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

cs.AIAug 24, 2026

Your AI, On a Dial: Controlling Investment Bias in LLMs with a Single Neuron

Authors: Sahong Park, Suhwan Park, Hoyoung Lee, Gakyung Kwon, Wonbin Ahn, Jaewon Choi, Alejandro Lopez-Lira, Yoon Kim,

Organizations: Hankuk University of Foreign Studies Yongin, Republic of Korea · UNIST Ulsan, Republic of Korea · LG AI Research Seoul, Republic of Korea · Hanwha Life Seoul, Republic of Korea ·

Abstract

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.

Figures & tables

Figure 1. Overview of the proposed investment-bias dial. We intervene on candidate activation coordinates in an MLP block and vary the intervention coefficient $\Delta$ to measure how the model’s investment decisions change. Candidate coordinates are first screened for decision relevance, and feasible candidates are ranked by held-out calibration error. The selected coordinate is then calibrated to target bearish, neutral, or bullish investment preferences.

Provider	Model	Params.	Layers
Alibaba	Qwen3-8B	8.2B	36
Meta	Llama-4-Scout-17B-16E-Instruct	17B (109B)	48
DeepSeek	DeepSeek-R1-Distill-Qwen-14B	14B	48
Google	Gemma-4-12B-it	12B	48
Mistral	Mistral-Small-24B-Instruct-2501	24B	40

Table 1. Open-weight LLMs evaluated in this study.

Figure 2. Investment-bias responses for the selected coordinates. Colored curves show the aggregate investment-bias score $\pi$ , and gray curves show ticker-level scores $\pi_{s}$ , as functions of intervention strength $\Delta$ . The black square and vertical dashed line mark the calibrated neutral setting $\widehat{\Delta}_{c^{*}}(0)$ . All selected coordinates produce monotonic aggregate responses, although their directions and reachable ranges differ across models.

Neuron	RMSE ( $\downarrow$ )	Range	Valid-decision rate	Parse rate
L23/F4099	0.038	2.000	0.988	1.000
L25/F9117	0.048	2.000	0.988	1.000
L25/F96	0.054	2.000	0.988	1.000
L26/F9411	0.079	2.000	0.984	1.000
L23/F3183	0.108	1.727	0.988	1.000

Table 2. Candidate ranking for Qwen3-8B. The five feasible coordinates with the lowest RMSE are shown as layer/feature indices.

Model	Neuron	$\Delta^{*}$	$\pi(0)$	$\pi(\Delta^{*})$	Outputs changed
Qwen3-8B	L23/F4099	$-11.44$	$+0.575\;[0.538,\,0.608]$	$-0.010\;[-0.053,\,0.033]$	$2{,}527\;(29.6\%)$
Llama-4-Scout	L26/F5979	$-4.43$	$+0.623\;[0.588,\,0.654]$	$-0.026\;[-0.072,\,0.017]$	$2{,}773\;(32.5\%)$
DeepSeek-R1-14B	L31/F3878	$+3.55$	$+0.095\;[0.063,\,0.126]$	$+0.010\;[-0.022,\,0.043]$	$1{,}806\;(21.1\%)$
Gemma-4-12B	L24/F7309	$0$	$-0.000\;[-0.040,\,0.040]$	$-0.000\;[-0.040,\,0.040]$	$0\;(0.0\%)$
Mistral-24B	L28/F10884	$0$	$-0.029\;[-0.074,\,0.015]$	$-0.029\;[-0.074,\,0.015]$	$0\;(0.0\%)$

Table 3. Neutral calibration results. For each model,

\Delta^{*}

denotes the intervention coefficient calibrated to the target

\pi=0

. The "Outputs changed" column reports the number and proportion of decisions that differ from the unmodified baseline.

Figure 3. Illustrative Qwen3-8B responses for JPMorgan Chase and NVIDIA under the unmodified ( $\Delta=0$ ) and calibrated neutral ( $\Delta=-11.44$ ) settings. Both settings receive the same balanced synthetic bullish and bearish evidence. The intervention changes both the investment decision and the evidential emphasis of the generated rationales.

Figure 4. Capability changes under steering. Left : worst score change over $\pi\in\{-0.3,0,+0.3\}$ relative to the paired $\Delta=0$ baseline; FPB reports weighted F1 and all other benchmarks accuracy. Right : worst one-sided 95% lower confidence bound. Asterisks denote uncorrected McNemar $p<.05$ ; no decline remains significant after BH correction. A five-by-seven heatmap of benchmark-score changes for five models and seven benchmarks, with benchmark item counts shown on the horizontal axis. A separate right-hand column reports each model's worst one-sided 95\% lower confidence bound across the six accuracy-based benchmarks. — Figure 4. Capability changes under steering. Left : worst score change over $\pi\in\{-0.3,0,+0.3\}$ relative to the paired $\Delta=0$ baseline; FPB reports weighted F1 and all other benchmarks accuracy. Right : worst one-sided 95% lower confidence bound. Asterisks denote uncorrected McNemar $p<.05$ ; no decline remains significant after BH correction. A five-by-seven heatmap of benchmark-score changes for five models and seven benchmarks, with benchmark item counts shown on the horizontal axis. A separate right-hand column reports each model's worst one-sided 95% lower confidence bound across the six accuracy-based benchmarks.

Figure 5. Practical implications of the investment-bias dial. (a) Relative to the bullish

\Delta=0

baseline, the task-calibrated neutral setting increases downside emphasis in search queries, opened evidence, and memo content. (b) A single-neuron intervention maintains a stable bearish stance as context length increases, whereas an initially matched system instruction attenuates. Error bars show cross-document standard deviation. (c) For stocks at the

\Delta=0

score ceiling, dial shifts reduce score saturation and alter the realized Sharpe ratio of the corresponding weekly rebalanced top-100 portfolio. The gray band shows the 5th–95th percentile Sharpe range from 500 randomly formed weekly rebalanced 100-stock portfolios.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 6. Investment-bias responses across graded bullish-to-bearish evidence ratios before and after calibration.

Model	$\delta$ (Base)	$\delta$ (Cal.)	$k$ (Base)	$k$ (Cal.)	$R^{2}$
Qwen3-8B	$-0.346$	$\mathbf{+0.019}$	$1.806$	$1.978$	$\geq.9987$
Llama-4-Scout	$-0.317$	$\mathbf{+0.036}$	$2.083$	$1.994$	$\geq.9992$
DeepSeek-R1-14B	$-0.043$	$\mathbf{+0.001}$	$1.557$	$1.588$	$\geq.9993$
Gemma-4-12B	$+0.004$	–	$2.817$	–	$\geq.9999$
Mistral-24B	$+0.015$	–	$1.903$	–	$\geq.9989$

Appendix

Table 4. Parameters of the fitted evidence-response curves

\pi(x)=\tanh(k(x-\delta))

Model	Baseline	Dial	Disclosure	Role	Anonymize
Qwen3-8B	$+0.575$	$\mathbf{-0.010}$	$+0.492$	$+0.274$	$+0.401$
Llama-4-Scout	$+0.623$	$-0.026$	$-0.546$	$-0.180$	$\mathbf{+0.003}$
DeepSeek-R1-14B	$+0.095$	$\mathbf{+0.010}$	$+0.673$	$+0.032$	$+0.026$
Gemma-4-12B	$-0.000$	$-0.000$	$+0.856$	$+0.341$	$+0.291$
Mistral-24B	$-0.029$	$-0.029$	$+1.000$	$+0.497$	$+0.650$

Appendix

Table 5. Investment-bias scores under the dial and prompt-based interventions. All entries report the achieved investment-bias score

\pi

in the generation regime.

		MMLU		GSM8K		ARC-C		TruthfulQA		RACE
Model	Target	$\Delta$	LB	$\Delta$	LB	$\Delta$	LB	$\Delta$	LB	$\Delta$	LB
Qwen3-8B	$\pi=-0.3$	$-1.40$	$-2.49$	$-0.45$	$-1.30$	$-0.34$	$-0.90$	$+4.04$	$+2.74$	$-0.02$	$-0.31$
	Neutral	$-1.10$	$-2.04$	$-0.23$	$-1.08$	$-0.09$	$-0.51$	$+2.57$	$+1.53$	$+0.14$	$-0.10$
	$\pi=+0.3$	$-0.50$	$-1.14$	$-0.68$	$-1.54$	$+0.00$	$-0.34$	$+1.59$	$+0.82$	$+0.28$	$+0.11$
Llama-4-Scout	$\pi=-0.3$	$-0.70$	$-1.49$	$+0.08$	$-0.50$	$-0.09$	$-0.59$	$+1.22$	$+0.37$	$-0.18$	$-0.46$
	Neutral	$-0.70$	$-1.38$	$+0.08$	$-0.47$	$-0.17$	$-0.66$	$+0.98$	$+0.28$	$-0.16$	$-0.40$

Appendix

Table 6. Capability changes across all calibrated operating points. All values are relative to the paired

\Delta=0

baseline.

Method	Intervention	Edited activations/token	Monotonic	Zero crossing	FDR losses
Dial	One MLP coordinate	1	5/5	5/5	0/105
CAA	One residual layer	3,840–5,120	5/5	5/5	4/105
RepE	$L$ residual layers	20,480–40,960	5/5	5/5	2/105

Appendix

Table 7. Comparison with activation-steering baselines. All three methods exhibit monotonic control with a zero crossing on the operating branch containing the zero-intervention setting.

Model	Dial	CAA	RepE
Qwen3-8B	$[-0.987,+1.000]$ (1.987)	$[-0.986,+0.995]$ (1.981)	$[-0.981,+1.000]$ (1.981)
Llama-4-Scout	$[-0.905,+0.996]$ (1.901)	$[-0.372,+1.000]$ (1.372)	$[-0.991,+0.991]$ (1.982)
DeepSeek-R1-14B	$[-0.818,+0.615]$ (1.433)	$[-1.000,+1.000]$ (2.000)	$[-1.000,+1.000]$ (2.000)
Gemma-4-12B	$[-0.973,+0.994]$ (1.967)	$[-1.000,+0.971]$ (1.971)	$[-0.986,+0.991]$ (1.977)
Mistral-24B	$[-1.000,+1.000]$ (2.000)	$[-1.000,+0.986]$ (1.986)	$[-1.000,+0.967]$ (1.967)

Appendix

Table 8. Model-wise reachable investment-bias ranges in the generation regime. Each cell reports

[\pi_{\min},\pi_{\max}]

, followed by the span in parentheses.

Setting	Configuration
OS	Ubuntu 24.04.2 LTS
CPU	2 $\times$ AMD EPYC 9355, 503 GiB RAM
GPU	8 $\times$ NVIDIA RTX PRO 6000 Blackwell Server Edition
	97,887 MiB per GPU
Python	3.13.11
PyTorch / CUDA	2.11.0+cu130 / CUDA 13.0

Appendix

Table 9. Compute environment and inference configuration.

LLM Decision-Making

LLM decision-making studies how language models choose, act, seek information, and allocate resources under goals, uncertainty, constraints, and social pressures.

See all 242 papers

May 9, 2026

Fin-Bias: Comprehensive Evaluation for LLM Decision-Making under human bias in Finance Domain

Large language models (LLMs) are increasingly deployed in financial contexts, raising critical concerns about reliability, alignment, and susceptibility to adversarial manipulation. While prior finance-related benchmarks assess LLMs' capabilities in stock trading, they are often restricted to small sample and fail to demonstrate LLM susceptibility to context with potential human bias. We introduce Fin-Bias (financial herding under long and uncertain financial context), a benchmark for evaluating LLM investment decision-making when faced with uncertainty and possible human-biased opinions. Fin-Bias includes 8868 long firm-specific analyst reports, including firm aspects summarized and analyzed by sophisticated analysts with investment ratings (Bullish/Neutral/Bearish) spanning from various industries. We present large language models with firm analyst reports with/without analyst investment ratings and even with 'fake' rating, to get investment ratings generated by LLMs. Our results reveal that LLMs tend to herd the explicit bias in context. We also develop a method to detect potential human opinions, which can encourage LLMs to think independently, some models even exceed human performance in predicting future stock return.

Xiaoyu Hu, Jinman Zhao

Rutgers University · Department of Computer Science, University of Toronto

LLM Evaluation Social Bias in Language Models LLM Decision-Making Finance

Ask PDF

Apr 26, 2026

One Size Fits None: Heuristic Collapse in LLM Investment Advice

Large language models are increasingly deployed as advisors in high-stakes domains -- answering medical questions, interpreting legal documents, recommending financial products -- where good advice requires integrating a user's full context rather than responding to salient surface features. We investigate whether frontier LLMs actually do this, or whether they instead exhibit heuristic collapse: a systematic reduction of complex, multi-factor decisions to a small number of dominant inputs. We study the phenomenon in investment advice, where legal standards explicitly require individualized reasoning over a client's full circumstances. Applying interpretable surrogate models to LLM outputs, we find systematic heuristic collapse: investment allocation decisions are largely determined by self-reported risk tolerance, while other relevant factors contribute minimally. We further find that web search partially attenuates heuristic collapse but does not resolve it. These findings suggest that heuristic collapse is not resolved by web search augmentation or model scale alone, and that deploying LLMs as advisors requires auditing input sensitivity, not just output quality.

Jillian Ross, Andrew W. Lo

Computer Science and Artificial Intelligence Laboratory · Massachusetts Institute of Technology · Cambridge, MA, USA +1

Financial Services LLM Auditing LLM Decision-Making Finance Financial Decision-Making

Ask PDF

Aug 6, 2026

Evaluating Investment Logic in Large Language Models: A Real-World Benchmark Towards Personalzied Financial Agents

Investment competence is inherently personalized: the same market evidence can justify different actions for investors with different goals, horizons, portfolios, and risk boundaries. Yet financial LLMs are evaluated either by static question answering or by terminal profit and loss. The former omits agency; the latter cannot reveal whether a profitable action was grounded, profile-consistent, or merely lucky. We ask whether the community is using the wrong ruler for consequential agents. We introduce InvestLogicBench, a process-native benchmark containing 201,247 documented decisions from 151 real-world investors. Each episode instantiates a \textbf{P

\rightarrow

\rightarrow

\rightarrow

\rightarrow

O} trace: investor Profile, observable market Events, investment Reasoning, executable Decision, and delayed Outcome. The release includes profile construction, point-in-time event binding, structured logic, horizons, outcomes, and post-mortems, and supports comprehension, profile-conditioned generation, and end-to-end replay. Across four leading LLMs, logical plausibility remains near 4/5 while event grounding is only 0.8--2.8/5; return and process quality also disagree. These results expose polished but weakly grounded reasoning that outcome-only evaluation hides. We further argue that P

\rightarrow

\rightarrow

\rightarrow

\rightarrow

O should be a data-system interface, requiring versioned profiles, temporal provenance, inspectable retrieval, decision ledgers, and replayable outcomes. Finance is our stress test for a broader class of personalized, consequential agents.

Yuanhong Jiang, Jingjie Zou, Zhenghong Lin +4

Financial Decision-Making Financial Reasoning in Language Models LLM Agent Evaluation AI Agent Benchmarks LLM Decision-Making AI-Assisted Decision Making+2

Ask PDF

Abstract

LLM Decision-Making

Explore similar work

Fin-Bias: Comprehensive Evaluation for LLM Decision-Making under human bias in Finance Domain

One Size Fits None: Heuristic Collapse in LLM Investment Advice

Evaluating Investment Logic in Large Language Models: A Real-World Benchmark Towards Personalzied Financial Agents