Travel Grant to Present Research on Robustness of LLM-Guided Optimisation at NeurIPS 2026

Berkeley, USA

papers.ssrn.com

Supporting research on when LLM-guided optimisation systems fail under changing specifications, constraints, and environments.

Project Details

I am requesting $3,000 toward the cost of attending NeurIPS 2026 in Sydney, where my paper, “Hyperparameter Optimisation of Convex Portfolio Trajectories using Large Language Models,” has been accepted for presentation at the EvoRobust workshop.

The estimated expenses are approximately:

$1,500-$1,800 international airfare
$900-$1,100 accommodation
$300-$450 conference registration and workshop participation
$200-$300 local transportation and other travel expenses

The requested $3,000 would therefore cover a substantial portion of the costs required to attend the conference in person.

The purpose of attending extends beyond presenting the accepted paper. I want to use the conference to obtain technical feedback on the proposed robustness and evaluation extension, engage with researchers working on LLM evaluations and AI safety, and identify the most valuable experiments for the next stage of the research.

Theory of Impact

As increasingly capable AI systems are used for reasoning, planning, optimisation, and tool-assisted decision-making, it becomes increasingly important to understand whether they remain reliable when the conditions surrounding their deployment change.

Many evaluations focus on performance under a fixed specification. However, a system may perform well against its nominal objective while behaving unexpectedly when the specification changes, constraints are modified, inputs shift, or the environment differs from the one in which it was evaluated. These failure modes could become safety-relevant when AI systems are embedded in consequential decision-making processes.

My work contributes to understanding these risks by providing a controlled setting in which LLM-guided optimisation can be systematically stress-tested. Because the optimisation objectives, constraints, and performance metrics are explicit, the framework allows failure modes such as specification sensitivity, constraint violations, distributional fragility, and stochastic instability to be measured quantitatively.

The potential path to reducing AI risk is through better evaluation methodology. By identifying which forms of LLM-guided behaviour are robust and which fail under seemingly small changes, this research can contribute to methods for testing AI systems before deploying them in more consequential environments.

People

Sayash Raaj

Team Member

Funding Details

Start Date: Mar 1, 2026
End Date: -
Expected Duration: 8 months
Funding Raised to Date: $0
Annual Budget: $4,180
Monthly Burn Rate: $348
Current Runway: 0 months
Funding Goal: -
Funding Stage: Seeking first dedicated funding for the NeurIPS 2026 research/travel stage
Fiscal Sponsor: -

Track Record

I am a UC Berkeley Master of Financial Engineering student with a background in mathematical optimisation, quantitative finance, machine learning, and statistical modelling. I graduated from IIT Madras with Department Rank 2/115 and have worked in quantitative research and portfolio engineering at Fidelity Investments and BlackRock.

My independent research has been accepted at NeurIPS 2025, ICML 2026, and NeurIPS 2026, and I have presented research at ICML in person. My technical work spans convex and linear optimisation, Gurobi, Monte Carlo simulation, statistical modelling, portfolio construction, and machine learning.

I have also previously received a $2,000 BlueDot Impact Rapid Grant supporting my transition toward technical AI safety research and successfully completed the objectives of that grant, including presenting my accepted ICML 2026 workshop research.

Grants Received– no grants recorded

Add grant

Funding Asks

Robustness Evaluation of LLM-Guided Optimisation

Open

Evaluating whether LLM-guided optimisation systems remain reliable when specifications, constraints, and environments change.

Minimum

$1,000

Ideal

$3,000

What this ask is

I am developing a research program on the reliability and robustness of LLM-guided optimisation systems.

The immediate foundation is my paper, “Hyperparameter Optimisation of Convex Portfolio Trajectories using Large Language Models,” which has been accepted for presentation at the EvoRobust workshop at NeurIPS 2026. The work studies how LLMs can guide hyperparameter optimisation in constrained, multi-period convex optimisation.

I am now extending this work from asking whether LLM-guided optimisation can improve an objective to systematically evaluating when and why such systems fail. I plan to study sensitivity to specification changes, constraint perturbations, distribution shift, stochastic instability across repeated runs, and pathological or adversarial inputs.

I am currently the sole researcher on the project. My background combines mathematical optimisation, machine learning, quantitative research, and statistical modelling. I graduated from IIT Madras with Department Rank 2/115 and have worked in quantitative research and portfolio engineering at Fidelity Investments and BlackRock. My research has previously been accepted at NeurIPS 2025 and ICML 2026, and I have previously presented research at ICML in person.

The project is currently at the stage of an accepted research result with a working experimental foundation and a defined research direction for the next iteration.

Theory of impact

As increasingly capable AI systems are used for planning, optimisation, tool use, and consequential decision-making, failures that are not captured by standard performance benchmarks may become increasingly important.

This project contributes to technical AI safety by developing methods for identifying such failures in a controlled setting. LLMs are used as components of an optimisation pipeline with explicit objectives and constraints, making it possible to quantitatively measure behaviours such as specification sensitivity, constraint violations, distributional fragility, and instability across repeated runs.

The causal chain is:

Develop a controlled framework for testing LLM-guided optimisation.
Systematically vary specifications, constraints, inputs, and environments to expose failure modes.
Develop quantitative metrics and evaluation procedures for these failures.
Share the resulting methodology and findings with researchers working on AI evaluations, robustness, and reliable AI systems.
Improve our ability to detect problematic behaviour in AI systems before similar capabilities are deployed in more consequential settings.

The portfolio-optimisation environment is primarily a testbed: the broader safety-relevant contribution is better methodology for evaluating AI systems whose outputs influence structured decision processes.

The immediate research is therefore not intended to solve existential risk directly. Its potential value comes from contributing to the broader technical toolkit for evaluating the reliability and robustness of increasingly capable AI systems.

How the money will be spent

The $3,000 funding goal will support in-person participation at NeurIPS 2026 in Sydney, where the underlying research has been accepted for presentation at the EvoRobust workshop.

The funding will primarily cover:

Airfare: international round-trip travel to Sydney
Accommodation: lodging during the conference period
Conference costs: NeurIPS registration and workshop participation
Local transportation: transportation within Sydney during the trip

The purpose of the travel extends beyond presenting the accepted paper. I will use the conference and surrounding AI-safety community to obtain technical feedback on the proposed robustness and evaluation work, discuss methodology with researchers working on LLM evaluations and optimisation, and identify the most valuable experiments for the next stage of the project.

Discussion

No comments yet. Be the first to share your thoughts.