aiqbench analysis
Programmable Semantic Interventions introduce a novel method for causal abstraction in neural networks by allowing investigators to define behavioral relationships directly via semantic programs rather than relying on curated counterfactual datasets. This approach learns low-rank activation subspaces that enable precise model steering and localization without ground-truth labels, achieving performance comparable to supervised alignment while fixing coordinate systems with known semantics.
Programmable Semantic Interventions: Learning Causal Carriers from Behavioral Programs
From arXiv
Full textarXiv:2611.21307v1 Announce Type: cross Abstract: Causal abstraction explains a frozen network by aligning its activations with the variables of a simpler, interpretable model, and checking that interventions on the two agree. Finding such an alignment currently requires a counterfactual dataset: for every pair of prompts, the answer the network ought to give once the variable has been changed. In practice, those labels are produced by an executable algorithm for the behavior, written before studying the model supposed to implement it.
However, the criterion these methods optimize only leverages the answer an intervention should produce. Here, we develop an approach that lets the investigator define that relationship directly: the answers the variable can produce, and which of them an intervention should yield. We call this interface a semantic program and introduce Programmable Semantic Interventions (PSI): a low-rank activation subspace learned so that interchanging it changes the model's answer as the program declares, with the expected answers read from the frozen model itself.
On the MIB benchmark PSI performs on par with supervised alignment at matched data and training budget, without curated pairs, and it needs no ground-truth labels; while on natural science questions with gold labels withheld, it tracks the model's own answers better than a subspace trained on gold labels. The declaration also fixes a coordinate system with known semantics, enabling donor-free steering: jointly training PSI for both localization and control realizes 94.8% of held-out requests for answers the investigator chooses, without the need for donor prompts.
Abstract:Causal abstraction explains a frozen network by aligning its activations with the variables of a simpler, interpretable model, and checking that interventions on the two agree. Finding such an alignment currently requires a counterfactual dataset: for every pair of prompts, the answer the network ought to give once the variable has been changed. In practice, those labels are produced by an executable algorithm for the behavior, written before studying the model supposed to implement it.
However, the criterion these methods optimize only leverages the answer an intervention should produce. Here, we develop an approach that lets the investigator define that relationship directly: the answers the variable can produce, and which of them an intervention should yield. We call this interface a semantic program and introduce Programmable Semantic Interventions (PSI): a low-rank activation subspace learned so that interchanging it changes the model's answer as the program declares, with the expected answers read from the frozen model itself.
On the MIB benchmark PSI performs on par with supervised alignment at matched data and training budget, without curated pairs, and it needs no ground-truth labels; while on natural science questions with gold labels withheld, it tracks the model's own answers better than a subspace trained on gold labels. The declaration also fixes a coordinate system with known semantics, enabling donor-free steering: jointly training PSI for both localization and control realizes 94.8% of held-out requests for answers the investigator chooses, without the need for donor prompts.
Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
arXiv:2611.21307 [cs.LG]
(or arXiv:2611.21307v1 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2611.21307 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 339 words
Related coverage
PlayGen2: Learning Real-Time Playable Worlds with Explicit 3D State
arXiv · Nov 6, 2026

How Can We Solve Animal Welfare's Evidence Problem? (Post 1)
Effective Altruism Forum · Oct 10, 2026

Researchers show how to automatically uncompute ancilla qubits with measurements
Quantum Zeitgeist · Nov 4, 2026
Training, learning and inference: unified dynamics of neural systems
arXiv · Aug 24, 2026