Building Virtual Environments for Training and Benchmarking Agentive Communicative Systems - Einzelansicht

Funktionen:

Veranstaltungsart	Seminar	Veranstaltungsnummer
SWS	2	Semester	SoSe 2026
Einrichtung	Department Linguistik	Sprache	englisch
Belegungsfrist	01.04.2026 - 10.05.2026

Gruppe 1:

Vormerken: Zur Zeit keine Belegung möglich

		Tag	Zeit	Rhythmus	Dauer	Raum	Lehrperson	Ausfall-/Ausweichtermine	Max. Teilnehmer/-innen
	Seminar	Do	14:00 bis 16:00	wöchentlich	16.04.2026 bis 23.07.2026	2.14.0.09	Prof. Dr. Schlangen

Kommentar	<p>Please note that this course starts in week 2; the first meeting will be on April 23rd.</p><p>Large language models have become what they are through the consumption of extremely large amounts of data. This has recently given rise to a set of different, but related concerns: When it comes to objectively testing the real abilities of these models beyond anecdotal evidence, the use of existing testing resources has become problematic, if it cannot easily be guaranteed that in their extreme hunger for data, these resources have not already been consumed during training -- and any attempts of testing with them would mean committing the cardinal sin of testing on the training data. When it comes to training, data of the kind used so far has to be seen as a finite resource, and there have been claims that this resource has been exhausted, possibly earlier than scaling effects through training on more data. Finally, there is an older claim that the kind of observational data (text as it has been used for whatever purposes the original author made it available for) that so far has been predominantly used is limited in what can be learned from it, and that experiential data that the learning agent is involved in creating is what is needed to unlock further capabilities.</p><div draggable="false" data-pm-slice="1 1 []" data-en-clipboard="true"> </div><div draggable="false">Synthetic interaction environments have been proposed as one possible solution to all three of these problems. In the course, we will look at the clembench environment [1] developed at the University of Potsdam which realises synthetic interactions in goal-oriented, game-like settings. More specifically, we will move beyond its use for benchmarking and towards use in learning, by realising larger scale interaction environments, in which LLM-based agents can simulate (and learn from) social interaction. A particular focus will be on the phenomenon of coordination between agents through the building of common ground. </div><div draggable="false">The course is particularly suitable if you want to gain experience in working on larger-scale software projects, on the cutting edge of development of software agents.</div><div draggable="false"> </div><div draggable="false">[1 ]<a draggable="false" href="https://github.com/clp-research/clembench">https://github.com/clp-research/clembench</a> ; <a draggable="false" href="https://github.com/clp-research/clemcore">https://github.com/clp-research/clemcore</a></div>

Kommentar

<p>Please note that this course starts in week 2; the first meeting will be on April 23rd.</p><p>Large language models have become what they are through the consumption of extremely large amounts of data. This has recently given rise to a set of different, but related concerns: When it comes to objectively testing the real abilities of these models beyond anecdotal evidence, the use of existing testing resources has become problematic, if it cannot easily be guaranteed that in their extreme hunger for data, these resources have not already been consumed during training -- and any attempts of testing with them would mean committing the cardinal sin of testing on the training data. When it comes to training, data of the kind used so far has to be seen as a finite resource, and there have been claims that this resource has been exhausted, possibly earlier than scaling effects through training on more data. Finally, there is an older claim that the kind of observational data (text as it has been used for whatever purposes the original author made it available for) that so far has been predominantly used is limited in what can be learned from it, and that experiential data that the learning agent is involved in creating is what is needed to unlock further capabilities.</p><div draggable="false" data-pm-slice="1 1 []" data-en-clipboard="true"> </div><div draggable="false">Synthetic interaction environments have been proposed as one possible solution to all three of these problems. In the course, we will look at the clembench environment [1] developed at the University of Potsdam which realises synthetic interactions in goal-oriented, game-like settings. More specifically, we will move beyond its use for benchmarking and towards use in learning, by realising larger scale interaction environments, in which LLM-based agents can simulate (and learn from) social interaction. A particular focus will be on the phenomenon of coordination between agents through the building of common ground. </div><div draggable="false">The course is particularly suitable if you want to gain experience in working on larger-scale software projects, on the cutting edge of development of software agents.</div><div draggable="false"> </div><div draggable="false">[1 ]<a draggable="false" href="https://github.com/clp-research/clembench">https://github.com/clp-research/clembench</a> ; <a draggable="false" href="https://github.com/clp-research/clemcore">https://github.com/clp-research/clemcore</a></div>

Strukturbaum

Die Veranstaltung wurde 2 mal im Vorlesungsverzeichnis SoSe 2026 gefunden:

Vorlesungsverzeichnis

Humanwissenschaftliche Fakultät

Department Linguistik

Master of Science

Cognitive Systems: Language, Learning and Reasoning (Prüfungsversion ab WiSe 2014/15)

Project Seminar

PM1 - Project in Computational Linguistics - - - 1

PM2 - Project in Machine Learning - - - 2