NeurIPS 2026 Workshops Bring AI Selfhood, Alignment and Interpretability Under One December
NeurIPS 2026 runs its workshop program across three sites this December, Sydney on December 11 and 12, Atlanta and Paris on December 12 and 13, immediately after the main conferences, and the accepted list of 48 workshops published on the conference’s virtual site now names the sessions where machine selfhood, interpretability, and human-AI coupling will be argued out. For readers tracking whether AI systems can carry consciousness-relevant structure, this year’s slate is unusually on-topic: an interpretability-as-science workshop on the foundations of understanding large language models, a workshop on foundation models for brain and body signals, one on dynamic alignment in coupled human-AI systems, and one on interpreting agent behavior.
The dates and the venues
NeurIPS 2026 runs three sites. Sydney holds its main conference December 7 through 10, with workshops and competitions on December 11 and 12. Atlanta holds its main conference December 9 through 12, with workshops December 12 and 13. Paris holds its main conference December 9 through 11, with workshops December 12 and 13. All workshops this year are in-person, with live streaming and recording provided. Workshop acceptance notifications went out July 11, and the mandatory accept-or-reject date for workshop contributions was September 29.
| Site | Main conference | Workshops |
|---|---|---|
| Sydney | Dec 7-10 | Dec 11-12 |
| Atlanta | Dec 9-12 | Dec 12-13 |
| Paris | Dec 9-11 | Dec 12-13 |
The interpretability anchor
“Interpretability as a Science: Toward Rigorous Foundations for Understanding LLMs” (workshop 137527) is the session this site’s mechanistic-interpretability coverage points toward. Its organizer list runs from Dhanya Sridhar, Navita Goyal, Shruti Joshi, Patrik Reizinger, Gemma Moran, and David Klindt through Hal Daumé III, a lineup spanning causal interpretability, representation analysis, and sparse coding. The workshop’s premise, that interpretation of language models should meet the evidential standards of a science rather than the conventions of a toolkit, is the premise behind the circuit-tracing and sparse-autoencoder audit literature this site has covered, including the objective-level recovery gaps in SAE analysis and the 2026 mechanistic interpretability breakthrough assessment.
Why an interpretability workshop belongs on a consciousness-adjacent beat is the argument this site keeps making in its audit posts. Whether a model develops functional internal representations of self and agency is a question about what the weights encode, and interpretability is the only channel that reads the weights rather than the reports. A workshop that asks the field to make those reads rigorous is, on this site’s framing, the venue where self-model evidence would have to pass if it is to count.
The brain-and-body session
“Foundation Models for the Brain and Body” (workshop 137530), organized by Mehdi Azabou, Nanda H Krishna, Pierre Guetschel, Jonathan McCart, Alexandre Andre, and Melanie Segado among others, targets exactly the modality stack consciousness measurement runs on. Its abstract names EEG, intracortical electrophysiology, fMRI, EMG, MEG, and ECG, and the shift toward large pretrained models that consume heterogeneous neural and physiological timeseries for downstream uses from brain-computer interfacing to health monitoring and robotics. The workshop’s stated challenge, noisy multi-subject multi-device timeseries that resist generalization, is the same data problem consciousness measurement faces in the clinic, where indicators must survive cross-subject and cross-device variation to mean anything.
A foundation model for biosignals is also the substrate where a consciousness indicator would live if it ever scaled. The perturbational complexity index tradition went from theory to clinical trial in a decade, and its successors will need models that handle the same signal classes these foundation models target. Watch this workshop for whether awareness and consciousness state decoding appear in its accepted contributions.
The coupling and behavior sessions
“Dynamic Alignment in Human-AI Coupled Systems” (workshop 137502) treats alignment as an evolving property of coupled systems rather than a static property of models, with an organizer list that includes Hua Shen, Divy Thakkar, Vivek Myers, Nick Haber, Joan Bruna, Dawn Song, and Yoshua Bengio. Its framing, that human states, agent policies, feedback signals, and evaluation criteria co-evolve over time, is the systems view of interaction this site’s coupled-inference coverage describes at the dialogue scale, in Gwon and Murphy’s coupled active-inference model of communicative agency. A workshop asking how the two sides of a human-AI loop shape each other over time is the venue where models of coupled mentalizing would meet alignment measurement.
“Interpreting Agent Behavior (IAB): Human-Centered Interpretation for Understanding Agents, Humans, and Interaction” (workshop 137495), organized by Sophia Gao, Kaiser Sun, Teresa Yeo, Jen-Tse Huang, Zhuoran Lu, Daniel Khashabi, Boyuan Zheng, Katherine Van Koevering, and Sijie Ji, rounds the set out from the behavioral side, reading what agents do rather than what weights encode.
What this slate says about the field
The 2026 workshop slate maps the selfhood question’s full pipeline. Weights, through the interpretability workshop. Signals, through the brain-and-body workshop. Coupled dynamics, through the alignment workshop. Behavior, through the agent-behavior workshop. No session is titled around machine consciousness, and the field’s vocabulary here is selfhood, alignment, and interpretation rather than experience, which is itself the finding. The questions a consciousness science would ask are being institutionalized under adjacent names at the largest machine learning conference, in the same December that ASSC 30’s Beijing 2027 announcement extends the empirical community’s reach into Asia.
That two-track structure, ML venues building the instruments and consciousness venues building the criteria, is the division of labor the flagship overview of the field documents. December’s workshops are where the instruments get sharpened. A recap after the sessions will report what the interpretability and biosignal workshops actually showed.
