The research team’s framework, described in a paper titled Dynamic risk budgeting via large language models: A context-aware framework for adaptive portfolio allocation, uses GPT-4.1 to infer the prevailing market regime and compute relative risk intensities for a universe of 55 U.S. equities spanning eleven sectors. The empirical results are eye-catching. Over the 2023 to 2024 evaluation window, the LLM-based approach achieved a Sharpe ratio of 2.150, a 46.4 percent improvement over conventional risk parity, which scored 1.469, and a clear lead over regime-switching strategies built on Hidden Markov Models, which reached 1.597. The Sharpe ratio, which measures excess return per unit of volatility, is the standard yardstick of risk-adjusted performance, and a gap of this magnitude is rare in a field where marginal gains are celebrated.

To understand why the approach works, it helps to see what it replaces. Classical risk parity allocates portfolio risk equally across assets under static budget constraints, a philosophy popularized in institutional finance for its diversification benefits and its relative indifference to return forecasts, which are notoriously difficult to estimate. The trouble is that volatility clusters, correlations shift, and the character of drawdowns changes with the economic weather. A budget that made sense in a low-volatility expansion can be badly miscalibrated when downside volatility spikes. The authors argue that what is needed is not a better static rule but a mechanism that continuously reinterprets market conditions and adjusts the budget accordingly.

Hidden Markov Models have long been the academic tool of choice for this problem. Since James Hamilton’s landmark 1989 work on regime-switching in economic time series, researchers have used HMMs to partition markets into discrete states, such as bull, bear, and choppy, with probabilistic transitions between them. The new paper acknowledges this lineage but identifies two structural limitations. HMMs operate over discrete state spaces, so they approximate a continuously evolving market with a finite set of snapshots. They also assume Markovian transitions, meaning the next state depends only on the current one, which constrains how richly the model can represent gradual, path-dependent changes in market character. An LLM, by contrast, can reason over a structured summary of recent market behavior and articulate a regime in continuous, nuanced terms.

The technical plumbing of the framework is as interesting as the headline numbers. The researchers encode financial data in Token-Oriented Object Notation, abbreviated TOON, a format designed to present structured information to language models efficiently. The system prompt establishes the model’s role: it is told it is a quantitative analyst specializing in risk-based portfolio construction, tasked with analyzing market conditions and assigning risk budgets to the 55-stock universe while considering volatility patterns, market regime, and cross-sectional characteristics. The model’s output, a set of time-varying risk budgets, then feeds into a standard risk budgeting optimization that translates budgets into portfolio weights. In effect, the LLM occupies the judgment layer while classical optimization handles the arithmetic.

Perhaps the most practically valuable finding concerns prompt design, the craft of instructing language models. The authors ran a series of scenarios varying what guidance the model received, and the results were sharply asymmetric. Focused directional guidance, embodied in a volatility asymmetry prompt, was the most consistent driver of risk-adjusted performance across specifications. That prompt instructed the model to prioritize assets with lower downside volatility relative to upside volatility, allocating proportionally greater risk budgets to stocks that tend to gain more in up-moves than they lose in down-moves, a principle grounded in the empirical observation that low-downside-volatility stocks deliver superior risk-adjusted returns over time. In the primary configuration, this guidance worked best when combined with input dimensionality reduction, a pattern the authors attribute to attention concentration effects: a narrower, more targeted information set helps the model focus on what matters.

By contrast, handing the model a comprehensive feature set without targeted guidance produced only marginal improvements over the baseline. This is a cautionary lesson for the growing crowd of practitioners experimenting with LLMs in finance. The model is not a universal oracle that extracts signal from any pile of data; its performance depends critically on how the problem is framed, which features are surfaced, and whether the prompt encodes sound financial priors. The difference between a well-designed prompt and a kitchen-sink one was, in this study, the difference between a substantial edge and statistical noise.

The authors are admirably explicit about the caveats, and they matter. The backtest period of 2023 to 2024 was characterized by favorable market conditions, which flatters any long-biased equity strategy and makes absolute performance figures hard to extrapolate to bear markets or prolonged drawdowns. More sobering still, preliminary experiments with alternative LLM architectures yielded substantially inferior results, suggesting that the framework’s success may be tightly coupled to the specific capabilities of GPT-4.1 and may not generalize across models. In other words, the finding is best read as a proof of concept for a class of techniques rather than a turnkey trading system. The researchers also note that temperature parameter sensitivity was examined in robustness checks, addressing concerns about the stochasticity of model outputs.

The study situates itself within a fast-moving literature on language models in finance, including work on FinBERT for financial sentiment analysis, BloombergGPT as a domain-specific foundation model, and zero-shot analyses of ChatGPT’s ability to forecast stock price movements. What distinguishes this contribution is its integration of LLM reasoning into a rigorous risk budgeting pipeline, rather than using the model to predict returns directly. By asking the model a question it can plausibly answer, namely how risk should be distributed given current conditions, and leaving return estimation out of the loop, the framework sidesteps some of the estimation error problems that have plagued mean-variance optimization since Chopra and Ziemba showed how devastating errors in means, variances, and covariances can be for optimal portfolio choice.

The broader significance extends beyond trading floors. The paper documents both the potential and the limitations of foundation models in financial decision-making, offering evidence relevant to any domain where decisions must adapt to shifting contexts that resist clean discretization. The idea of using a language model as a context-aware layer atop classical optimization could transfer to supply chain risk, energy grid balancing, or insurance underwriting, wherever static rules meet dynamic environments. For investors, the message is tempered but real: the era in which a language model can serve as a disciplined, promptable component of a quantitative investment process has arrived, at least in favorable markets and with the right model. The datasets analyzed derive from publicly available Yahoo Finance data, and the authors report that processed data and code are available from the corresponding author upon reasonable request, which should make replication and stress-testing under harsher market conditions feasible for the research community.

Subject of Research: Using large language models for dynamic, context-aware risk budgeting in portfolio allocation

Article Title: Dynamic risk budgeting via large language models: A context-aware framework for adaptive portfolio allocation

Article References: Lim, S., Ahn, H., Lee, T., Kim, G., Lim, S., & Choi, I. (2026). Dynamic risk budgeting via large language models: A context-aware framework for adaptive portfolio allocation. Applied Intelligence, 56(15), Article 454. https://doi.org/10.1007/s10489-026-07501-w

Image Credits: AI Generated

DOI: 10.1007/s10489-026-07501-w

Keywords: large language models, risk budgeting, risk parity, portfolio optimization, GPT-4.1, regime detection, prompt engineering, quantitative finance, Hidden Markov Models, Sharpe ratio, dynamic asset allocation, Applied Intelligence

Liver biopsy is a cornerstone of cancer diagnosis, but it is far from trivial. The liver is a crowded, vascular organ hemmed in by ribs, lungs, bowel, and major blood vessels, and the needle must reach a target lesion while avoiding all of these hazards. Today, that planning rests almost entirely on the experience and judgment of individual interventional radiologists, which introduces meaningful variability between operators. Standardized AI-assisted planning has been proposed as a way to reduce that variability, particularly at centers with less experienced staff, but the field remains preliminary. The Memorial Sloan Kettering team set out to test the most basic building block of such a system: could a generative AI model, given nothing but a single axial CT slice and a standardized prompt, propose an entry angle that experienced reviewers would rate as potentially safe?

The study design was deliberately simple. The researchers retrospectively collected de-identified CT images from 30 consecutive CT-guided liver biopsies performed between February and June 2025 by sixteen board-certified interventional radiologists with a median of eight years of experience. The patient group included 20 women and 10 men with a median age of 56, and their lesions spanned liver segments 1 through 8 under the Couinaud classification. Crucially, the AI never touched a patient. It worked only on the initial planning CT image, after the fact, and only in two dimensions, while the actual procedures had already been completed safely by human hands.

The preprocessing workflow was strikingly low-tech for a study of cutting-edge AI. Each CT image was cropped to remove patient identifiers, magnified without rescaling, and overlaid with a radial grid created in GIMP, the free open-source image editor. The grid consisted of 36 radial lines at 10-degree intervals with concentric circles spaced 10 millimeters apart, calibrated to known anatomical landmarks. The annotated images were saved as PNG files and fed to GPT-4.1 through a standardized prompt. The model operated entirely on its baseline training, with no fine-tuning on medical data, and a fresh chat was started for every case to prevent contamination between cases.

The researchers tested two conditions that differed in how much hand-holding the AI received. In the first condition, two study authors manually annotated each image with a blue circle marking the lesion and a red line delineating sensitive structures to avoid, essentially handing the model a pre-drawn map of the danger zones. In the second, more demanding condition, only the blue circle was drawn, forcing the AI to identify critical structures on its own. For each case, the model was allowed up to four iterations: if its proposed trajectory failed safety criteria, it was re-prompted within that limit. Two independent reviewers then rated every AI-generated trajectory using consensus methodology, judging safety, avoidance of organs, vessels, and bone, whether the path crossed the midline or excessive tissue, and whether the approach would be comfortable and clinically reasonable.

The headline numbers were encouraging, with important caveats. In the first condition, the AI needed a median of two iterations and produced trajectories rated potentially safe in 27 of 30 cases, or 90 percent. In the second condition, without the red-line annotations, it needed a median of just one iteration and scored 28 of 30, or 93 percent. The difference between conditions was not statistically significant, and the model’s proposed entry angles differed significantly from the angles actually used by clinicians, with mean differences reaching statistical significance in both conditions. Only 23 percent of cases in condition one and 33 percent in condition two showed strong concordance, defined as entry angles within 10 degrees with similar liver traversal patterns, meaning the AI was often safe but not necessarily matching human choices.

The failures are as instructive as the successes. In three cases in the first condition, the AI proposed paths crossing bone, the midline, or the marked red line, even after four attempts. One of these involved a patient with ascites, fluid in the abdomen that complicated the anatomy; another involved a patient whose prior surgery had removed the left hepatic lobe, creating atypical anatomy; and a third targeted a high-lying segment 2 lesion where the model routed the needle through costal cartilage. Cartilage proved a recurring blind spot: in the second condition, both unsafe trajectories involved bone or cartilage, suggesting the model struggles to recognize structures with intermediate radiodensity. The authors note that excluding these inadequate cases from distance calculations introduces selection bias that may flatter the AI’s performance.

On the metric of liver tissue traversed, the AI actually looked competitive. In condition one, its paths crossed a median of 31.6 millimeters of liver versus 48.25 millimeters for the manual paths, and in condition two, 39.35 versus 48.25 millimeters, though neither difference reached statistical significance. Interestingly, the AI’s longer median distances in the unannotated condition hint at a conservative streak: without predefined avoidance zones, it may favor longer but safer routes. A power analysis indicated the 30-case sample could only detect large effect sizes, so smaller but clinically meaningful differences would require larger cohorts to detect.

The authors are refreshingly blunt about what these results do not mean. This was a proof of concept, nothing more. The workflow does not address a real clinical need: interventional radiologists can identify safe trajectories in seconds using native console capabilities with full three-dimensional visualization, whereas the AI pipeline required manual image export, GIMP annotation, and chatbot prompting, all while discarding the three-dimensional spatial information that real planning depends on. A human planner scrolls through multiple slices, uses multiplanar reconstruction, and mentally constructs 3D anatomy, capabilities entirely absent from a static 2D analysis. The AI also required continuous human supervision, with iterative re-prompting in most cases, which undermines any claim of autonomy. And because the AI trajectories were never actually executed, they remain theoretical, while the manual paths were proven safe by successful, complication-free procedures, a comparison that cannot establish equivalence.

Looking forward, the researchers see a path, if a long one. Future systems would need automated lesion identification, since current generative models lack the segmentation capabilities required for precise localization, and integration with dedicated segmentation architectures such as U-Net could enable fully autonomous planning from raw CT images. Three-dimensional volumetric analysis, quantitative vascular proximity measurements, phantom studies, and prospective multicenter trials would all be prerequisites before any clinical consideration. The authors also point toward applications in tumor ablation and margin assessment, and note that many liver biopsies worldwide are performed under ultrasound guidance, where real-time, operator-dependent imaging poses an even harder challenge for AI. For now, the message is measured: generative AI can look at a liver CT slice and propose a plausible needle angle most of the time, but the 7 to 10 percent of cases where it fails, sometimes by routing a needle through bone or cartilage, are exactly the cases where a patient would be harmed. In medicine, that margin is everything.

Subject of Research: Feasibility of generative AI for planning CT-guided liver biopsy needle trajectories

Article Title: Feasibility of generative AI for CT-guided liver biopsy trajectory planning: a pilot study

Article References: Feasibility of generative AI for CT-guided liver biopsy trajectory planning: a pilot study. (n.d.). https://doi.org/10.1007/s44343-026-00041-7

Image Credits: AI Generated

DOI: 10.1007/s44343-026-00041-7

Keywords: generative AI, GPT-4.1, liver biopsy, CT guidance, interventional radiology, trajectory planning, pilot study, medical imaging, needle biopsy, proof of concept, Memorial Sloan Kettering, artificial intelligence