I have papers to review by November 17

Speaking of reviews: Prometheus: Inducing Fine-grained Evaluation Capability in Language Models

• Recently, using a powerful proprietary Large Language Model (LLM) (e.g., GPT-4) as an evaluator for long-form responses has become the de facto standard. However, for practitioners with large-scale evaluation tasks and custom criteria in consideration (e.g., child-readability), using proprietary LLMs as an evaluator is unreliable due to the closed-source nature, uncontrolled versioning, and prohibitive costs. In this work, we propose Prometheus, a fully open-source LLM that is on par with GPT-4’s evaluation capabilities when the appropriate reference materials (reference answer, score rubric) are accompanied. We first construct the Feedback Collection, a new dataset that consists of 1K fine-grained score rubrics, 20K instructions, and 100K responses and language feedback generated by GPT-4. Using the Feedback Collection, we train Prometheus, a 13B evaluator LLM that can assess any given long-form text based on customized score rubric provided by the user. Experimental results show that Prometheus scores a Pearson correlation of 0.897 with human evaluators when evaluating with 45 customized score rubrics, which is on par with GPT-4 (0.882), and greatly outperforms ChatGPT (0.392). Furthermore, measuring correlation with GPT-4 with 1222 customized score rubrics across four benchmarks (MT Bench, Vicuna Bench, Feedback Bench, Flask Eval) shows similar trends, bolstering Prometheus’s capability as an evaluator LLM. Lastly, Prometheus achieves the highest accuracy on two human preference benchmarks (HHH Alignment & MT Bench Human Judgment) compared to open-sourced reward models explicitly trained on human preference datasets, highlighting its potential as an universal reward model. We open-source our code, dataset, and model at this https URL.

GPT Agents

• Had a good discussion with Jimmy and Shimei yesterday about bias and the chess model. In chess, white always moves first. That’s bias. Trying to get the model to get black to move first is hard and maybe impossible. That, and other chess moves that are more common and less so might be a good way to evaluate how successful treating bias in a model could be, without destroying them.
• My personal thought is that there may need to be either “mapping functions” that are attached to the prompt vector that steer the machine in certain ways, or even entire models who’s purpose is to detect and mitigate bias.
• Started on getting ecco data to build maps. I need to install the project so that it can be edited, since I’m going to have to tweak. Here’s how: pip.pypa.io/en/latest/topics/local-project-installs

SBIRs

• Need to add references to figures in the white paper – done
• 10:30 IPT meeting – just getting ducks in a row

SBIRs

• Review fixes to white paper
• 9:00 standup – done
• 11:30 CSC touch point – done
• 12:00 SEG ROM discussion – done
• 1:00 JSC meeting
• 3:30 M30 meeting

GPT Agents

• 2:00 Weekly meeting
• Created a Huggingface model card for the chess model to use with map research. I think I’m going to try to build color maps for each layer as tokens are generated and see how they change as a game is generated

Clean bathroom!

SBIRs

• 2:00 Weekly MDA meeting
• Had a good discussion yesterday with Rukan about the demo app. Made the cutest wireframe you have ever seen using emojis
• Got the Organizational Lobotomy story accepted into Future Tense!

GPT Agents

• Finished my IUI bidding

This weekend’s fun:

SBIR’s

• 9:00 Demos – need to make slides!
• 11:30 Rocket review
• 2:30 AI Ethics
• 3:00 M30 meeting
• 4:00 Sprint planning

GPT Agents

• Started bidding on the IUI papers. Need to see if there are any more “reluctant”

Had a good discussion with Shimei and Jimmy yesterday about Language Models Represent Space and Time. Basically the idea that the model itself should have the relative representation if information in it and that could be available. The token embeddings are a kind of direction, after all.

Tasks

• Call Jim Donnie’s – done
• Call Nathan – done
• Chores
• Load Garmin (done) and laptop (on thumb drive)
• Pack!
• Bennie note

SBIRs

• Cancelled group dinner

GPT Agents

• Wrote up some thoughts about mapping using the LLM itself here

Take a look at the IUI abstracts and maybe put together a sortable spreadsheet?

SBIRs

• 9:00 standup
• See if the (relative?) ship position data Loren used to create his FOM curves can be incorporated as input data in our app
• First read of Language Models Represent Space and Time – done. Boy is there a backlash on Xitter
  • Found this in the citations: Mapping Language Models to Grounded Conceptual Spaces
    • A fundamental criticism of text-only language models (LMs) is their lack of grounding—that is, the ability to tie a word for which they have learned a representation, to its actual use in the world. However, despite this limitation, large pre-trained LMs have been shown to have a remarkable grasp of the conceptual structure of language, as demonstrated by their ability to answer questions, generate fluent text, or make inferences about entities, objects, and properties that they have never physically observed. In this work we investigate the extent to which the rich conceptual structure that LMs learn indeed reflects the conceptual structure of the non-linguistic world—which is something that LMs have never observed. We do this by testing whether the LMs can learn to map an entire conceptual domain (e.g., direction or colour) onto a grounded world representation given only a small number of examples. For example, we show a model what the word “left” means using a textual depiction of a grid world, and assess how well it can generalise to related concepts, for example, the word “right”, in a similar grid world. We investigate a range of generative language models of varying sizes (including GPT-2 and GPT-3), and see that although the smaller models struggle to perform this mapping, the largest model can not only learn to ground the concepts that it is explicitly taught, but appears to generalise to several instances of unseen concepts as well. Our results suggest an alternative means of building grounded language models: rather than learning grounded representations “from scratch”, it is possible that large text-only models learn a sufficiently rich conceptual structure that could allow them to be grounded in a data-efficient way.
  • Understanding intermediate layers using linear classifier probes
    • Neural network models have a reputation for being black boxes. We propose to monitor the features at every layer of a model and measure how suitable they are for classification. We use linear classifiers, which we refer to as “probes”, trained entirely independently of the model itself. This helps us better understand the roles and dynamics of the intermediate layers. We demonstrate how this can be used to develop a better intuition about models and to diagnose potential problems. We apply this technique to the popular models Inception v3 and Resnet-50. Among other things, we observe experimentally that the linear separability of features increase monotonically along the depth of the model.
• Slides for demos

GPT agents

• 2:00 Meeting
• Send story to CACM and see if they would like to pursue and what the lead times are – done
• Worked a bit on Neema’s Senate testimony

The bidding phase of IUI 2024 is now open. Now my present/future self has to live up to the commitments made by me in the past.

Just got back from the excellent Digital Platforms and Societal Harms IEEE event at American University. Some of the significant points that were discussed over the past two days:

• Moderation is hard. Determining, for example, what is hate speech in the ten seconds or so allocated to moderators is mostly straightforward but often complicated and very dependent of locale and culture. I get the feeling that – based on examining content alone – machine learning could easily take care of 50% or so, particularly if you just decide to lump in satire and mockery. Add network analysis and you could probably be more sophisticated and get up to 70%? Handling the remaining 30% is a crushing job that would send most normal people running. Which means that the job of moderating for unacceptable content is its own form of exploitation.
• Governments that were well positioned to detect and disrupt organizations like ISIS are no better prepared than a company like Meta when it comes to handling radical extremists from within the dominant culture that produced the company. In the US, that’s largely white and some variant of Christian. I’d assume that in China the same pattern exists for their dominant group.
• There is a sense that all of our systems are reactive. That they only come into play when something has happened, not before something happens. Intervention for someone who is radicalizing requires human intervention. Which means it’s expensive and hard to scale. Moonshot is working to solve this problem, and has made surprisingly good progress, so there may be ways to make this work.
• Militant accelerationism, or hastening societal collapse, is a thing. The exploitation of vulnerable people to become expendable munitions is being attempted by online actors. Generative AI will be a tool for these people, if it isn’t already.
• There are quite a few good databases, but they are so toxic that they are largely kept in servers that are isolated from the internet to a greater or lesser degree. Public repositories are quite rare.
• The transformation of Twitter to X is a new, very difficult problem. Twitter built up so much social utility as, for example, early warning, or reports from disaster areas that it can’t be removed from an App Store in the same way that an app that permits similar toxic behavior but only has 25 users can be. No one seems to have a good answer for this.
• The Fediverse also appears to complicate harm tracking and prevention. Since there is no single source, how do you pull your Mastodon App if some people are accessing (possibly blacklisted) servers hosting hate speech? Most people are using the app for productive reasons. Now what?
• Removing content doesn’t remove the person making the content. Even without any ability to post, or even with full bans from a platform, they can still search for targets and buy items that can enable them to cause harm in the real world. This is why moderation is only the lowest bar. Detection and treatment should be a goal.
• Of course all these technologies are two edged swords. Detection and treatment in an authoritarian situation might mean finding reporters or human rights activist and imprisoning them.
• The organizers are going to make this a full conference next year, with a call for papers and publication, so keep an eye on this space if you’re interested: https://tech-forum.computer.org/societal-harms-2023/

SBIRs

• The War Elephants paper got a hard reject. Need to talk to Aaron to see How to proceed. Done
• Add ASRC to letterhead – Done
• Expense report! Done
• Had a good chat with Rukan about using the SimAccel for interactive analysis of trajectories and FOM curves
• Work on Senate story

GPT Agents

• 3:00 Alden meeting Nope
• Gotta get back to maps. Found this:
• Language Models Represent Space and Time
• The capabilities of large language models (LLMs) have sparked debate over whether such systems just learn an enormous collection of superficial statistics or a coherent model of the data generating process — a world model. We find evidence for the latter by analyzing the learned representations of three spatial datasets (world, US, NYC places) and three temporal datasets (historical figures, artworks, news headlines) in the Llama-2 family of models. We discover that LLMs learn linear representations of space and time across multiple scales. These representations are robust to prompting variations and unified across different entity types (e.g. cities and landmarks). In addition, we identify individual “space neurons” and “time neurons” that reliably encode spatial and temporal coordinates. Our analysis demonstrates that modern LLMs acquire structured knowledge about fundamental dimensions such as space and time, supporting the view that they learn not merely superficial statistics, but literal world models.
• Because the world is mean, the paper cites two papers from 2022 on reconstructing the game board from knowledge in the model with Chess and Othello. My paper did this in 2020. Grumble

Day 2 of the Digital Platforms and Societal Harms IEEE event

My poster:

Logically combines advanced AI with one of the world’s largest dedicated fact-checking teams. We help governments, businesses, and enterprise organizations uncover and address harmful misinformation and deliberate disinformation online.

SBIRs

• 4:00 Technical Fellows meeting. Try to remember this time. Done! Went… ok? I think they are looking for something very specific.
• Made a cover

GPT Agents

• Fold in the new Informed Consent and add a link to a pdf of the doc – done!

Different ways to do learning that used to be RL. Need to look at this, and there is a repo.

SBIRs

• 9:00 standup – done
• 11:30 Touchpoint – nope
• Finish poster and submit. Need a V2 – done!

GPT Agents

• Finish training – done!
• 2:00 Meeting – need to update the Informed consent – Rewrote it. Now I need to stuff all that text on the webpage. Ugh.

Dinner with Greg at 6:00

SBIRs

• Continue poster. Made a lot of synthetic art “in the style of Francis Bacon,” which captures the mood nicely. Really like this one, even though it’s not what I was after:

• No meetings today?!
• Fix headers and send WP to Lauren

GPT Agents

• Finish CITI training
• Start fixing IRB submission

The radar patterns are still pretty confusing. Going to try to get out at 11:00

SBIRs

• Ping Bob S. About contacting SEG – done
• Added some text to the M30 doc and made a tabel of options
• Start poster
• 12:30 JSC discussion
• 2:30 AI Ethics
• 3:00 M30 Meeting

GPT Agents

• Do more CITI training. One more down

It’s Fall! Ping Nathan

SBIRs

• Make a test_harness.tex file – done
• Finish venues – done
• Slides! – done
• 9:00 Sprint demos – done
• Stories! – done
• 2:00 MDA
• 3:00 Sprint planning

GPT Agents

• Test out on people?

Very rainy weekend. Worked a bit on the book, and on trimming long CVs

Connecting Large Language Models with Evolutionary Algorithms Yields Powerful Prompt Optimizers

• Large Language Models (LLMs) excel in various tasks, but they rely on carefully crafted prompts that often demand substantial human effort. To automate this process, in this paper, we propose a novel framework for discrete prompt optimization, called EvoPrompt, which borrows the idea of evolutionary algorithms (EAs) as they exhibit good performance and fast convergence. To enable EAs to work on discrete prompts, which are natural language expressions that need to be coherent and human-readable, we connect LLMs with EAs. This approach allows us to simultaneously leverage the powerful language processing capabilities of LLMs and the efficient optimization performance of EAs. Specifically, abstaining from any gradients or parameters, EvoPrompt starts from a population of prompts and iteratively generates new prompts with LLMs based on the evolutionary operators, improving the population based on the development set. We optimize prompts for both closed- and open-source LLMs including GPT-3.5 and Alpaca, on 9 datasets spanning language understanding and generation tasks. EvoPrompt significantly outperforms human-engineered prompts and existing methods for automatic prompt generation by up to 25% and 14% respectively. Furthermore, EvoPrompt demonstrates that connecting LLMs with EAs creates synergies, which could inspire further research on the combination of LLMs and conventional algorithms.