iiinova — where your digital self keeps evolving

Mandana Samiei

@mandanas.bsky.social

2.6K followers 442 following 18 posts

PhD candidate at McGill and Mila - Quebec AI Institute w/ Blake Richards and Doina Precup. Interested in Reinforcement Learning, Continual Learning and Human Cognition. Based in Montreal. 🇨🇦

Posts Replies Media

Mandana Samiei @mandanas.bsky.social · 19/06/2026

While the top humans still outperform models in proposing interventions that bear the greatest cumulative information gain and most efficient elimination of hypotheses, most models approximate average human performance.

141 0

Mandana Samiei @mandanas.bsky.social · 19/06/2026

Conjunctive causal reasoning is more challenging for both LLMs and humans, with GPT-5 approaching top-human level performance

140 0

Mandana Samiei @mandanas.bsky.social · 19/06/2026

We recently investigated how humans and AI reason through causal rules: arxiv.org/abs/2606.06464 Accepted at #CogSci2026. We gave adults and LLMs agency to actively experiment using a Blicket causal game, breaking down their process into the iterative loop shown in the following figure:

13812 0

Mandana Samiei @mandanas.bsky.social · 19/06/2026

Conjunctive causal reasoning is more challenging for both LLMs and humans, with GPT-5 approaching top-human level performance.

000 0

Mandana Samiei @mandanas.bsky.social · 04/06/2026

These parallels led us to a broader point: Declarative memory may lie along a continuous spectrum of abstraction: Episodic → autobiographical → semantic → schematic Read here: www.cell.com/neuron/fullt...

110 0

Mandana Samiei @mandanas.bsky.social · 04/06/2026

Inspired by older connectionist theories like harmony theory, we propose: • Matching existing activation patterns → minimal synaptic change • Partial matches → distortions or false recall • Strong mismatches → widespread synaptic reorganization

100 0

Mandana Samiei @mandanas.bsky.social · 04/06/2026

We revisit Piaget’s classic ideas of integration, assimilation, and accommodation through the lens of modern neuroscience and AI. • Integration: new info fits existing knowledge • Assimilation: new info is reshaped by prior knowledge • Accommodation: representations reorganize for novelty

100 0

Mandana Samiei @mandanas.bsky.social · 16/05/2025

Many LMs default to disjunctive inferences—even when given conjunctive evidence. Unlike children, LMs’ exploration is shaped by the underlying causal rule. Could promoting child-like curiosity help LMs reason more effectively about causality? More details: arxiv.org/abs/2505.09614

030 0