What Goes Where? Information Flow in Long-Horizon AI Agents

Seminar is open to everyone. 

Submit email for Teams link: Registration for Teams link(open to everyone)-OxW*CS Seminar: Prof. Niloofar Mireshghallah – Fill in form

Abstract:
An AI agent's behavior depends on what enters its context, what
persists in memory, and how observations from tools become actions.
What goes where, and what changes along the way? In this talk, we
examine language models as components of software systems that gather,
synthesize, transform, and pass on information, focusing on
long-horizon, open-ended tasks where state accumulates and feedback is
often incomplete.

We begin with CIMemories (https://arxiv.org/abs/2511.14937), examining
when information remembered about a user is appropriate to disclose in
a new context. Through Et Tu, Brute?
(https://arxiv.org/abs/2609.24927), we then trace how personal
information obtained through profiles, inboxes, and MCP-style tool
access influences downstream decisions, sometimes steering
recommendations against the user's stated interests. We next turn to
Persimmon and the Trickle Test
(https://persimmon.humansand.ai/blog/persimmon.html), asking whether
user models capture how people reveal information gradually and
selectively as an interaction unfolds.

Finally, we connect these questions to scientific discovery. Both
social and scientific tasks require deciding what to ask or measure
next, maintaining evolving state, and acting on partial information
with imperfect feedback. Using SMDD-Bench
(https://arxiv.org/abs/2605.21740) and our work on harness engineering
(https://r-suresh07.github.io/writing/is-human-taste-overrated-in-harness-engineering/),
we examine how the representation of evidence and persistence of state
shape scientific search. We close with directions for models and
harnesses that better reason about what information to seek, retain,
abstract, and pass on across a workflow.