BEGIN:VCALENDAR VERSION:2.0 PRODID:https://evenements.uqam.ca X-PUBLISHED-TTL:P1W BEGIN:VEVENT UID:34430@https://evenements.uqam.ca DTSTART:20260917T103000Z SEQUENCE:4 TRANSP:OPAQUE URL:https://evenements.uqam.ca/evenements/seminaire-au-dic-mechanistic-emer gence-of-grounding-par-ziqiao-ma/34430?date=2026-09-17_10-30-00 LOCATION:UQAM - Pavillon Président-Kennedy (PK) (201\, avenue du Présiden t-Kennedy\, Montréal ) SUMMARY:Séminaire au DIC: «Mechanistic Emergence of Grounding» par Ziqia o Ma CLASS:PUBLIC DESCRIPTION:Séminaire ayant lieu dans le cadre du Doctorat en informatique cognitive\, en collaboration avec le centre de recherche CRIA             \n\n\n \n\n\nTITRE : Mechanistic Emergence of Grounding\n\n \n \n\n\nZiqiao MA\n\n\nJeudi le 17 septembre 2026 à 10h30\n\n\nLocal PK -5115 (Il est possible d'y assister en virtuel en vous inscrivant ici)               \n\n\n \n\n\nRÉSUMÉ\n\n\nWhat does it mean fo r a language model to actually ground a word in the world? Much of the cur rent discussion treats grounding as an observed correspondence: a word lik e “horse” aligns with the right image region\, so the model appears gr ounded. But correlation alone leaves a deeper question unanswered: how doe s this connection arise during learning\, and what inside the model actual ly implements it? In this talk\, I will approach grounding as a process ra ther than a property of a finished model. Starting from a minimal setting inspired by child language learning\, we trace how models learn to connect linguistic symbols with corresponding information from the environment. I nterestingly\, models initially rely heavily on simple co-occurrence stati stics\, but later develop mechanisms that go beyond these surface correlat ions. By following information flow across training and intervening on ind ividual attention heads\, we find that grounding becomes concentrated in s pecialized aggregation mechanisms in the model's middle layers. I will the n show how this picture extends from controlled experiments to vision-lang uage models\, and discuss what it suggests about language learning\, multi modal model design\, hallucination\, and the broader symbol grounding deba te.\n\n\n \n\n\nBIOGRAPHIE\n\n\nZiqiao MA is a Member of Technical Staff at Thinking Machines Lab. He obtained his Ph.D. at the University of Michi gan. His research stands at the intersection of language\, interaction\, a nd embodiment from a scalable and cognitive perspective\, with the goal of grounding and aligning language agents to non-linguistic modalities and r ich interactive contexts. He received an Outstanding Paper Award at ACL 20 23\, and an Amazon Alexa Prize Award.\n\n\n \n\n\nRÉFÉRENCES\n\n\n Wu\ , S.\, Ma\, Z.\, Luo\, X.\, Huang\, Y.\, Torres-Fonseca\, J.\, Shi\, F.\, &\; Chai\, J. (2025). The Mechanistic Emergence of Symbol Grounding in Language Models. ICML. https://arxiv.org/abs/2510.13796 \n\n\nBick\, A.\, Xing\, E.\, &\; Gu\, A. (2025). Understanding the Skill Gap in Recurre nt Language Models: The Role of the Gather-and-Aggregate Mechanism. Procee dings of the 42nd International Conference on Machine Learning\, PMLR 267\ , 4324–4344. https://arxiv.org/abs/2504.18574\n\n\nWang\, L.\, Li\, L.\, Dai\, D.\, Chen\, D.\, Zhou\, H.\, Meng\, F.\, Zhou\, J.\, &\; Sun\, X . (2023). Label Words are Anchors: An Information Flow Perspective for Und erstanding In-Context Learning. Proceedings of the 2023 Conference on Empi rical Methods in Natural Language Processing\, 9840–9855. https://aclant hology.org/2023.emnlp-main.609/ \n\n\nBisk\, Y.\, Holtzman\, A.\, Thomaso n\, J.\, Andreas\, J.\, Bengio\, Y.\, Chai\, J.\, Lapata\, M.\, Lazaridou\ , A.\, May\, J.\, Nisnevich\, A.\, Pinto\, N.\, &\; Turian\, J. (2020). Experience Grounds Language. Proceedings of the 2020 Conference on Empiri cal Methods in Natural Language Processing\, 8718–8735. https://aclantho logy.org/2020.emnlp-main.703/\n\n\nBousselham\, W.\, Petersen\, F.\, Ferra ri\, V.\, &\; Kuehne\, H. (2024). Grounding Everything: Emerging Locali zation Properties in Vision-Language Transformers. Proceedings of the IEEE /CVF Conference on Computer Vision and Pattern Recognition\, 3828–3837. https://arxiv.org/abs/2312.00878\n\n\nSzot\, A.\, Mazoure\, B.\, Attia\, O .\, Timofeev\, A.\, Agrawal\, H.\, Hjelm\, D.\, Gan\, Z.\, Kira\, Z.\, &am p\; Toshev\, A. (2025). From multimodal LLMs to generalist embodied agents : Methods and lessons. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition\, 10644–10655. https://arxiv.org/abs/2412 .08442 \n\n\nGoulet\, N.\, Massé\, A. B.\, &\; Abdendi\, M. (2025). A pproaching the Source of Symbol Grounding with Confluent Reductions of Abs tract Meaning Representation Directed Graphs. arXiv preprint arXiv:2508.1 1068.\n\n\nVincent‐Lamarre\, P.\, Massé\, A. B.\, Lopes\, M.\, Lord\, M .\, Marcotte\, O.\, &\; Harnad\, S. (2016). The latent structure of dic tionaries. Topics in cognitive science\, 8(3)\, 625-659.\n\nMot-clés : Séminaire\, LLMs\, Neurolab\, LATECE UQAM INFORMATIQUE\, LATECE\, CRIA\, Département de Neuroscience\, Sciences cognitives\, Philosophie\, Science s cognitive\, École de langues\, neurosciences cognitives\, Institut des sciences cognitives\, Apprentissage du langage\, sciences du langage\, app rentissage machine\, apprentissage profond\, langage automatique\, langage cognitif\, metacognition\, Cognition humaine\, cognition\, intelligence a rtificielle\, IA\, intelligence artificielle\, chatGPT\, enseignement sup érieur\, IA \; intelligence artificielle\, société\, IA\, intelligence artificielle\, IA\; intelligence artificielle\; société\, département d e linguistique\, maîtrise en psychologie\, doctorat en psychologie\, dép artement de psychologie\, Département d'informatique\, Faculté des scien ces de l'UQAM\, Faculté des sciences humaines\, Faculté des sciences\, d octorat en informatique cognitive\, doctorat en informatique\n\nPrix : Gra tuit\n\n CATEGORIES:Séminaire,Conférence DTSTAMP:20260914T003648Z CREATED:20260910T142302Z LAST-MODIFIED:20260911T124713Z END:VEVENT END:VCALENDAR