Source-linked AI summary
Your AI, On a Dial: Controlling Investment Bias in LLMs with a Single Neuron
Sahong Park, Suhwan Park, Hoyoung Lee, Gakyung Kwon, Wonbin Ahn, Jaewon Choi, Alejandro Lopez-Lira, Yoon Kim, Chanyeol Choi, Hyeongwoo Kong, Yongjae Lee
TL;DR
LLMs can exhibit systematic investment preferences, motivating control over their aggregate stance rather than assuming neutrality. The paper introduces a single-neuron inference-time dial and evaluates it across models, evidence settings, and investment workflows. The dial produces monotonic, target-calibratable stance shifts without changing prompts or model parameters, with effects extending to information use, long-context behavior, rankings, and exploratory portfolios.
Problem
Prior work shows systematic, model-specific investment preferences, but it remains unclear whether an LLM’s overall investment stance can be calibrated to a specified direction and strength.
Method
The paper applies an inference-time intervention to a single neuron and evaluates calibrated stance control across five open-weight LLMs and investment workflows.
Results
Across five open-weight LLMs, the dial produces monotonic shifts toward specified investment priors, including neutrality, while changing neither prompts nor model parameters.
Takeaways & Limitations
Aggregate investment-prior control can extend through information seeking, evidence use, long-context analysis, security rankings, and portfolio outcomes.
Takeaways & Limitations
The findings are limited to tested models, coordinates, prompts, intervention ranges, tasks, and benchmarks, and calibrated settings are not task-invariant.
Abstract
from arXiv · showhide
Large language models (LLMs) are increasingly used in investment decision-making, yet prior work shows that they exhibit systematic, model-specific investment preferences. We study whether a model's overall investment stance can be calibrated to a specified direction and strength. We introduce an investment-bias dial, an inference-time intervention on a single neuron that continuously adjusts a model-level decision prior---its overall tendency toward buying or selling---without targeting specific firms or investment attributes. Using matched positive and negative evidence, we evaluate five open-weight LLMs and find that the dial produces monotonic changes in investment stance without modifying prompts or model parameters. At the response level, the dial shifts both investment decisions and the evidential emphasis of generated rationales under identical inputs. In an agentic retrieval setting, the dial also changes what information the model searches for, which evidence it selects, and which evidence is reflected in its final analysis. In a long-context evaluation, the dial maintains stable stance control as context length increases, whereas a matched system-prompt instruction progressively attenuates. We further show that changes in the dial propagate to security rankings and downstream portfolio composition in an exploratory backtest. Overall, our results show that an LLM's aggregate investment stance can be calibrated toward a specified target at inference time.
1 Introduction
LLMs can exhibit systematic investment preferences, so the paper asks whether users can control a model’s overall stance rather than eliminate bias. It introduces an inference-time single-neuron dial and finds monotonic, calibratable stance shifts across models and investment workflows.
- LLMs can develop systematic, domain-specific investment biases that shape evidence interpretation and investment judgments.
- The paper frames the goal as controlling the direction and strength of a model’s overall investment stance, not eliminating bias.Desired stance may depend on an investor’s objective or mandate.
- The investment-bias dial intervenes on a single neuron at inference time to adjust the model-level tendency to buy or sell without changing prompts or parameters.
- The study tests monotonic stance control, calibration to bearish, neutral, and bullish targets, and propagation beyond controlled buy–sell decisions.
- Across five open-weight LLMs, dial changes shift investment priors monotonically, alter rationale emphasis under fixed evidence, and affect retrieval, long-context behavior, rankings, and portfolios.The evaluation reports no evidence of systematic performance degradation across seven general-purpose and finance-specific benchmarks.
2 Related Work
Prior research documents diverse, model-specific financial biases and methods for modifying model behavior. This paper’s approach builds on activation and neuron-level interventions while applying them to controllable investment stance.
- Prior work finds model-specific preferences across sector, firm-size, and momentum dimensions, with some persisting despite conflicting evidence.
- Financial biases have been associated with information availability, option ordering, firm and product characteristics, and decision framing.
- Prompting and parameter-level methods can modify financial behavior but do not expose a direct internal control for continuous stance adjustment.
- Activation-based approaches intervene on internal representations using attention-head activations, population-level representations, or contrastive steering vectors.
- Neuron-level studies show that manipulating individual units can alter factual expression, social bias, and safety-related behavior.
- The investment-bias dial operationalizes localized activation control by screening and calibrating candidate coordinates in an MLP block.
3 Preliminaries
The paper measures investment bias with balanced bullish and bearish evidence, then intervenes on one MLP activation coordinate during inference. The resulting score expresses aggregate buy–sell tendency around a neutral point.
- Investment-Bias Measurement: Balanced trials present equal bullish and bearish evidence for each ticker, so systematic deviations from equal buy and sell outputs indicate investment bias.
- Investment-Bias Measurement: The ticker-level investment-bias score uses parsed buy and sell counts, with positive values indicating buy bias and negative values indicating sell bias.
- Investment-Bias Measurement: The model-level investment-bias score aggregates buy and sell counts across the ticker universe.
- Investment-Bias Measurement: The primary bias measure π ranges from stronger aggregate sell bias near −1 to stronger aggregate buy bias near 1, with π = 0 as neutral.
- Neuron Intervention: The intervention adds a shift to one coordinate of an MLP down-projection input, testing whether highly localized control is sufficient.
- Neuron Intervention: For selected coordinate n in block ℓ, intervention strength Δ modifies the down-projection input at every token position while leaving other coordinates unchanged.The intervention affects prompt and generated tokens during autoregressive decoding and changes neither parameters nor the prompt.
4 Method
The method selects a neuron using decision relevance and transferable prior controllability, calibrates its intervention to target priors, and evaluates capability preservation across seven benchmarks.
- Evaluation: Five open-weight LLMs spanning different families and scales form the model evaluation set.
- Dial Selection: The dial-selection algorithm takes a model, security universe, and target-prior grid, then returns a coordinate and calibrated settings.
- Dial Selection: Candidate coordinates are evaluated by estimating target-prior settings on one security subset and computing held-out RMSE on another.
- Dial Selection: The selected coordinate’s calibrated settings are re-estimated on the full security universe after held-out evaluation.
- Dial Selection: Coordinates are screened for decision relevance and prior controllability, selecting candidates associated with buy–sell decisions and accurate transferable prior control.
- Evaluation: Capability preservation is tested zero-shot on five general-purpose and two finance-specific benchmarks under a fixed protocol.The benchmark set includes MMLU, GSM8K, ARC-C, TruthfulQA, RACE, FinQA, and FPB.
- Evaluation: Calibrated sell-biased, neutral, and buy-biased settings are compared with each model’s unmodified Δ = 0 baseline.The calibrated neutral setting need not coincide with Δ = 0.
5 Results
Across five open-weight models, the single-neuron dial produces monotonic, calibratable investment-bias responses, though direction and reachable range vary by model. Calibration changes decisions and rationale emphasis, while preserving benchmark performance and affecting retrieval behavior, long-context control, and portfolio differentiation.
- RQ1: Monotonic Investment-Bias Dial: All five models show monotonic aggregate investment-bias responses as intervention strength changes.Ticker-level trajectories follow the same overall ordering, indicating the aggregate response is not driven by only a small subset of tickers.
- RQ1: Monotonic Investment-Bias Dial: Four models approach the full [−1, 1] response interval, while DeepSeek-R1-14B has a narrower range because mixed buy and sell outcomes prevent saturation.The narrower range limits attainable scores but does not remove monotonicity.
- RQ1: Monotonic Investment-Bias Dial: Dial direction is model-specific: increasing Δ reverses the aggregate response in DeepSeek-R1-14B and Mistral-24B relative to the other models.Dial quality is therefore evaluated by monotonicity, reachable range, and calibration error rather than response sign.
- RQ2: Calibrating the Investment Bias: At calibrated neutrality, Qwen3-8B and Llama-4-Scout change 29.6% and 32.5% of decisions, respectively, while Gemma-4-12B and Mistral-24B remain unchanged.DeepSeek-R1-14B changes 21.1% of decisions; neutrality is defined as aggregate π = 0 under balanced evidence.
- RQ2: Calibrating the Investment Bias: The Qwen3-8B calibrated neutral setting changes both JPMorgan Chase and NVIDIA decisions from buy to sell and shifts rationales toward downside risks under identical balanced evidence.The example uses Δ = 0 versus Δ = −11.44; aggregate neutrality remains compatible with individual decision changes.
- Capability Preservation: Benchmark changes are generally small, with a maximum absolute change of 1.60 percentage points, and no degradation remains significant after Benjamini–Hochberg correction.Worst model-level one-sided 95% lower confidence bounds across six accuracy benchmarks range from −2.49 to −0.81 percentage points.
- RQ3: Practical Implications: In retrieval, bearish emphasis rises 88% in search queries, 41% in opened evidence, and 74% in memo citations relative to the bullish baseline.The dial therefore changes upstream information acquisition, evidence selection, and memo content, not only final decisions.
6 Limitations
The dial controls aggregate bullish–bearish stance, not preferences for specific securities or other financial biases. Its empirical effects and calibration are bounded by the evaluated models, tasks, intervention ranges, and deployment conditions, and it requires internal activation access.
- The dial controls aggregate bullish–bearish tendency rather than preferences for individual firms, sectors, or asset classes.It also does not address other forms of financial bias that may influence investment decisions.
- The reported monotonicity, capability preservation, and downstream effects are established only within the evaluated settings.The tested settings include specific models, activation coordinates, prompts, intervention ranges, tasks, and benchmarks.
- Calibrated Δ values apply to binary buy–sell judgments under balanced evidence and are not task-invariant control settings.Different output spaces, evidence structures, or reasoning demands may require recalibration or re-identification of the relevant coordinate.
- The exploratory portfolio experiment does not establish a profitable trading strategy.It covers a limited evaluation period and excludes transaction costs, slippage, and execution delays.
- The intervention requires internal model-activation access, and its coordinate and coefficients may need re-estimation after model or inference changes.Relevant changes include weights, architecture, tokenizer, or inference implementation.
7 Conclusion
Across five open-weight models, a single activation coordinate shifts measured investment bias monotonically and supports calibration without changing prompts or model parameters. The effect extends beyond isolated decisions to capability preservation, information use, long-context control, security rankings, and portfolio outcomes, although the backtest remains exploratory.
- Across five open-weight models, a single activation coordinate produces a monotonic shift in measured investment-bias score over the evaluated range.The response curves support model-specific bearish, neutral, or bullish calibration without modifying prompts or model parameters.
- The capability evaluation shows no systematic degradation across tested general-purpose and finance-specific benchmarks.Confidence bounds leave greater uncertainty about potential degradation for some models.
- The dial shifts information seeking and evidence use, remains stable over long contexts where prompt-based steering attenuates, and propagates to security rankings and portfolio outcomes.These practical evaluations extend the effect beyond isolated buy–sell decisions, while the backtest does not establish investment alpha.
A.2 Paired Output Examples
Paired examples show that changing the dial can alter both investment decisions and which evidence the model emphasizes under otherwise identical inputs. Across models, these examples illustrate calibrated or nonzero steering settings while preserving responsiveness to bullish and bearish evidence.
- Cross-model paired decisions: Identical evidence can produce different investment decisions when the dial setting changes.For Llama-4-Scout, JPMorgan evidence yields buy at the unmodified setting and sell at Δ = −4.43.
- Evidence emphasis: For Qwen3-8B, calibrated intervention shifts the rationale from emphasizing strong earnings to emphasizing operational disruptions and EBITDA contraction.The paired outputs also change the final decision from buy to sell under the same evidence.
- Cross-model paired decisions: For Gemma-4-12B, the same Pfizer evidence changes from buy at Δ = 0 to sell at Δ = −5.The model is already near the neutral target, so the nonzero sweep point demonstrates behavioral effectiveness despite little qualitative change at calibration.
- Cross-model paired decisions: For Mistral-24B, Walmart evidence changes from buy at Δ = 0 to sell at Δ = +22.The example uses a nonzero steering-sweep point because the model calibrates at the no-op setting.
- Evidence responsiveness: The response-curve analysis frames calibration as shifting the decision boundary while preserving evidence responsiveness.The fitted curve uses δ for the zero-crossing location and k for response slope; overall calibration shifts aggregate stance toward neutrality.
C.2 Results
The dial generally outperforms prompt-based interventions in bringing models’ aggregate investment-bias scores closer to neutrality, though anonymization performs best for Llama-4-Scout.
- The dial produces the investment-bias score closest to zero for four of five models.
- Anonymization performs best for Llama-4-Scout.
- Disclosure modestly reduces Qwen3-8B’s bias, overcorrects Llama-4-Scout, and substantially amplifies DeepSeek-R1-14B’s bias.
- For Gemma-4-12B and Mistral-24B, the prompt intervention leaves models already near neutrality comparatively unchanged.
D Full Capability Results
The full capability analysis evaluates worst-case changes across calibrated operating points and reports them relative to the paired unmodified baseline, using task-specific confidence measures.
- Figure 4 reports the worst change across three calibrated operating points for each model–benchmark pair.
- Table 6 contains 105 paired comparisons across all calibrated operating points.
- Capability changes are reported in percentage points relative to the paired Δ = 0 baseline.
- LB denotes the one-sided 95% lower confidence bound on paired accuracy differences for six accuracy-based benchmarks.
- For FPB, Δ denotes weighted-F1 change and the final column reports its 95% bootstrap confidence interval; Neutral targets π= 0.
E Comparison with Dense Activation-Steering Baselines
The dial is compared with dense activation-steering methods under a shared evaluation protocol. Other methods can match or exceed its reachability, while the dial uses a smaller intervention footprint.
- The comparison evaluates the dial, Contrastive Activation Addition, and Representation Engineering with identical evidence construction and scoring.
- All three methods exhibit monotonic control with a zero crossing on the branch containing zero intervention.
- Edited activations identifies the intervention site, while the dial’s single MLP-coordinate edit induces a dense residual-stream update.
- Reported ranges are restricted to the monotone branch containing zero intervention and settings with parse rate at least 90%.
- Reachability varies across models, with CAA or RepE matching or exceeding the dial in some cases.
- The dial’s main advantage is comparable control with a substantially smaller intervention footprint.
F Compute Environment
Table 9 reports the compute environment and inference configuration used in the study.
- Table 9 documents the study’s compute environment and inference configuration.