PAHF: Learning Personalized Agents from Human Feedback¶
L1 memory, clarification gate, and drift overwrite. No hidden persona, gold action, GPT calls, robotics, or paper shopping simulator.
Paper information¶
| Field | Details |
|---|---|
| Paper | arXiv |
| Institution | Meta Superintelligence Labs / Princeton University (first author; work done at Meta). |
| First public release | 2026-02-18 (arXiv v1) |
| Author code | facebookresearch/PAHF. |
| Local adapter or method key | pahf |
| Local reproduction code | src/auto_research/agent_research/official_meta_backfill_20261004.py |
Paper method and evidence¶
Retrieve explicit user preferences before acting and ask for clarification when unknown or uncertain. Accept corrections after acting even when memory was confident, replacing stale preferences across new users, contexts and preference drift.
flowchart LR
M[Retrieve explicit memory] --> Q[Clarify uncertainty]
Q --> A[Act]
A --> F[Accept feedback]
F --> U[Overwrite stale preference]Original paper figure¶
The original figure shows pre-action clarification and post-action correction. From the original paper; copyright remains with the authors.
Core equation¶
\(a_t=\pi_{act}(I_t,O_t,m_t,q_t,f_t^{pre})\); theoretical dynamic regret \(O(K+\gamma Tm^{-k})\).
Reported results¶
Phase-2/4 scores are 70.5%/68.8% embodied and 41.3%/70.3% shopping, ahead of no-memory and single-channel feedback. Paper offline results, not attributable production A/B or local measurements.
Local reproduction¶
Compare with mechanism disabled/default. Three-seed diagnostics are diagnostic_only=true and excluded from capability ranking.