aiqbench analysis
This study challenges the standard interpretability assumption in selective state-space models by demonstrating that top-activating heads in Mamba-2 do not uniquely define the causal circuit for tasks like needle retrieval. The research reveals a dissociation between representational signatures and functional impact, showing that distinct head sets can produce identical activation patterns while yielding vastly different intervention effects. Consequently, relying on simple probe activations to identify computational circuits in these architectures is insufficient without accounting for non-unique causal localizations.
A Circuit, Not The Circuit: Non-Unique Causal Localisation of the Mamba-2 State Sink
arXiv:2606.00930v2 Announce Type: replace-cross Abstract: Mechanistic interpretability routinely reads a probe and labels its top-activating units as the circuit executing the computation. We test the move in Mamba, on the state sink: the selective state-space analogue of the Transformer attention sink, where the Delta-gate fires disproportionately on boundary tokens such as BOS and newline. At Mamba-1 channel granularity the probe's units carry the causal effect. At Mamba-2 head granularity the label-to-locus link breaks in three ways. The causal set is not unique: a near-disjoint set of top-activation heads, sharing almost none of the specialists, matches or exceeds their ablation effect. Representation does not track function: dual heads, a quarter to over a third of all heads, respond alike to BOS and newline yet do not reproduce the specialists' causal profile. And the effect dissociates by intervention surface: perturbing the input-dependent Delta the probe reads moves the loss by at most 0.16 nats, while suppressing the same heads' output moves it by 1.5 to 7.3 nats. The heads still matter behaviourally, dropping needle-retrieval success on a RULER-style probe from 30/30 to 0/30 at 1024 context on both flagship checkpoints. A representational signature in a selective state-space model thus marks a behaviourally important head set without pinning down the circuit: a single probe recovers a circuit, not the circuit.
Abstract:Mechanistic interpretability routinely reads a probe and labels its top-activating units as the circuit executing the computation. We test the move in Mamba, on the state sink: the selective state-space analogue of the Transformer attention sink, where the Delta-gate fires disproportionately on boundary tokens such as BOS and newline. At Mamba-1 channel granularity the probe's units carry the causal effect. At Mamba-2 head granularity the label-to-locus link breaks in three ways. The causal set is not unique: a near-disjoint set of top-activation heads, sharing almost none of the specialists, matches or exceeds their ablation effect. Representation does not track function: dual heads, a quarter to over a third of all heads, respond alike to BOS and newline yet do not reproduce the specialists' causal profile. And the effect dissociates by intervention surface: perturbing the input-dependent Delta the probe reads moves the loss by at most 0.16 nats, while suppressing the same heads' output moves it by 1.5 to 7.3 nats. The heads still matter behaviourally, dropping needle-retrieval success on a RULER-style probe from 30/30 to 0/30 at 1024 context on both flagship checkpoints. A representational signature in a selective state-space model thus marks a behaviourally important head set without pinning down the circuit: a single probe recovers a circuit, not the circuit.
v2: substantially revised. The earlier detection/execution dichotomy is withdrawn and replaced by non-unique causal localisation, following additional controls: probe-set intersection/difference ablations, a size- and depth-matched cosine null with a generalist comparator, a threshold sweep, and a joint/marginal decomposition. 24 pages, 3 figures
Computation and Language (cs.CL); Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
arXiv:2606.00930 [cs.CL]
(or arXiv:2606.00930v2 [cs.CL] for this version)
Full text under the publisher's open licence · 272 words
Related coverage
J++ Lens: Jacobian Filtering Enables More Faithful Workspace Lenses
LessWrong · Oct 8, 2026
Quantum agents can reason consistently, new framework shows
Quantum Zeitgeist · Sep 10, 2026
Does preservation make sense before we know how to revive?
LessWrong · Jun 15, 2026
Closing the Loop: Non-Causal Computation, Partial Traces, & Postselected Entanglement
arXiv · Sep 9, 2026