TWed Talk: "Detecting Ambiguity in Question Answering over Financial Documents using LLMs"

Key Takeaways

Oh, isn't that >> nice? I welcome everybody to our first TWED of the spring 2026 semester. Thank you, Kelsey and Claire for being great and and submitting to Tom's persistence and getting you to do it. Um, without further ado, why don't you take over take over control. >> All right. >> And I should I should say my usual thing. This is being this is being recorded. It'll be made available u through our uh our usual means >> and next week we're going to have Tisiana leading its web >> some >> and whoever else you want to rope into it. I I'll contact you later in in the week to get some information there. >> I should we should see your slides here shortly. Yeah, you would think so. >> Oh, I know it's here. It's I understand. It says >> Oh, okay. Fine. All right. In slideshow. Is that going to work? Okay, cool. All right. Uh today we're going to be talking about uh detecting ambiguity and question answering over financial documents using LLMs. Um and I'm Kelsey Rook and Deb is my advisor >> and I'm Claire Arrington and Sabella Doll is my advisor which is not so >> Yeah. >> Yep. So we're happy to have Claire here. All right. So we are currently uh funded uh by AFL under TOMC and last semester we were basically given uh the assignment to do something to improve uh the domain of finance with LLMs. Uh so we thought that we would look into financial document question answering. Uh and basically what that is is uh you have some kind of financial report like up here uh you can see an annual report from Apple. Uh you have that all parsed and and ready to go uh for an LLM to process and then uh you ask a question about the document and hopefully the LLM answers you correctly. Uh so multi hierarchy is the data set that we've been primarily looking at. Uh the question types in this data set are direct retrieval. So basically uh retrieving a single value from a table or the text. Uh numerical reasoning which involves some kind of math or comparison. and uh hybrid text plus table reasoning where uh you might need context from the text to interpret something that's in the table. Um and each uh sample includes the entire document uh including however many uh tables were in it. Um there are other data sets for this task like uh TAT QA and FIN QA but they typically um only include like a couple paragraphs of context surrounding one table and all of their questions are just about that snippet. Um so multihiret is kind of the most uh realistic uh data set for kind of a real life financial document question answering setting. So uh we started just trying to establish a baseline. So in the original uh multih higher tee paper they used um uh they evaluated some BERTbased methods initially and what worked best for them was this kind of uh Roberta largebased pipeline where they did um information extraction and then had kind of a reasoning module that those values went into. Um, and uh, as you can see, it uh, didn't do fantastic. Um, I'm actually sorry. I'm going to refresh here because some of these values are not correct. Um, let's see. I don't Okay, I'm I apologize. Something got overwritten here. Um, but I'll just tell you what the values actually should be. Um, so we ran uh Gemini 2.5 Pro and GPT 5.2 uh which are pretty recent. We released LLMs on this data set. Um, and for uh span selection questions, which uh involve little to no mathematical reasoning, it's just grabbing a section from the texture table and returning it. um we get a fairly dis uh decent like obviously there's a lot of improvement over the uh Roberta method. Um and for arithmetic problems um those values are are wrong. It should actually be like 57 uh% or uh sorry 47 47ish%. Um this these were the original values with like no tolerance for for rounding. But if you increase that tolerance by like half a percent, the performance bumps up a whole bunch just random data processing processing stuff. Anyways, uh those values should be like 47 to 50%. Um so we thought about it and we were kind of like okay so most of these questions are just basic extract the correct values from the tables, plug it into a mathematical equation. Why are the uh like why is it still unsuccessful at such a large uh subset of problems? Um well in the paper uh the human performance is only like 80 to 90%. Um and this is uh the human performance averaged over uh five financial experts who um you know uh gave their answers for a sample of the data set and interotator agreement is about 93%. So we decided to uh look at the questions themselves. Um and we have someone in our group who is pretty interested in the problem of uh query ambiguity and uh looking at multi hierarchy. We found a good amount of that. So here's a question from the data set with a little snippet of a table. Um this is just average train speed uh in miles per hour for some company like Amtrak or something over a few different years. And the question is what is the mathematical range for average train speed for 2008 to 2010? Uh ask let's yeah let's ask an LLM to uh do this problem for us. Uh it returns the range is 23.5 to 26.2 to uh miles per hour. Uh but that is wrong. That goes against what the uh data set's answer is. Um the data set's answer is 3.8 miles hour because it didn't want the boundaries of that range. It wanted the actual difference between the two ranges. But potentially those answers are still correct. in a conversational setting, if I asked an LLM uh for this answer and I let it give me a couple different answers, I would be very happy with that. Um, but for a QA data set where you're only allowing for just a single input and, uh, evaluating on that, uh, it doesn't it doesn't work out. Here's another example that I think is is really fun. Um so these two tables are for uh from a document that describes the number of uh crude oil and natural gas wells as well as uh developed and undeveloped acreage in a lot of different areas. Um and one of them is the North Sea. And the question in the data set is what is the highest total amount of North Sea? >> I don't know. Uh I guess the highest number in these tables is 213. But what if they meant to say what is the highest total amount of uh like wells in I don't know. Um it doesn't it doesn't make sense. There are uh different if you ask an LLM to just give you one answer and you evaluate that it's uh probably going to be wrong, right? Um so so yeah uh we went through uh a bit of the multih highier data set and found a lot of instances of these ambiguities. Uh so let's talk about what query ambiguity actually is. Um, an ambiguous query is basically one um that is not specific enough or uh contains some phrase such that there are multiple plausible answers. And we can contrast this with unanswerable unanswerable queries um which are either like a horribly malformed query or a query that just cannot be answered uh given the information available in the document or um in the LLM's training data that I can look at. Um, so the problem in a lot of these QA data sets is that uh, well, well, not necessarily a problem, but it's just a fact of life that ambiguity is a feature of human language. Uh, when we're just chatting, we kind of assume that we share some context with the people we're talking to. Um, and sometimes those assumptions are a little bit wrong. Uh but because of how human dialogue works, we can kind of adjust in in real time and ask clarification questions. Um and when uh communicating with LLMs, we tend to uh retain that habit. Um so to actually disambiguate in a lot of conversations, uh there needs to be some cooperative element. Um however in writing uh and evaluating data sets there kind of is not this cooperative element. Um so where has this kind of seeped into uh real research situations? Uh so in 2020 these uh researchers uh were looking at the natural questions data set which is a fairly widely used data set released by Google and it consists of actual um submitted searches and answers and they found that uh about 50% of this data set is ambiguous in some way. Uh that might be a little bit of of a reach um because as we'll see the definition of ambiguity can be a little bit loose but there certainly is a good amount of ambiguity in there. Um and there was a study just released that looked at 15 uh QA data sets in uh tabular and data analysis and text to SQL uh settings that found um also very high rates of ambiguity. Uh so this is uh kind of a a common issue. All right. Is this where you were going to take over? >> Yeah. Or you can do this one if you want to. >> Yeah. I mean >> Okay. You can go for it. >> Um so hi everyone jumping in. Um so there are a lot of different areas that people are looking into when they're talking about specifically question answering. The biggest one being um looking at domains. So things like trying to think how to put this. um just any sort of stuff related to the field of finance. You're asking questions maybe not on specific documents but you just want to know pieces of information or you might be looking at a document but not trying to do too much extra information beyond that. Um there's also things like the task oriented settings where you're trying to do something that's a little bit more complex. Um, and then you also have things looking at ambiguity types, but a lot of these that we've seen are very limited in terms of what kind of ambiguities they've defined specifically for finance, but a lot of them are also just very general forms of ambiguity. Um, so you get things like pronouns or under specificity, like not kind of clarifying the context that it's looking for information, and a lot of them stop up to that point, which is understandable. it's very hard to define ambiguity and kind of model that in a larger data set. But we looked into some additional taxonomies for that which I'll get into in a minute. Um but first I just wanted to discuss specifically in this time period currently where you have LLM, what kind of ambiguity shows up in those instances. um and what kinds of problems that we've been thinking about addressing as we work towards um looking at some of these instances of ambiguity ourselves. So, one of the biggest things is overconfidence that LLM might have when answering a question. They're not necessarily going to state that it doesn't understand exactly. It's going to give the best possible answer that it can unless you're trying to push it very strongly to make those kind of statements. And you see that in particular with things like chain of thought prompting anywhere you're trying to get it to reason. It especially might not say, "Okay, I've hit a wall. I'm going to completely stop, you know, trying to do this reasoning." Uh, additionally, similar to that notion, um, you're not going to particularly see follow-up questions unless you're again pushing that very hard. Uh, like we talked about, we were looking at multihire where it is a single instance of answering. So you aren't even been opening up that kind of conversation for a question to be asked. So that's another complication there. There's also just weak performance when identifying ambiguous questions. Just because something is wrong doesn't necessarily mean that it's ambiguous. And these models aren't necessarily, at least at the moment, although we'll talk about the different taxonomies, it isn't being built in to identify, you know, when context is missing, when jargon's being used, things of that nature. And then even more specifically, not all ambiguity comes from the query itself. It might come from the document. It might come from the LLM, things like that. So there's even worse performance when you're trying to figure out where is that ambiguity happening and trying to resolve that specifically. So, one of the big things with this is that a lot of this research is trying to take regular taxonomies of ambiguity, whether that's in different areas of research, not necessarily computer science, and trying to borrow those and then apply them more specifically. And so you have a lot of we'll go through three different examples but different ways that people have defined ambiguity taxonomies and the different kinds of um intents of the reason why they you know make them more granular, make them more specific, things like that. So one of the main things is purpose. Whether you're doing an evaluation data sets, you want to come up with a very difficult taxonomy, for example, um or you might be doing something where you're trying to have them help with reasoning and so you might not necessarily want there to be too too much because it gets a little bit harder to train something. Uh similarly, you're going to have different forms of granularity, which is similar to that point. And it might be if you're really interested in, you know, a specific domain like finance or something else like history, things like that, and the type of ambiguity that's occurring is more interesting. Um, and then finally, the method of construction. We talked about this a bit that a lot of these data sets and taxonomies are kind of iteratively being built. So the multihire was based on a RD table data set which was already based on a simple question answering data set. So you get this kind of rolling issue, you know, which is good and bad of building up data sets and introducing new instances of taxon or um of ambiguities and things like that. So there's a little bit with that that can also be complicated. Um so the first taxonomy that we looked at um or amongst the ones that we'll talk about is confuse bench. This one is a very simple three highle one where there are ambiguity types and the main thing that they were focusing on is the possible source where that ambiguity was occurring. So they have document scarcity where if you're trying to get information from a particular document that would be missing um and you're not able to answer the question for that reason. There's limited capacity where the query is too complex. there's some aspect that the LN is not able to fully grasp about the reason that somebody's asking something or about just too many small pieces of information that they're looking for. And then finally, there's the regular query ambiguity which as we're talking about this, this is kind of the main thing that's coming up um where the query itself is unclear for some particular reason. And this is discussed mostly which is we've also gone through this where it tends to be the thing that is blamed over investigating any other alternate source just because it's kind of easy to say you wrote the question wrong. There's not anything that can be done. Um but that's not necessarily always the reason why something's being uh answered incorrectly. Um on the other side you have something like the clamber data set. So this is designed to test input understanding and task completion and it has a much more granular level of information. It still has a three highlevel version where there is epistemic misalignment. And I will note that the categories are a little bit more complex because they're trying to fit a lot of different forms in there. Um again you have the linguistic ambiguity which is mostly that query ambiguity again where it might just have you know terms itself that are ambiguous with different meanings. um or the context is missing. And then this aliatoric output where a lot of it is just I'm trying to think how to say this. Um there's different forms of pieces of information that are missing contextually. So this who when where the person is assuming that the model might know that bit of information about them but that's not the case. And then finally we have this ATCOT which is the ambiguity types uh chain of thought reasoning model and this is again a very simple highle model and this one is meant to be integrated into chain of thought reasoning and so it is a very just is it something with semantics is it too generalized and is it too specific which this one for example like I mentioned with the ones before it's leaving out a lot of stuff so if you're really trying to force that the issue is one of these you might be skipping over some of those other sources of information, for example. But if you're particularly just caring about something where you're not necessarily getting documents, then maybe that's not an issue. Um, and so these kind of three different taxonomies are things that we've looked at and are working through ourselves in terms of determining, you know, what would be the right level of information as well for our specific instance. So, Okay. So, uh continuing to talk about the uh data set uh multi hierarchy, we decided to do a little bit of annotation just to uh get to analyze the amount of ambiguity in the data set uh and figure out how those ambiguities align with uh current taxonomies. uh do they cover everything in this data set and then um potentially either come up with more uh granular uh classifications of ambiguity that are specific to uh domain specific and and document question answering. Um so this isn't completely done uh at this point but we have about uh 200 samples of uh ambiguous uh data points within this data set. Um and we've been able to kind of go through and uh certain categories have emerged uh that are uh really frequent. Uh so first under specification this has kind of already occurred in uh this has been covered in other taxonomies. Uh we have a lot of colloquial phrasing that uh could align with multiple financial terms. Uh so for example uh for the question in which section does the sum of net product revenues have the highest value? Uh just the word section could refer to uh like the the quarter, the time period. It could refer to like um the the the segment like the kind of uh assets in the company that this is about. Um and then there's a lot of colloquial phrasing that could align with uh several different mathematical operations. So for example, the train example that we looked at before, uh there are also a lot of occurrences uh where it's difficult to tell um whether the answer should return some literal value from a table or interpret it uh with some more I guess uh like financial context to give a more insightful answer. Um there are also uh frequent quote unquote conflicts between um uh tables in the document versus the text where the table will give a very precise figure but the text will um present a figure that is different for uh like some particular reason or it's just rounded. Um and there's also uh a lot of instances of asking for like a rate of change. Um and the if it's a decline the model might uh give the absolute value like the magnitude of the rate of change whereas uh the gold dancer in the data set wants the uh percent as a negative. So again, this is an issue when you're just evaluating this QA data set as like a oneanddone thing uh with and like assuming that there is zero ambiguity in any of the questions. Uh so we did a um a little bit of an experiment with a sample of about uh 50 uh samples that we labeled as unambiguous. So very straightforward questions that should only have one answer. And we adapted this from the uh condig QA uh benchmark data set paper where they were trying to evaluate whether LLMs could say yes or no this question is um ambiguous. Uh so the prompt basically asks um it's kind of a modified chain of thought thing. Uh analyze the question for ambiguity. Explain different interpretations. Uh analyze for multiple possible answers. provide examples and answer in uh in JSON format. Uh and if the question contains no ambiguities, multiple answers should be a list of length one. Um in that original paper, uh we mentioned earlier that um that paper found that even with chain of thought and uh few shot prompting, it was really difficult to uh actually get good results. Um and not shockingly this happened for us too. Um in this sample of unambiguous questions uh it found multiple answers for about like 30 out of the 50 um 10 were just unparable and then only 12 of them uh did it only uh basically assess that they were unambiguous. Um so one example um we have this table uh that's uh different uh debts and assets uh relevant to apartment homes owned by some business and the question is what's the increasing rate of secured debt in 2006 uh at the end of the year? Uh so the answer is pretty clearly um I guess I should put in a caveat here about pretty clearly. We are not financial experts. Um we don't have any yeah we don't really have anyone in our group who is like specifically a financial expert. So so take that that caveat. Um but to us um the answer is pretty unambiguously just the 2006 amount of debt minus the 2005 amount of debt uh divided by the 2005 amount to get the uh increasing rate which is about 6%. Um, but given uh that prompt that we just showed uh and the LLM spat out uh four different answers uh one the first being the correct one which happens frequently. Um so that's good I guess. Um the the second answer being the absolute increase which was not what the question was asking for. It was asking for a rate. Um, and the other two are kind of just doing whatever. Uh, kind of just answering a completely different question, which again, it it isn't shocking. Even though um the the prompt said if the question contains no ambiguities, multiple answers should be a list of length one, still the LLM has been asked to find ambiguities. It's going to find ambiguities. Uh so um our our future work kind of will probably involve adding some kind of uh rating or ranking uh to hopefully differentiate between when it's just reaching for something versus uh when it actually uh like sees the the first answer it gives as the the most plausible or or the least plausible or or whatever. Uh so did you want to talk about future plans? >> You can go ahead. >> All right. Um so again we need to uh finish doing this analysis of how um existing taxonomies apply to this uh more specific domain. Um and we also want to run a few more of these experience uh experiments on actually detecting the presence of ambiguities with a few different uh prompting methods. Um another thing is uh we've been using multihiret over and over and over because it's the again it's the most realistic realistic data set for this setting. Um but there are some faults with it other than the uh ambiguous questions. there are some poorly formed questions um and maybe even a few incorrect answers. Um, additionally, this data set was created by giving uh financial experts a document and asking them to write questions uh with a single numeric or a span extraction answer based on those documents. So, they're not necessarily the kinds of questions that are actually going to pop up in when in like a situation where questions are being naturally elicited. Um, so something that we would like to do is hunt down some naturally elicited uh questions that go with documents. And we're not totally sure where to find that right now. Um but I think oh I forget what they're called but a lot of uh public companies have um like quarterly calls where they basically it's they kind of give a summary of like these financial documents over an actual call and then usually they'll take uh questions from whoever is in the call and there are publicly available transcripts of these calls for a ton of companies. So that's one potential source. Um uh we are trying to wrap this uh project up fairly soon. Uh so there's only so much that we can do but um there's a lot of other stuff that we've thought about. So for example, how does user expertise affect how ambiguity is perceived? Um again, we're not financial experts. Um so some questions look very ambiguous to us. uh they might not to a financial expert. They might look ambiguous to uh just whatever LLM, but they might not look ambiguous to a language model fine-tuned on a financial data set. Uh so there also is that factor to consider. Um additionally, we've only really been considering uh the task of ambiguity detection. So is this question ambiguous? Yes or no. But to actually kind of resolve that um you need to like one of the uh one interesting problem is uh actually like classifying the ambiguity and then generating follow-up questions uh for if you're in a conversational setting. Um and then this leads into kind of uh user satisfaction. Um, do users want multiple answers? Do they just want the one precise answer? Um, do they prefer uh many follow-up questions to get the exact precision to answer the question with certainty? Um, I don't know. It's kind of an an interesting uh problem. Um, and yeah, did you have any comments on this? >> No. >> All right. Uh, yeah. So that's that's basically the synthesis of what we've been doing in the the last semester. Um hopefully uh we're trying to write a paper uh do some more experiments for that. So uh so yeah uh thank you for listening. >> Yes, good talk. >> Um how much do you think this might generalize beyond just financial uh ambiguity? like what if it's over spreadsheets of uh tables like scientific experiments that have a lot of numbers that are not like financial reports. >> Uh yeah. So, um, a lot of the ambiguity types that we've been encountering are do kind of, um, like pertain to the same kind of mathematical and like data or like information extraction issues that you would find in in like scientific reports, tables or whatever. So, I think it could easily extend to that. I almost um again we haven't like done a quantitative analysis with our classification of ambiguity types but um just from what we've been working on I feel like most of the ambiguities in this data set in particular have more to do with the document and hybrid tabular textual element than they actually do with uh the financial domain element. I'll add a little bit of most of these we've seen are just kind of general ambiguity like taxonomies to begin with and just data sets for um that we've seen a few more recent data sets where it specifically is on like jargon and things like that in the financial domain but most of these you don't really get something that says it specifically is tough because it's looking for like financial rate of change and things like that. So it is something where >> it would be nice almost to have that more specifically. Um, we haven't looked too heavily, I guess, kind of to that point of like things like medical. There's probably a decent bit towards that of medical terminology would probably be the closest like >> adaptable area, but we haven't dived too much into that. >> Yeah, like some of these papers have tried applying their methodologies to like domain specific data sets with like not any particular success. Um, and I do like the idea of keeping the taxonomy like as generalizable as possible. Our goal here is just to figure out like does introducing a specific domain, does introducing a document with text and tables um, introduce any like the possibility of ambiguities that are outside of any of the taxonomies that already exist. >> Yeah. >> And another one you keep saying taxonomy, any ontologies are really just That's not >> a oneast we know of. Yeah. >> I was actually gonna ask you know what's the manifestation? Are you building like a graph of the annotation? What how are you documenting this >> uh big fat spreadsheet? How are how are you actually >> spread? >> Yeah, it's a spreadsheet. >> It's like a Yeah, we haven't >> That's depressing. I think it would be we gave like the high level I guess a couple slides back of just general observations but really to figure out like certain ones we've caught and how much we'd categorize it we'll probably have to just do back and forth to get like true numbers on that. >> Yeah. >> Yeah. I have just a clarification question. So you were working with the financial data set and then >> you noticed the questions were sometimes ambiguous and then you kind of drift from that direction or did you start refing beauty at the beginning? I >> think so. >> So yeah I don't really remember exact like I'm sure we were pushed towards that because we have someone in our group who is very interested in this issue of ambiguity. Um but I think >> I think it was mostly like figuring out what the gap is for financial like data specifically for these kind of tabular data sets and then I think we just looking at the samples we're like a lot of these make no sense and then started from there which yeah happens kind of naturally in data sets anyway but >> that was the question that that earth was posed once how come the LMS are so bad at at those financial questions. >> Okay. >> And so so you launched that and and I found the ambiguities is the one big cause of that, right? >> And and it has many different manifestations. >> Yeah. >> It's came towards the end. So maybe I'm just being completely ignorant when asking come. which so hold me if it's the case. Um you talk about ambiguity. Um with regards how um how would you tackle the question of vagueness? I mean do you do you imply vagueness? Because sometimes LM they uh so ambiguity is as I understand um one thing meaning two uh to concepts right whereas vagueness has insufficient definitions. Ambiguity is two competing ones. So um is how would you like how would you tackling ambiguity distinguish from tackling vagueness? uh and how do LMS perform in terms of over um over hallucinating some some concept that is maybe vaguely explained but rather they um they add some things to it in order to make it sound like everything is not ambiguous but unvague or >> I think make sense >> yeah um I think that's a good question honestly yeah we are particularly when you're or talking about multiple taxonomies, you end up with the problem where you're saying something and everyone's defined it like a bunch of different ways. I think this one is the closest towards what you were asking of detangling those two where you have the linguistic ambiguity where you have the lexical where it's multiple meanings overlapping versus the semantic where it's like some aspect of the the sentence itself like missing parts of that. So they'll say ambiguity but that means I guess like 100 things under the sun specifically but that is something of trying to catch and like why is the sentence performing poorly and like you said the hallucinations too it's like >> did the model understand and then it still just wanted to come up with something or it didn't have the the particular information it needed in its you know um knowledge base or is there something else going on there. So I think that's one of the biggest things that's particularly interesting of can it tell why and what part of it and and address those actually but >> I I I'm just interested in one couple of things. One one is uh to to what degree some of the ambiguity or maybe disconnect right between the question and the the the tables or data comes from not being financial expert. that is which is that whoever asked the question right may ask it using language that they know exactly how to interpret that we're looking at the table but LLM obviously takes it literally what is being asked >> and therefore makes all kinds of weird mistakes so there could be years and profits and something and somebody would say you know some kind of whatever in this jargon of financial jargon Mhm. >> Right. What is this or that? And and any financial person would immediately know what to do looking at these tables even though to me they mean nothing or just columns of I don't know profits and losses and something else. So it's how much of that is a problem. Right. And and presumably that could be shortcut by training this on that data. Right. Right. Um and and it's possible that there isn't enough of that in the training of the LLM, so they don't quite get it. It's like a different lingo. >> Yeah. Yeah, >> it's not the same as medical because medical has kind of very specific terms that but yeah the the terms are I don't know if they they are so um strict they they're a little more kind of offhand you know people use them in their they coming from this school they use this term coming from that school they this term but they all understand each other but we don't understand them right >> I don't know that they all understand each other within like a certain click they probably all understand each other. >> Yeah. I actually I was working with now this is in medicine but I was working with somebody at Mount Si and she had come from Yale and the vernacular at Mount Si despite the fact that you know they're all trained MDs the vernacular in her um emergency room was very different than at Yale. >> Okay. So that could be one other thing that you mentioned. I found it very interesting that you mentioned the um having those um financial reporting to I don't know through um stakeholders or whatever meetings like >> and then the people asking questions which is cool because that there's people who are there they may have completely different things they want to know >> how long you you you will last you know. >> Yeah. Um so this could be good question. Now the the answers are given presumably the question is whether they should be considered ground truth because then the company may not give you the straight answer. They may try to give you a little misleading answers. I think >> everything is great don't worry. But actually, I think that's a really nice data set because there's strict reporting requirements for what they they start with. Maybe not in how they answer the questions, but for what they start with on those quarterly or whatever it is, >> there's kind of templates that they're following, >> right? But then then presumably that it could be that the numbers in tables that I have to report I mean assuming they're accurate and not cheating or anything but then how you interpret those numbers right >> that that is matter of of interpretation right and and this the stakeholders or or shareholders they may they they probably don't care you know about details they want to know how much profit I'm going to break in in next year >> and they want to know if they should sell or or hold. >> Exactly. And the company say, "Oh, yeah. Yeah, you can you hold on to it. It's going to go up." Right. >> Whatever. >> Yeah. >> I was looking at some of these some of these examples and uh zero shot. These are good examples, but in the context of a of a like a chat, >> they aren't >> right. >> Yeah. Because like who's Neil Armstrong? Okay. When did he land on the moon? That makes perfect sense. >> Yeah. >> Yeah. I was How much of that was like really context or just the previous question >> is missing in that case for example? >> Yeah. I think this I think this um Oh, shoot. I can't remember if the clamber benchmark because some of them um were like there was one that was designed with conversational question answering in mind. Do you remember which one it is? Because I don't right now. >> Yeah. Not off the top of my head. >> Yeah, cuz Yeah, that is an important distinction for sure. >> I guess what what uh where what formed that question? I was I started thinking a few minutes ago about should should I imagine a chatbot that sort of evaluated your question and I'm an ambiguity you know like I'm not going to accept that question because you have element to it right >> I I can't answer that question >> yeah you would you would want that happening in the in background and then you would want them to generate a nice >> when did he land? >> Who are you talking about? Which moon? >> Yeah. >> Um but you see what I mean? You know, there's I'm wondering if if there's a de disambiguator. >> Yeah. >> Type of thing. >> Yeah. in back in linguistic terms would be you actually would it be also considered a vagueness rather than ambiguity well that that's in their in their categorization somewhere I guess I feel like >> oh when did you >> I feel like you addressed >> semantic ambiguity >> like I mean we came to this because Sasha asked about vagueness >> yeah can you like reiterate what your definition of vagueness is because we haven't like run into that in this research as like a precise Right. So I guess it's these are all linguistic terms and I think March has like made a great distinction between them. So vagueness is when it's too fuzzy. So for example animal well is it a fox? Is it a dog? Is it a cat? Right. But ambiguity is let's say um school. Is it No, actually this is also a bad idea. Um being too big. >> Exactly. >> Exactly. It's uh it's when a word has multiple meanings which are not a subcategory. I'm blanking right now but I think but it has two meanings. Uh one one is one meaning and but the other one is the other meaning >> cat or car. >> Yeah. Yeah. I think um I I would say that for most of these taxonomies the idea of vagueness does fall under the ambiguity umbrella. So like here um I mean you were mentioning like this lexical ambiguity like um the question tell me about the source of the Nile um that in uh I think in a in like the appendix um they explain that like the two possible interpretations of that could have been like the river Nile or like some book called the Nile >> there's ambiguity. >> Uhhuh. Because uh which because because one one is one concept the other one is the other right they are disjunct where they talk about animal and I say okay he's talking about animal but does he mean a dog a fox or a cat specificity when the concept is too vague and insufficient to make a call between one of the subcategories that is part of it. Yeah. I mean >> this question is also vague because you don't know what tell me about the source of of Nile, right? >> Yes. >> What do you want to know? >> Where does the water >> What do you want to know? >> Yes. Yes. >> What do you want to know? That's vague. >> Yes. >> What do you want to know is this? How much water comes out of it? You know, whatever. >> Yes. Exactly. But if you um so this is we've been focusing on questions but if you turn it around what would a vague answer be what would an ambiguous answer just like you through to through the um effects right tell me about it will tell you whatever it knows but it will make some kind of selection right probably doesn't tell you everything right that's right you know pick something and say hey this is what I know and then if you're not happy then say hey tell me more or tell me about this or that right >> there's another thing what if it gives a well a vague answer but yeah if it's too general would that be vague >> yes maybe >> nebulous where does nebulous nebula >> well actually vague if if you don't know how to answer that yes >> but if you simply answer this somehow then I mean it's not vague anymore right >> well it could be then ambiguous do you mind me searching your car? No. Oh, yes. Right. What What does the answer mean? No, I do mind you searching your car. No, I don't mind you searching my car. >> This is an ambiguous >> talking about my car. >> Take one. hopefully yours. I think big aspect to the conversation of the answer being ambiguous really goes into the user themselves which we were talking about of like if the user is an expert or if they're a novice in terms of are they able to properly state what information that they actually want to get out of the question or out of the answer because if they're just leaving it which I feel like I've done plenty of times before like I'm not really sure like exactly what information I want to learn about something and so I might just be looking around for a while on Wikipedia or something until I some particular, you know, aspect that I thought was interesting. Um, but I think that's a big >> part of it where they're not properly defining the thing that they need solved in that moment. um in terms of what piece of information they're trying to get which I was I was going to comment on for the meeting transcripts I think would be interesting just because those are questions that are being answered like in the moment which I think is very different than someone curating a data set and curating those questions it's like someone join you know you know putting their microphone on and asking something real quick and then I I have no idea how that would happen that the people in the back are like scrambling to find the pieces of information that'll >> not make their socks like fall apart. But um yeah, I think that that's particularly interesting there like the interaction between the person asking the question and how well the the model can resolve that. So >> but yeah, we're looking at a lot of single attempts at the moment. >> So what is the army's paid for this? What what are they going to do with this? >> It's just okay. What are they going to do with this? >> Air Force is generally interested in how far you can push LLM to do stuff. >> I don't know where that came from. There was there was someone from within the air force the research labs this this came out. Somebody had a problem like this and hey what what what can we do about this? Right? So this is the one kind of thing you know the other thing that that that landed on on us is uh you know how can we how can we can push LLMs to do wargaming >> which is something that we we're doing right now. It's another another part of a project, right? And so well something like yeah there are two forces and they they incomplete information and something is damaged and you have to go around that you know whatever and basically they want to know how far you can push those agentic whatever things to be able to solve this kind of problems. Okay. And they were particularly interested in damage assessment. Meaning when you your ship is is damaged and it's not sunk yet, you know, it's it's still a float but but can't quite do everything that it could do before. So they want to somehow assess that based on whatever observation you can make. some have to send a drone, you know, whatever. Maybe the ship can be rep repurposed for some other use. Like, okay, you can't move. So, the only thing I can do with you can use your radar and and we're going to use you as a, you know, to see who is around because there otherwise there's nothing you can do because you can't move, right? That kind of stuff. And then then take this in the totality of some assets that you have and and there's some kind of adversary. And so you have to something is damaged so you have to replan everything and then some of these assets you have can be used for something else now because they they no longer function in that kind of kind of thing. And so so we actually building for them we said we're not going to go into classified world. They they have they have a um a very the um realistic model of the Middle East. It was a company that that did that for them. Um but it's static. It it covers like three or four weeks of of activities in the Gulf of Aen detail to the single vehicles moving around whatever. But the problem is that it's all static. And so if you want to play a scenario where things are changing, you can't because it's it goes on one particular line. So and building that again to be kind of it's it's too much. Even building that static word was was we didn't do that. Somebody else did. So, so we said that we can build a a little game, but it's going to be um really very simp kind of a schematic where you can basically play those swarms against one another. And so, so that's what we're doing. And then now they they give us a list of tasks. I say, "Okay, so it's great. This is going to be public domain. We can put it out there." Uh we we can even get people play that and then here's you know we want you to check that scenario this scenario kind of situation whatever one one thing they're interested in is uh whether the LLMs can LLM specifically right they can learn from their own mistakes right so they you go and they try to execute some plan and you fail and then if you go into the same situation again, you know, would you do the same mistake again or not? Like can you can you learn from that and somehow avoid doing this? >> Wasn't one of the papers we talked about at the 3:00 meeting about that? memory the memory. Yeah, >> I think I represented something like from using episodic memory to which I can share it with everyone >> and actually I mean the general game playing literature is it's old like chzareth is kind of a player in it um for learning you know basically playing games and then learning from your game playing learning strategy so it's it predates LLMs by decades actually but there's literature are there on >> Oh, no. Yeah. So, we we using some of that at least. I mean, there's uh Joe is working on that and there's another another student at UNC. Um and we we try to get them also when they wrap up this to help a little bit with with some of the aspects of that. Um >> I like this potentially dynamic asset reallocation angle, you know. >> Yeah, I like that too. partially damaged >> suggested to them and they grab that that it's um the the what the difference from games like game theories and game thing is that information is incomplete. You have things like you you only are aware of certain part of the environment and and right visibility and it's usually fragmented because you may have some you know may have some drone and we have some radar and and then you you see this piece and that piece and you don't see what is in between right so you have to have make some assumptions about your adversary also there are some other things like merchant ships there's weather moving you know and and you can predict which way this goes. You can get weather forecast but you cannot always trust it, right? said some kind of stuff like that and and somehow you have to function in this in in the in this kind of fog because what a fog of war is called right and you have to function in that right and and you have limited assets right there's only so much that you have right if something gets destroyed then you have fewer you know and and so forth so it's it's the economic aspect you know that kind of thing come into place and and then we're doing this with LLMs because because it's cheap to do this with LLMs. you don't have to do this the visualizations and and and vision and all that stuff which they don't want to invest in then until they know what the kind of capabilities are then maybe then you can invest in some additional >> well one of the reasons I asked you what they were doing with this is because I think the idea of having um an evaluation of the question and how to improve the question might be something that you could still use the LLM if you could say well this question has you know some problems would you like to refine it say this way you know would you like to refine it or here I can give you five different suggestions on how to refine it because you know that's historically what a lot of the search query engines used to do like varity ancient search engine you know used to evaluate whether I was going to get way too many answers or way too few answers and so you know you either make your question more general or more specific so they were doing kind of a query suggestion mechanism Yeah. Yeah. We we we did that kind of things in the past. Yes. It was there was >> but like it might be a next step for this if they wanted to use something like this. If you could just do a one round of refinement on the query make it less of a certain kind of ambiguity and maybe you could get a good answer. I don't know you know this actually it's funny because this came from air force and then then they ended up awarding this through khaki subcontract >> and these people are cocky I have they have absolutely no idea what why we're doing this because there's a total disconnect they just give us money and and they don't care what we're doing >> people >> yeah yeah khaki bought well khaki bought LGS right >> and then basically most of the good people who didn't have kids in my school left. Uh but yeah, and redirected. Khaki's buying everybody and kind of ruining everybody they >> Yeah. Well, what this we have this whole my program manager and this whole team left at Air Force. >> Yeah, >> they're gone. He he ended up working for for Nvidia. >> Oh. Oh, they have options. >> So, >> oh, are we still reporting? We should stop. I thought let's let's >> Oh, it's five o'clock. Better go, right? >> Good job. >> Thank you. >> Thank you guys. This is will be for being here >> helpful for um putting together some kind of a report up to them. Are