Dial M—for Markets: Brain‑Scanning and Steering LLMs for Finance

TL;DR for operators

This paper is not mainly about whether an LLM can forecast stock moves from news. That storyline is already crowded, noisy, and full of people discovering that backtests look unusually handsome when nobody has yet met execution costs. The more useful contribution is different: it shows a way to inspect and adjust the internal concepts an LLM activates while processing financial text.

The mechanism is sparse autoencoding. Instead of treating an LLM’s internal state as a dense black-box vector, the authors insert an interpretable sparse representation into Gemma-2-9B-IT. Individual sparse features are labeled with human-readable concepts, then used as inputs to return-prediction models or as steering levers during generation.

The result is operationally interesting for three reasons. First, interpretability does not destroy performance in the paper’s main news-based portfolio exercise: the sparse representation reaches an annualized Sharpe ratio of 5.51, compared with 4.91 for a dense embedding benchmark from the same model. Second, the interpretable features can be grouped into 17 economic clusters, making it possible to ask what kinds of concepts drive the model’s signal. Sentiment, finance/markets, technical analysis, temporal concepts, and risk sit near the top. Third, steering a positivity feature exposes what looks like optimism bias in the LLM’s financial classification; reducing positivity improves the long-short strategy relative to the unsteered baseline.

The business interpretation is not “deploy this as a trading system”. Please, let us leave the backtest-to-Bloomberg-terminal pipeline to the overcaffeinated. The more durable lesson is that financial AI governance can move from output review to concept-level inspection. A firm could ask not only “what did the model say?” but “which concepts were active, which ones mattered, and what happens if we attenuate the suspicious ones?”

The boundary is equally important. The evidence is built on Reuters after-hours U.S. equity news from 2015–2024, Gemma-2-9B-IT, pretrained Google DeepMind sparse autoencoders, and specific portfolio and prompt tests. The authors argue that look-ahead bias is likely limited, but it is not magically abolished by interpretability. Any operational use would need fresh validation by model, task, market regime, execution setting, compliance requirement, and risk appetite. That last phrase is not decoration. In finance, it is often where the bodies are buried.

The useful trick is not prediction; it is opening the panel

Financial firms already know how to use models that produce scores. They have credit scores, fraud scores, churn scores, risk scores, sentiment scores, and enough dashboards to simulate a small weather service. The harder problem is not getting another number. The harder problem is knowing whether the number was produced for a reason that survives scrutiny.

That is where A Financial Brain Scan of the LLM becomes interesting.¹ The paper applies sparse autoencoders to an instruction-tuned Gemma-2-9B model and uses them to expose labeled internal features while the model processes financial news. The metaphor of a “brain scan” is slightly theatrical, but the mechanism is sober enough: take an internal residual stream, encode it into a sparse set of feature activations, decode it back so the model can continue operating, then study which sparse features lit up.

The distinction matters. A normal embedding tells you that two texts are close in vector space. A sparse autoencoder feature tries to tell you that a particular internal direction corresponds to something like positive sentiment, financial risk, technical analysis language, or temporal framing. It is not perfect mind reading. It is more like installing a labeled diagnostic panel on a machine that previously only emitted a finished answer and a vague aura of confidence.

For business use, that changes the governance question. Instead of merely asking whether the model’s output was accurate, a firm can begin asking whether the model relied on acceptable concepts. Did a research assistant over-weight optimism? Did a market-news classifier lean on entity memory? Did a customer-facing adviser become more risk-seeking after small prompt changes? Output review sees the symptom. Concept inspection starts looking at the machinery.

The mechanism: sparse concepts between hidden state and output

The paper’s core mechanism begins with the transformer residual stream. During inference, an LLM passes information through layers of internal representations. These representations are dense: many dimensions are active, and no single coordinate carries an obvious human-readable meaning. That density is efficient for prediction but deeply annoying for anyone who must explain the model to a risk committee, regulator, auditor, client, or their future self.

A sparse autoencoder changes the representation. It learns to reconstruct the model’s residual stream while forcing most latent features to be inactive for any given input. The sparsity constraint encourages a structure where a relatively small number of features activate strongly, making it easier to associate those features with concepts.

The paper then uses labeled sparse features from Google DeepMind’s Gemma Scope work on Gemma-2-9B-IT. Feature meanings come from inspecting the text contexts that strongly activate them and assigning human-readable descriptions through a labeling process. This matters because the paper is not training a finance-specific LLM from scratch. It is using a pretrained model and an interpretability layer to make the model’s internal processing more legible.

The operational flow is simple enough to sketch:

Financial news
   ↓
Gemma-2-9B-IT residual stream
   ↓
Sparse autoencoder
   ↓
Sparse labeled concept features
   ↓
Two uses:
   1. interpretable embeddings for prediction
   2. concept steering for controlled intervention

The first use is diagnostic: convert financial text into labeled sparse features, then examine which features drive predictions. The second use is interventional: choose a feature, such as positivity or risk aversion, and push the model’s generation process along that direction.

That second step is where the paper becomes more than interpretability theatre. A dashboard that says “the model used sentiment” is useful. A lever that says “reduce this positivity feature and observe whether the model becomes less optimistic” is much more useful. It moves the method from explanation into controlled experimentation.

The Reuters test asks whether interpretability carries a performance tax

The first empirical question is brutally practical: does interpretability make the model worse?

To test this, the authors use Reuters news articles related to U.S.-listed firms, matched to CRSP returns. The sample covers 2015–2024 and contains 3,664,197 after-hours news articles after cleaning and filtering. The forecasting setup is standard enough to be interpretable: use news-derived embeddings to predict the sign of subsequent intraday returns, then form a long-short portfolio by buying the top 20% of stocks by predicted probability and shorting the bottom 20%.

The model begins with 131,000 sparse features. Most of those features are irrelevant for finance; a general-purpose SAE does not politely restrict itself to markets, earnings, and macro gloom. The authors therefore rank features by predictive relevance using a principal-component logistic regression procedure, back-project the coefficients into the original feature space, and retain reduced feature sets from 5 up to 5,000 features. The rolling procedure is re-estimated through time, which is important because otherwise feature selection itself could smuggle future information into the exercise.

The headline result: the 5,000-feature sparse embedding produces an annualized Sharpe ratio of 5.51. A dense final-layer embedding benchmark from the same LLM produces 4.91. The reported difference is statistically significant at the 5% level. Accuracy numbers look small in absolute terms, as financial prediction usually does: 51.55% for the 5,000-feature sparse model versus 51.27% for the benchmark. In markets, tiny classification edges can become large portfolio metrics when applied repeatedly. This is also why one should read Sharpe ratios in papers with both interest and suspicion, preferably in that order.

The purpose of this test is main evidence, not a cute appendix detour. It supports the claim that sparse interpretability does not necessarily require sacrificing predictive performance. In fact, in this setting, the interpretable sparse representation outperforms the dense benchmark.

But the more subtle finding is the feature-count pattern. Five features already produce a nontrivial Sharpe ratio. Three hundred features recover much of the full model’s performance. Yet performance continues to improve as more features enter. That is the paper’s version of the “virtue of complexity”: the signal is not a single magic neuron called “alpha”, sadly unavailable for licensing. It is distributed across many interpretable pieces.

The clusters explain what the model seems to be using

Five thousand labeled features are technically interpretable and practically unreadable. Nobody wants a governance meeting where slide 47 is “feature 3,812 seems important, thoughts?” Sensible societies have collapsed over less.

The authors solve this by embedding the feature labels, clustering them with k-means, selecting 25 clusters through a silhouette criterion, and then manually merging economically similar clusters into 17 groups. This topic-merging appendix is best read as an implementation detail for interpretability, not as a second thesis. The key point is that the authors turn thousands of feature labels into categories that finance people can reason about.

The resulting clusters include sentiment, finance/markets, finance/corporate, technical analysis, temporal concepts, risk, fixed effects, legal, healthcare, technology, governance and politics, quantitative, punctuation and symbols, and other categories. Some labels are intuitive. Others require care. “Fixed effects” is used for features associated with named entities such as firms, people, and locations. “Punctuation and symbols” partly reflects coding-like activation patterns in the feature-labeling process, not a claim that commas secretly run Wall Street. Tempting though that may be.

The paper then asks two different questions.

First, if a model uses only one cluster, how well does it perform? That gives the cluster’s standalone predictive power. Second, if the full model excludes one cluster, how much performance is lost? That gives a leave-one-feature-group-out contribution, described in the paper as a Shapley-like Sharpe contribution.

These two questions are easy to confuse, and the difference is important.

The top marginal contributors are economically unsurprising in the best possible way. Sentiment ranks first with a Shapley Sharpe contribution of 0.54. Finance/Markets follows at 0.42. Technical Analysis and Temporal Concepts both appear at 0.41, as does Risks. Finance/Corporate follows at 0.38.

The temporal result is the most instructive. Temporal Concepts has one of the lowest standalone Sharpe ratios, 1.27, yet a high marginal contribution. That means timing concepts do not tell the model whether news is good or bad on their own. They help the model decide whether the information should matter now. A product launch “today”, a regulatory decision “next quarter”, and a strategic plan “over five years” do not carry the same trading horizon. Timing is not direction. It is context for direction.

The Quantitative cluster tells the opposite story. It does not show a meaningful marginal contribution and produces below-median standalone performance. The authors interpret this as consistent with evidence that LLMs remain weak at quantitative reasoning. A more cautious business reading is that, in this specific news-to-return setup, the model’s useful signal appears much more qualitative than numerical. It reads tone, market framing, finance language, timing, and risk. It does not become a valuation engine just because numbers appear in the article. Tragic news for anyone hoping that “LLM + earnings table” equals analyst desk in a box.

Steering turns explanation into intervention

The paper’s second act uses the same sparse representation not just to inspect the model but to steer it.

The steering idea is conceptually simple. Identify a sparse feature associated with a concept, use the decoder to map additional activation of that feature back into the residual stream, and generate with the model pushed in that conceptual direction. In the paper’s sentiment experiment, the authors use a positivity-related feature from Gemma Scope and vary the steering coefficient in both positive and negative directions.

They then prompt the model to classify each news item as positive or negative for the firm’s return. The proportion of positive classifications rises monotonically with positive steering. At a steering level of -100, the model classifies 56.0% of news as positive. With no steering, it classifies 64.5% as positive. At +100, the positive classification share reaches 77.2%.

That is the basic mechanism validation. The dial moves the thing it claims to move.

The return patterns are also consistent with the intervention. Under negative steering, positive classifications have higher average subsequent returns; under positive steering, the model becomes looser with positive labels, and those positive labels become less return-informative. This is what one would expect if the base model is somewhat over-optimistic: make it even more optimistic, and it says “positive” too often; reduce positivity, and the remaining positive calls become more selective.

The portfolio exercise makes this sharper. The unsteered long-short strategy has an annualized Sharpe ratio of 3.87. Negatively steered versions produce higher Sharpe ratios: 4.07, 4.11, and 4.28. Positively steered versions are no better and can be much worse, with reported Sharpe ratios of 3.90, 3.74, and 2.71. The paper also reports statistically significant alpha for the negatively steered portfolios relative to the baseline.

This is the most business-relevant experiment in the paper because it shows a full loop:

1. identify a concept;
2. intervene on that concept;
3. observe model behavior change;
4. test whether the changed behavior improves an operational metric.

That loop is much more useful than a post-hoc explanation pasted under a model output like a polite apology.

The paper’s evidence has different roles, and that matters

Not every experiment in the paper should be weighted equally. Some tests establish the main result. Some validate the mechanism. Some are robustness checks. Some are exploratory extensions.

The appendix neural-network table is worth noting but not over-selling. It repeats the feature-count exercise with a neural-network prediction model and reports a similar broad pattern: the full sparse-feature representation performs strongly, and performance increases with richer feature sets. This supports robustness of the embedding idea, but it is not a separate claim that neural networks solve the finance problem. We have tried that optimism before. It bought a lot of GPUs.

The business value is concept governance, not a trading bot

For a financial institution, the most valuable version of this method is not “an LLM that trades Reuters headlines”. It is a concept-level governance layer around LLM-mediated decision support.

That layer could matter in several workflows.

In investment research, sparse features could help analysts understand whether a model’s recommendation comes from sentiment, sector language, timing clues, risk framing, named entities, or something suspiciously close to memorized event context. The output becomes inspectable before it becomes influential.

In compliance and model risk, steering can test whether a model’s advice changes excessively when risk, optimism, loss, or wealth-related concepts are amplified or suppressed. That is closer to stress testing than prompt testing. Prompt testing asks whether a different wording breaks the model. Concept testing asks whether a different internal emphasis changes the decision.

In customer-facing advisory systems, concept steering could support controlled persona variation, risk tolerance simulation, or bias attenuation. The key phrase is “could support”. Nobody should directly steer a retail-investment assistant’s internal risk dial without validation, logging, suitability checks, and a very alert legal department. Preferably one that has slept.

In enterprise AI governance, the paper points toward a useful separation of responsibilities:

The ROI case, if one exists, would not come from replacing analysts. It would come from reducing diagnostic cost. When a model behaves badly, firms currently spend too much time debating prompts, examples, evaluation sets, and vibes. Sparse concept inspection gives them a more structured way to ask where the behavior came from.

The boundaries are not footnotes; they define the product risk

The paper is careful enough to provide useful boundaries, and those boundaries should travel with the result.

First, the model is Gemma-2-9B-IT with open sparse autoencoders. That is good for reproducibility, but it also means the results should not be casually transferred to larger proprietary models, smaller local models, or finance-tuned systems. Different models may encode concepts differently. A positivity feature in one model is not a universal positivity organ floating around the machine-learning ether.

Second, the dataset is Reuters after-hours firm news matched to U.S. equity returns. That is a clean setting for studying news-driven short-horizon predictability. It is not the same as credit underwriting, macro research, portfolio construction with transaction costs, private-market diligence, ESG controversy monitoring, or client suitability advice.

Third, look-ahead bias is addressed but not annihilated. The authors argue that risk is limited because Gemma-2-9B-IT is relatively small, because the analyses are comparative within the same model, and because much of the paper is about interpretability rather than raw forecasting. That is reasonable. It is not equivalent to proving that memorized information plays no role. The Fixed Effects cluster, associated with named entities, is a useful reminder that entity-specific representations deserve scrutiny.

Fourth, steering is powerful precisely because it intervenes inside the model. That makes it attractive for controlled experiments, but also operationally sensitive. A steering coefficient is not a compliance policy. It is an intervention that must be validated against downstream behavior, failure modes, stability, and user context.

Finally, the financial metrics are research metrics. Sharpe ratios in clean academic exercises do not automatically survive trading frictions, capacity limits, latency, shorting constraints, market impact, data entitlements, corporate-action handling, and the general market habit of punching elegant ideas in the face.

What Cognaptus would take from this

The paper’s deeper lesson is architectural. LLM oversight should not stop at the answer layer.

A mature AI stack for financial work needs at least three levels of control. The first is output evaluation: was the answer accurate, compliant, and useful? The second is process evaluation: did the model use the right tools, data, and retrieval sources? The third is representation evaluation: which internal concepts shaped the answer?

Most current deployments are still fighting at level one. Some serious teams are building level two with retrieval logs, tool traces, and evaluation harnesses. This paper points toward level three.

That does not mean every business needs sparse autoencoders in production next quarter. It does mean the direction is clear. As LLMs move into higher-stakes decision support, firms will need controls that are more precise than “improve the prompt” and less expensive than “fine-tune the whole model again and hope the weird behavior leaves quietly”.

Concept-level inspection and steering offer a middle path. They make model behavior more diagnosable, interventions more controlled, and evaluation more connected to the mechanisms that produced the output.

The best business use of this paper is therefore not to chase the trading result. It is to ask a harder operating question: where in our AI workflow do we need a concept panel rather than another confidence score?

Because when an LLM says “positive”, “safe”, “material”, “low risk”, or “probably fine”, the interesting question is not whether the word sounds professional. The interesting question is which dial inside the model just moved.

Cognaptus: Automate the Present, Incubate the Future.