Shared Fisher-Rao Geometry: A Universal Approach to LLM Safety Control

When Controlling Large Models Is Art Rather Than Science

For a long time, steering the behavior of large language models (LLMs) has felt more like a craft than a rigorous science. Whether guiding model outputs, editing internal knowledge, or performing feature attribution, researchers have typically relied on parameter updates in Euclidean space — straightforward, but lacking a unified theoretical foundation. The problem with this approach is that every intervention can introduce unpredictable "collateral damage," reducing safety control to a patchwork of ad hoc engineering fixes.

A new study published on arXiv aims to change that. It proposes that large models across different architectures and families converge, at the output level, to an almost identical geometric structure — the Fisher-Rao metric. This finding provides a solid mathematical anchor for "universal control."

Activation Geometry vs. Output Geometry: A Crucial Distinction

One of the paper's core insights is the explicit distinction between activation geometry — the distributional structure of a model's internal activations — and output geometry — the geometric properties of the next-token probability distributions.

The paper argues that what truly correlates with model behavior and controllability is output geometry, not activation geometry. This is a critical point: many past control methods operated in activation space, but activations vary enormously across different models, making universal methods elusive. When researchers shifted their focus to output space, a surprising regularity emerged.

Remarkable Convergence Across Architectures

The study finds that Transformers, state-space models (SSMs), and recurrent neural networks (RNNs) — architecturally very different — exhibit far greater alignment in their output geometry than in their activation geometry. In other words, regardless of the underlying architecture, the geometric rules that models follow when "predicting the next word" are strikingly consistent, all approximating the same Fisher-Rao metric.

This convergence is precisely what makes universal control possible. Given a shared metric, it becomes theoretically feasible to design a "minimal-perturbation update" scheme applicable to any Transformer, state-space, or recurrent model.

---

The Fisher-Rao metric originates from the field of Information Geometry and was first introduced by statistician C.R. Rao in 1945. It is essentially a Riemannian metric on the space of probability distributions, measuring the "statistical distance" between two distributions via the Fisher information matrix. Unlike Euclidean distance, the Fisher-Rao metric naturally respects the intrinsic structure of probability distributions: when two distributions are statistically hard to distinguish, their distance is small, and vice versa. For LLMs, each prediction step outputs a probability distribution over the vocabulary, so the Fisher-Rao metric provides a mathematically "natural" tool for measuring and manipulating these output distributions. This also explains why perturbations based on this metric can be genuinely "minimal" — they seek the shortest path in the intrinsic geometry of probability distributions, rather than making crude linear moves in an artificially defined Euclidean space.

State Space Models (SSMs) are a recently emerged class of sequence modeling architectures, with the Mamba series as a prominent example. Unlike Transformers, which rely on attention mechanisms, SSMs capture sequential dependencies through linear recurrence equations, offering linear time complexity advantages for long sequences. RNNs are an older class of sequence models that propagate information across time steps via hidden states. These three architectural families — Transformers, SSMs, RNNs — differ fundamentally in their internal computation mechanisms, and their activation statistics naturally diverge considerably. This makes the researchers' observation that their output geometries are highly aligned all the more surprising: it suggests that regardless of the internal pathway taken to reach the goal of "predicting the next word," models are ultimately shaped into systems that follow the same geometric rules at the level of probability outputs.

---

One Correction, Multiple Tasks Improved

The correction method based on Fisher-Rao geometry has a remarkably broad scope of application. According to the paper [1], the same correction can simultaneously improve performance across multiple tasks:

• Steering: More precisely adjusting the direction of model outputs
• Knowledge editing: Modifying facts stored within the model
• Feature attribution: Explaining the basis for model decisions
• Dictionary learning: Extracting interpretable features
• Fine-tuning: Optimizing the model for specific tasks

The most striking reported result is that this geometry-based correction can reduce "off-target changes" by up to two orders of magnitude. In other words, while achieving the intended intervention, unintended disruption to the model's other behaviors is compressed by nearly a hundredfold. For AI systems that prioritize safety and controllability, this is a highly significant improvement.

The paper explicitly attributes this benefit to the shared Fisher-Rao geometry — it is precisely because the output geometries of different architectures are highly aligned that a universal "minimal perturbation" metric becomes viable.

---

Dictionary Learning in interpretability research refers to decomposing the high-dimensional activation vectors inside a model into sparse linear combinations of semantically interpretable "atomic features" — these atoms constitute the so-called "dictionary." The core motivation is that individual neurons in neural networks typically exhibit polysemanticity, responding simultaneously to multiple unrelated inputs and resisting straightforward interpretation. Sparse features extracted through dictionary learning, by contrast, tend to correspond to cleaner semantic concepts, making it one of the key tools in Mechanistic Interpretability research today. Anthropic and other organizations have applied it at scale to analyze the internal representations of models like Claude. Introducing Fisher-Rao geometry corrections into dictionary learning means the extracted features carry stricter geometric constraints at the probability output level, potentially improving feature quality and stability.

---

Costs and Open Questions

There is no such thing as a free lunch. Estimating the Fisher-Rao metric requires substantial gradient computation, which could mean significant additional computational overhead for very large-scale models or multimodal models.

The experiments reported in the paper cover several model families, but scalability remains an open question. The authors candidly acknowledge that validating whether the method generalizes to trillion-parameter vision-language systems will require further research. This means that while the current findings are convincing at moderate model scales, whether they hold for the most cutting-edge giant multimodal systems remains unknown for now.

A Potential New Benchmark for Safety Control

If this geometry-based correction method proves broadly effective, it has real potential to become the new starting point for future safety control benchmarking. Compared to the ad hoc solutions that have relied on intuition and trial-and-error, a principled, cross-architecture universal approach undoubtedly better aligns with AI safety research's pursuit of rigor.

A prospect worth contemplating: will future LLM toolchains embed Fisher-Rao-driven controllers into every release? If so, "safe and controllable" may shift from an after-the-fact engineering challenge into foundational infrastructure built into model design from the start.

Conclusion

The greatest value of this research lies not only in providing a concrete control tool, but in revealing a deep underlying regularity: despite the enormous diversity of large model architectures, they share — at the level of information geometry — a kind of learned, controllable universal structure. This unifying property opens a new door for LLM interpretability and safety control research. Of course, computational costs and scalability to very large models remain questions for future work to answer.