Detection vs. Execution: Single-Bucket Probes Miss Half the Mamba-2 State Sink

2026-05-30 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Emerging Technologies & Innovation · Depth: Expert, quick

Summary

A study on Mamba-2 reveals that the common mechanistic interpretability assumption—that probes identifying a representational signature also identify the executing circuit—can systematically fail. Researchers found that single-bucket probes for the Mamba-2 state sink, analogous to the attention sink, recover only a small execution layer while missing a much larger detection layer with the same representational signature. The state sink decomposes into two functional head sets: BOS-specialist heads (about 5% of heads at 2.7B) causally support BOS-context and newline-target predictions, while dual heads (27-35% of heads) show stronger representational similarity but weaker causal effects. Ablating BOS-specialist heads collapsed RULER NIAH retrieval accuracy from 1.00 to 0.00 at 1024 context length in both Mamba-1 2.8B and Mamba-2 2.7B, confirming their functional importance. This distinction, implicating Mamba-2's head-shared Delta projection, highlights that separating execution from detection circuits requires class-conditional ablation.

Key takeaway

For Machine Learning Engineers interpreting Mamba-2's internal mechanisms, recognize that single-bucket probes may identify detection layers without corresponding execution circuits. Your interpretability efforts should incorporate class-conditional ablation to differentiate functional head sets, as representational similarity alone does not guarantee causal effect. This distinction is critical for accurately understanding and modifying model behavior, especially for tasks like RULER NIAH retrieval, where BOS-specialist heads are crucial.

Key insights

Representational similarity in Mamba-2 does not imply functional equivalence for mechanistic interpretability probes.

Principles

• Probes can identify detection without execution.
• Ablation is key to distinguish functional circuits.
• Mamba-2 state sink has two functional head sets.

Method

Distinguish detection from execution circuits by using class-conditional ablation, rather than just class-conditional cosine similarity, especially when probes recover both at coarse granularity.

In practice

• Use class-conditional ablation for Mamba-2 interpretability.
• Evaluate probe findings with causal effect tests.
• Beware of representational similarity alone.

Topics

• Mechanistic Interpretability
• Mamba-2
• State Sink
• Neural Network Probes
• Causal Ablation
• Large Language Models

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• TN-SHAP-G: Graph-Structured Tensor Network Surrogates for Shapley Values and Interactions — TN-SHAP-G efficiently computes Shapley values for graph-structured inputs using a graph-aligned tensor network surrogate.
• The Case for Model Science: Verify, Explore, Steer, Refine — The AI community needs a systematic "Model Science" to understand why complex models behave as they do, beyond benchmarking what they do.
• TriLens: Per-Layer Logit-Lens Entropy for White-Box Hallucination Detection — Internal model uncertainty, specifically per-layer logit-lens entropy, can effectively signal hallucination before final output.
• Beyond Task Success: Behavioral and Representational Diagnostics for WAM and VLA — WAMs improve object-level robot behavior, but architectural choices impact predictive representation and inference cost.
• Explainable deep reinforcement learning reveals energy-efficient control strategies for turbulent drag reduction — XDL-guided MARL significantly improves turbulent drag reduction and energy efficiency in fluid dynamics.
• Tiny but Trusted: Efficient Vision-Language Reasoning for Time-Series Anomaly Detection — Fine-tuning VLMs with natural-language rationales significantly improves time-series anomaly detection performance and interpretability.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.