems – a type of AI system that enables human feedback – are often framed as a promising way to reduce risks in settings where automated models cannot be relied upon to make decisions alone. But what if the humans are unsure?

“Uncertainty is central in how humans reason about the world but many AI models fail to take this into account,” said first author Katherine Collins from Cambridge’s Department of Engineering. “A lot of developers are working to address model uncertainty, but less work has been done on addressing uncertainty from the person’s point of view.”

We are constantly making decisions based on the balance of probabilities, often without really thinking about it. Most of the time – for example, if we wave at someone who looks just like a friend but turns out to be a total stranger – there’s no harm if we get things wrong. However, in certain applications, uncertainty comes with real safety risks.

“Many human-AI systems assume that humans are always certain of their decisions, which isn’t how humans work – we all make mistakes,” said Collins. “We wanted to look at what happens when people express uncertainty, which is especially important in safety-critical settings, like a clinician working with a medical AI system.”

“We need better tools to recalibrate these models, so that the people working with them are empowered to say when they’re uncertain,” said co-author Matthew Barker, who recently completed his MEng degree at Gonville & Caius College, Cambridge. “Although machines can be trained with complete confidence, humans often can’t provide this, and machine learning models struggle with that uncertainty.”

For their study, the researchers used some of the benchmark machine learning datasets: one was for digit classification, another for classifying chest X-rays, and one for classifying images of birds. For the first two datasets, the researchers simulated uncertainty, but for the bird dataset, they had human participants indicate how certain they were of the images they were looking at: whether a bird was red or orange, for example. These annotated ‘soft labels’ provided by the human participants allowed the researchers to determine how the final output was changed. However, they found that performance degraded rapidly when machines were replaced with humans.

“We know from decades of behavioural research that humans are almost never 100% certain, but it’s a challenge to incorporate this into machine learning,” said Barker. “We’re trying to bridge the two fields so that machine learning can start to deal with human uncertainty where humans are part of the system.”

̽��ֱ��researchers say their results have identified several open challenges when incorporating humans into machine learning models. They are releasing their datasets so that further research can be carried out and uncertainty might be built into machine learning systems.

“As some of our colleagues so brilliantly put it, uncertainty is a form of transparency, and that’s hugely important,” said Collins. “We need to figure out when we can trust a model and when to trust a human and why. In certain applications, we’re looking at probability over possibilities. Especially with the rise of chatbots, for example, we need models that better incorporate the language of possibility, which may lead to a more natural, safe experience.”

“In some ways, this work raised more questions than it answered,” said Barker. “But even though humans may be miscalibrated in their uncertainty, we can improve the trustworthiness and reliability of these human-in-the-loop systems by accounting for human behaviour.”

̽��ֱ��research was supported in part by the Cambridge Trust, the Marshall Commission, the Leverhulme Trust, the Gates Cambridge Trust and the Engineering and Physical Sciences Research Council (EPSRC), part of UK Research and Innovation (UKRI).

Reference:
Katherine M Collins et al. ‘Human Uncertainty in Concept-Based AI Systems.’ Paper presented at the Sixth AAAI/ACM Conference on Artificial Intelligence, Ethics and Society (AIES 2023), August 8-10, 2023. Montréal, QC, Canada.

Researchers are developing a way to incorporate one of the most human of characteristics – uncertainty – into machine learning systems.

Uncertainty is central in how humans reason about the world but many AI models fail to take this into account

Katherine Collins

PeopleImages via Getty Images

Futuristic image of a doctor looking at brain scans

̽��ֱ��text in this work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License. Images, including our videos, are Copyright © ̽��ֱ�� of Cambridge and licensors/contributors as identified. All rights reserved. We make our image and video content available in a number of ways – as here, on our main website under its Terms and conditions, and on a range of channels including social media that permit your use and sharing of our content under their respective Terms.

Yes