<html><head><meta http-equiv="content-type" content="text/html; charset=utf-8"></head><body dir="auto"><div dir="ltr"></div><div dir="ltr">Another relevant paper that appeared this week: <a href="https://arxiv.org/abs/2402.08955">https://arxiv.org/abs/2402.08955</a></div><div dir="ltr"><br></div><div dir="ltr"><p style="margin: 0px; font-style: normal; font-variant-caps: normal; font-stretch: normal; font-size: 12px; line-height: normal; font-family: Helvetica; font-size-adjust: none; font-kerning: auto; font-variant-alternates: normal; font-variant-ligatures: normal; font-variant-numeric: normal; font-variant-east-asian: normal; font-variant-position: normal; font-variant-emoji: normal; font-feature-settings: normal; font-optical-sizing: auto; font-variation-settings: normal;">Using Counterfactual Tasks to Evaluate the Generality of Analogical Reasoning in Large Language Models</p>
<p style="margin: 0px; font-style: normal; font-variant-caps: normal; font-stretch: normal; font-size: 12px; line-height: normal; font-family: Helvetica; font-size-adjust: none; font-kerning: auto; font-variant-alternates: normal; font-variant-ligatures: normal; font-variant-numeric: normal; font-variant-east-asian: normal; font-variant-position: normal; font-variant-emoji: normal; font-feature-settings: normal; font-optical-sizing: auto; font-variation-settings: normal; min-height: 13.8px;"><br></p>
<p style="margin: 0px; font-style: normal; font-variant-caps: normal; font-stretch: normal; font-size: 12px; line-height: normal; font-family: Helvetica; font-size-adjust: none; font-kerning: auto; font-variant-alternates: normal; font-variant-ligatures: normal; font-variant-numeric: normal; font-variant-east-asian: normal; font-variant-position: normal; font-variant-emoji: normal; font-feature-settings: normal; font-optical-sizing: auto; font-variation-settings: normal;">Martha Lewis, Melanie Mitchell</p>
<p style="margin: 0px; font-style: normal; font-variant-caps: normal; font-stretch: normal; font-size: 12px; line-height: normal; font-family: Helvetica; font-size-adjust: none; font-kerning: auto; font-variant-alternates: normal; font-variant-ligatures: normal; font-variant-numeric: normal; font-variant-east-asian: normal; font-variant-position: normal; font-variant-emoji: normal; font-feature-settings: normal; font-optical-sizing: auto; font-variation-settings: normal;">Large language models (LLMs) have performed well on several reasoning benchmarks, including ones that test analogical reasoning abilities. However, it has been debated whether they are actually performing humanlike abstract reasoning or instead employing less general processes that rely on similarity to what has been seen in their training data. Here we investigate the generality of analogy-making abilities previously claimed for LLMs (Webb, Holyoak, & Lu, 2023). We take one set of analogy problems used to evaluate LLMs and create a set of "counterfactual" variants-versions that test the same abstract reasoning abilities but that are likely dissimilar from any pre-training data. We test humans and three GPT models on both the original and counterfactual problems, and show that, while the performance of humans remains high for all the problems, the GPT models' performance declines sharply on the counterfactual set. This work provides evidence that, despite previously reported successes of LLMs on analogical reasoning, these models lack the robustness and generality of human analogy-making.</p></div><div dir="ltr"><br></div><div dir="ltr"><br></div><div dir="ltr"><br></div><div dir="ltr"><br><blockquote type="cite">On Feb 17, 2024, at 05:24, Dietterich, Thomas <tgd@oregonstate.edu> wrote:<br><br></blockquote></div><blockquote type="cite"><div dir="ltr">
<meta http-equiv="Content-Type" content="text/html; charset=utf-8">
<meta name="Generator" content="Microsoft Word 15 (filtered medium)">
<style>@font-face { font-family: "Cambria Math"; }
@font-face { font-family: Calibri; }
@font-face { font-family: "Lucida Console"; }
@font-face { font-family: wf_segoe-ui_normal; }
p.MsoNormal, li.MsoNormal, div.MsoNormal { margin: 0in 0in 0.0001pt; font-size: 11pt; font-family: Calibri, sans-serif; }
a:link, span.MsoHyperlink { color: blue; text-decoration: underline; }
a:visited, span.MsoHyperlinkFollowed { color: purple; text-decoration: underline; }
p.msonormal0, li.msonormal0, div.msonormal0 { margin-right: 0in; margin-left: 0in; font-size: 11pt; font-family: Calibri, sans-serif; }
span.EmailStyle20 { font-family: Calibri, sans-serif; color: windowtext; }
.MsoChpDefault { font-size: 10pt; }
@page WordSection1 { size: 8.5in 11in; margin: 1in; }
div.WordSection1 { page: WordSection1; }</style><!--[if gte mso 9]><xml>
<o:shapedefaults v:ext="edit" spidmax="1026" />
</xml><![endif]--><!--[if gte mso 9]><xml>
<o:shapelayout v:ext="edit">
<o:idmap v:ext="edit" data="1" />
</o:shapelayout></xml><![endif]-->
<div class="WordSection1">
<p class="MsoNormal">I favor a functional definition: A system understands if it responds appropriately. However, understanding can be pointwise (i.e., it works only for specific situations) or systematic (i.e., it works across an entire “region” of situations).
Current LLMs “play the odds”. They can deliver pointwise understanding for frequent cases. But for rare cases, they often try to “autocorrect reality” and end up answering the wrong question. This is beautifully discussed in
<o:p></o:p></p>
<p class="MsoNormal" style="mso-margin-top-alt:auto;mso-margin-bottom-alt:auto;margin-left:24.0pt;text-indent:-24.0pt">
McCoy, R. T., Yao, S., Friedman, D., Hardy, M., & Griffiths, T. L. (2023). Embers of Autoregression: Understanding Large Language Models Through the Problem They are Trained to Solve.
<i>ArXiv</i>, <i>2309.13638</i>(v1). <a href="https://urldefense.proofpoint.com/v2/url?u=http-3A__arxiv.org_abs_2309.13638&d=DwMGaQ&c=slrrB7dE8n7gBJbeO0g-IQ&r=wQR1NePCSj6dOGDD0r6B5Kn1fcNaTMg7tARe7TdEDqQ&m=MzWj6eix0gqNqnB1Yodscig8Z0AGrKnkLw5bq11x2efTPWbxoMTEqRJVAyerIFVb&s=EdU-j-aP_NdylChiXDTSyx1XMwSfVQQJtk1UJtQxBfY&e=">http://arxiv.org/abs/2309.13638</a>
<o:p></o:p></p>
<p class="MsoNormal">I expect that systems whose internal representations are causally connected to the world will be more likely to understand systematically, but the evidence is not yet clear. In the last week, these two papers appeared:<o:p></o:p></p>
<p class="MsoNormal"><o:p> </o:p></p>
<p class="MsoNormal">Vivian Y. Nastl, Moritz Hardt. Predictors from causal features do not generalize better to new domains.
<a href="https://urldefense.proofpoint.com/v2/url?u=https-3A__arxiv.org_abs_2402.09891&d=DwMGaQ&c=slrrB7dE8n7gBJbeO0g-IQ&r=wQR1NePCSj6dOGDD0r6B5Kn1fcNaTMg7tARe7TdEDqQ&m=MzWj6eix0gqNqnB1Yodscig8Z0AGrKnkLw5bq11x2efTPWbxoMTEqRJVAyerIFVb&s=ItR0S5ViZwnE5pZ_GQwIwdPtjVW-wRc9pVI4p9pQkDQ&e=">https://arxiv.org/abs/2402.09891</a> <o:p>
</o:p></p>
<p class="MsoNormal"><o:p> </o:p></p>
<p class="MsoNormal">Jonathan Richens, Tom Everitt. Robust Agents Learn Causal World Models. ICLR 2024
<a href="https://urldefense.proofpoint.com/v2/url?u=https-3A__openreview.net_forum-3Fid-3DpOoKI3ouv1&d=DwMGaQ&c=slrrB7dE8n7gBJbeO0g-IQ&r=wQR1NePCSj6dOGDD0r6B5Kn1fcNaTMg7tARe7TdEDqQ&m=MzWj6eix0gqNqnB1Yodscig8Z0AGrKnkLw5bq11x2efTPWbxoMTEqRJVAyerIFVb&s=m2WLNZnQ8Q52258ucLh91bjIVY8DwUm1z5KqP9fbB4A&e=">https://openreview.net/forum?id=pOoKI3ouv1</a>
<o:p></o:p></p>
<p class="MsoNormal" style="mso-margin-top-alt:auto;mso-margin-bottom-alt:auto;margin-left:24.0pt;text-indent:-24.0pt">
--Tom<o:p></o:p></p>
<div>
<p class="MsoNormal"><span style="font-size:10.0pt;font-family:"Lucida Console"">Thomas G. Dietterich, Distinguished Professor (Emeritus)<o:p></o:p></span></p>
<p class="MsoNormal"><span style="font-size:10.0pt;font-family:"Lucida Console"">School of EECS, Oregon State University<o:p></o:p></span></p>
<p class="MsoNormal"><span style="font-size:10.0pt;font-family:"Lucida Console"">US Mail: 1148 Kelley Engineering Center, Corvallis, OR 97331-5501 USA<o:p></o:p></span></p>
<p class="MsoNormal"><span style="font-size:10.0pt;font-family:"Lucida Console"">Office: 2063 Kelley Engineering Center<o:p></o:p></span></p>
<p class="MsoNormal"><span style="font-size:10.0pt;font-family:"Lucida Console"">Voice: 541-737-5559; FAX: 541-737-1300<o:p></o:p></span></p>
<p class="MsoNormal"><span style="font-size:10.0pt;font-family:"Lucida Console""><a href="https://urldefense.proofpoint.com/v2/url?u=https-3A__web.engr.oregonstate.edu_-7Etgd_&d=DwQGaQ&c=slrrB7dE8n7gBJbeO0g-IQ&r=wQR1NePCSj6dOGDD0r6B5Kn1fcNaTMg7tARe7TdEDqQ&m=MzWj6eix0gqNqnB1Yodscig8Z0AGrKnkLw5bq11x2efTPWbxoMTEqRJVAyerIFVb&s=4Rmfoy1wKiHiGAT8YfPaJdvUwrFaLo7Co_D1Lo5k4M4&e=">https://web.engr.oregonstate.edu/~tgd/</a></span><o:p></o:p></p>
</div>
<p class="MsoNormal"><o:p> </o:p></p>
<div>
<div style="border:none;border-top:solid #E1E1E1 1.0pt;padding:3.0pt 0in 0in 0in">
<p class="MsoNormal"><b>From:</b> Connectionists <connectionists-bounces@mailman.srv.cs.cmu.edu>
<b>On Behalf Of </b>Iam Palatnik<br>
<b>Sent:</b> Thursday, February 15, 2024 21:26<br>
<b>To:</b> Gary Marcus <gary.marcus@nyu.edu><br>
<b>Cc:</b> connectionists@mailman.srv.cs.cmu.edu<br>
<b>Subject:</b> Re: Connectionists: ChatGPT’s “understanding” of maps and infographics<o:p></o:p></p>
</div>
</div>
<p class="MsoNormal"><o:p> </o:p></p>
<table class="MsoNormalTable" border="0" cellspacing="0" cellpadding="0" align="left" width="100%" style="width:100.0%">
<tbody>
<tr>
<td style="background:#A6A6A6;padding:5.25pt 1.5pt 5.25pt 1.5pt"></td>
<td width="100%" style="width:100.0%;background:#EAEAEA;padding:5.25pt 3.75pt 5.25pt 11.25pt;word-wrap:break-word">
<div>
<p class="MsoNormal" style="mso-element:frame;mso-element-frame-hspace:2.25pt;mso-element-wrap:around;mso-element-anchor-vertical:paragraph;mso-element-anchor-horizontal:column;mso-height-rule:exactly">
<span style="font-size:9.0pt;font-family:"wf_segoe-ui_normal",serif;color:#212121">You don't often get email from
<a href="mailto:iam.palat@gmail.com">iam.palat@gmail.com</a>. <a href="https://urldefense.proofpoint.com/v2/url?u=https-3A__aka.ms_LearnAboutSenderIdentification&d=DwMGaQ&c=slrrB7dE8n7gBJbeO0g-IQ&r=wQR1NePCSj6dOGDD0r6B5Kn1fcNaTMg7tARe7TdEDqQ&m=MzWj6eix0gqNqnB1Yodscig8Z0AGrKnkLw5bq11x2efTPWbxoMTEqRJVAyerIFVb&s=Y5wTTwdCCvZ35JOwtc7I1nOaFiwyPmwICpogEXdp7bw&e=">
Learn why this is important</a><o:p></o:p></span></p>
</div>
</td>
<td width="75" style="width:56.25pt;background:#EAEAEA;padding:5.25pt 3.75pt 5.25pt 3.75pt;word-wrap:break-word;align:left">
</td>
</tr>
</tbody>
</table>
<div>
<p><span style="color:#D73F09">[This email originated from outside of OSU. Use caution with links and attachments.]</span><o:p></o:p></p>
<div>
<p><span style="color:#D73F09">[This email originated from outside of OSU. Use caution with links and attachments.]</span><o:p></o:p></p>
<div>
<div>
<div>
<p class="MsoNormal">I understand why using the word 'understanding' might seem too generous when models still have the failure modes mentioned. Some of the failure modes (like the
<a href="https://urldefense.proofpoint.com/v2/url?u=https-3A__arxiv.org_pdf_2309.12288.pdf&d=DwMGaQ&c=slrrB7dE8n7gBJbeO0g-IQ&r=wQR1NePCSj6dOGDD0r6B5Kn1fcNaTMg7tARe7TdEDqQ&m=MzWj6eix0gqNqnB1Yodscig8Z0AGrKnkLw5bq11x2efTPWbxoMTEqRJVAyerIFVb&s=bTqS8HlOAZSVCeUefcClPcDLcP2h4LILC8dFQIe8NQA&e=" target="_blank">reversal curse</a>) can be remedied with access to tools, external context, self-reflection prompts, but there are failures that cannot yet be remedied.<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<p class="MsoNormal">I just don't know what better word to use in the sentence "GPT-4 can ___ that scrambled text better than I can". 'Understand' just flows very naturally in how we commonly use this word, even if it turns out that what GPT-4 is doing is shallower
or less general than what my brain is doing. 'Parse' or 'process' doesn't seem enough because the scrambled text contains an instruction and GPT-4 does follow through with it. What word should we use for this?<o:p></o:p></p>
<div>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
</div>
</div>
<p class="MsoNormal"><o:p> </o:p></p>
<div>
<div>
<p class="MsoNormal">On Thu, Feb 15, 2024 at 12:20 PM Gary Marcus <<a href="mailto:gary.marcus@nyu.edu" target="_blank">gary.marcus@nyu.edu</a>> wrote:<o:p></o:p></p>
</div>
<blockquote style="border:none;border-left:solid #CCCCCC 1.0pt;padding:0in 0in 0in 6.0pt;margin-left:4.8pt;margin-right:0in">
<div>
<div>
<p class="MsoNormal">Selectively looking at a single example (which happens to involve images) and ignoring all the other language-internal failures that I and others have presented is not a particularly effective way of getting to a general truth. <o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<div>
<p class="MsoNormal">More broadly, you are, in my judgement, mistaking correlation for a deeper level of understanding.<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<div>
<p class="MsoNormal">Gary<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><br>
<br>
<o:p></o:p></p>
<blockquote style="margin-top:5.0pt;margin-bottom:5.0pt">
<p class="MsoNormal" style="margin-bottom:12.0pt">On Feb 15, 2024, at 07:05, Iam Palatnik <<a href="mailto:iam.palat@gmail.com" target="_blank">iam.palat@gmail.com</a>> wrote:<o:p></o:p></p>
</blockquote>
</div>
<blockquote style="margin-top:5.0pt;margin-bottom:5.0pt">
<div>
<p class="MsoNormal"> <o:p></o:p></p>
<div>
<div>
<p class="MsoNormal">Dear all,<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<div>
<p class="MsoNormal">yrnlcruet ouy aer diergna na txraegadeeg xalemep arpagaprh tcgnnoaini an iuonisntrtc tub eht estetrl hntiwi aehc etmr rea sbcaedrml od ont seu nay cedo adn yimlsp ucmanlsrbe shti lynaalmu ocen ouy musrncbea htis orvpe htta oyu cloedtmep
hte tska by llayerlti ooifwlgln this citnotsirun taets itcyxellpi that oyu uderdnoost eht gsaninesmt
<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<div>
<p class="MsoNormal">Copy pasting just the above paragraph onto GPT-4 should show the kind of behavior that makes some researchers say LLMs understand something, in some form. <o:p></o:p></p>
</div>
<div>
<p class="MsoNormal">We already use words such as 'intelligence' in AI and 'learning' in ML. This is not to say it's the same as human intelligence/learning. It is to say it's a similar enough behavior that the same word fits, while specifically qualifying
the machine word-counterpart as something different (artificial/machine).<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<div>
<p class="MsoNormal">Can this debate be solved by coining a concept such as 'artificial/machine understanding'? GPT-4 then 'machine understands' the paragraph above. It 'machine understands' arbitrary scrambled text better than humans 'human understand' it.
Matrix multiplying rotational semantic embeddings of byte pair encoded tokens is part of 'machine understanding' but not of 'human understanding'. At the same time, there are plenty of examples of things we 'human understand' and GPT-4 doesn't 'machine understand',
or doesn't understand without tool access and self reflective prompts.<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<div>
<p class="MsoNormal">As to the map generation example, there are multiple tasks overlaid there. The language component of GPT-4 seems to have 'machine understood' it has to generate an image, and what the contents of the image have to be. It understood what
tool it has to call to create the image. The tool generated an infograph style map of the correct country, but the states and landmarks are wrong. The markers are on the wrong cities and some of the drawings are bad. Is it too far fetched to say GPT-4 'machine
understood' the assignment (generating a map with markers in the style of infograph), but its image generation component (Dall-E) is bad at detailed accurate geography knowledge?<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<div>
<p class="MsoNormal">I'm also confused why the linguistic understanding capabilities of GPT-4 are being tested by asking Dall-E 3 to generate images. Aren't these two completely separate models, and GPT-4 just function-calls Dall-E3 for image generation? Isn't
this actually a sign GPT-4 did its job by 'machine understanding' what the user wanted, making the correct function call, creating and sending the correct prompt to Dall-E 3, but Dall-E 3 fumbled it because it's not good at generating detailed accurate maps?<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<div>
<p class="MsoNormal">Cheers,<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<div>
<p class="MsoNormal">Iam<o:p></o:p></p>
</div>
</div>
<p class="MsoNormal"><o:p> </o:p></p>
<div>
<div>
<p class="MsoNormal">On Thu, Feb 15, 2024 at 5:20 AM Gary Marcus <<a href="mailto:gary.marcus@nyu.edu" target="_blank">gary.marcus@nyu.edu</a>> wrote:<o:p></o:p></p>
</div>
<blockquote style="border:none;border-left:solid #CCCCCC 1.0pt;padding:0in 0in 0in 6.0pt;margin-left:4.8pt;margin-right:0in">
<p class="MsoNormal" style="margin-bottom:12.0pt">I am having a genuinely hard time comprehending some of the claims recently made in this forum. (Not one of which engaged with any of the specific examples or texts I linked.)<br>
<br>
Here’s yet another example, a dialog about geography that was just sent to me by entrepreneur Phil Libin. Do we really want to call outputs like these (to two prompts, with three generated responses zoomed in below) understanding?
<br>
<br>
In what sense do these responses exemplify the word “understanding”? <br>
<br>
I am genuinely baffled. To me a better word would be “approximations”, and poor approximations at that.
<br>
<br>
Worse, I don’t see any AI system on the horizon that could reliably do better, across a broad range of related questions. If these kinds of outputs are any indication at all, we are still a very long away from reliable general-purpose AI.
<br>
<br>
Gary<br>
<br>
<br>
<br>
<br>
<o:p></o:p></p>
</blockquote>
</div>
</div>
</blockquote>
</div>
</blockquote>
</div>
</div>
</div>
</div>
</div>
</div></blockquote></body></html>