Current alignment techniques, such as RLHF (Reinforcement Learning from Human Feedback), typically rely on a centralized set of "safety guidelines" that reflect the cultural norms of the model's developers. This often leads to cultural misalignment, where the LLM may inadvertently erase local nuances, apply inappropriate social taboos, or fail to recognize regional linguistic etiquette.
This thesis aims to move beyond "cultural awareness" toward Cultural Verification. The student will design a systematic framework to evaluate and verify whether an LLM’s outputs adhere to the specific socio-cultural norms, value systems, and historical contexts of a target demographic.

Key Responsibilities & Tasks:

The research will be structured around four primary pillars, beginning with the development of a comprehensive Taxonomy of Cultural Correctness to define the specific dimensions—such as social hierarchy, religious sensitivities, and gender roles—necessary for nuanced LLM evaluation. Building upon this foundation, the study will propose a Verification Methodology centered on a "Cultural Guardrail" architecture, utilizing both programmatic checks and human-in-the-loop oversight to validate model responses. To test the efficacy of this framework, a specialized Benchmark will be curated, featuring a localized "red-teaming" suite designed to expose cultural hallucinations and latent insensitivities. Finally, a Comparative Analysis will be conducted to measure the performance of leading models, including GPT-4, Llama 3, and Claude, against these proposed designs across two or more distinct cultural contexts.

This thesis investigates the intersection of federated learning and parameter-efficient fine-tuning of Large Language Models (LLMs), with a focus on Low-Rank Adaptation (LoRA) techniques. The goal is to develop and evaluate strategies for fine-tuning LLMs across decentralized clients, each using their own local LoRA modules tailored to their data distributions.

## The central research questions are: ##
1. How can clients collaboratively fine-tune a shared LLM backbone using local, heterogeneous LoRA modules without sharing raw data?
2. How can we effectively aggregate model updates in the presence of statistical heterogeneity across clients?
3. What are the privacy, communication, and performance trade-offs of federated LoRA-based fine-tuning?

The thesis will build upon recent work that shows it is possible to train LLMs in federated environments by exchanging only adapter weights. This project will extend that idea by exploring heterogeneous LoRA configurations—i.e., each client may choose its own rank, structure, or optimization setup—and study how such diversity affects convergence, model utility, and communication efficiency.

## Key components of the thesis: ##
1. Implementation of federated fine-tuning protocols using LoRA and related adapter methods (e.g., QLoRA, AdaLoRA).
2. Support for heterogeneous client-side LoRA configurations (e.g., different ranks, tasks, or target layers).
3. Evaluation on benchmark NLP tasks (e.g., sentiment analysis, question answering) using common federated learning setups (e.g., LEAF, FedNLP).
4. Investigation of aggregation strategies that accommodate non-iid client updates (e.g., FedAvg with normalization, FedProto, FedAdam).
5. Optional exploration of privacy-preserving extensions (e.g., differential privacy, secure aggregation).

## Expected outcomes: ##
1. A practical system that enables federated, privacy-conscious, and communication-efficient fine-tuning of LLMs.
2. Empirical insights into the effects of heterogeneity in LoRA configurations across clients.
3. A set of recommendations for scalable and modular federated LLM training pipelines.

The tactile interpretation of images requires the transformation of two-dimensional visual information into three-dimensional representations accessible through touch. This conversion presents significant challenges, particularly in the preservation of crucial expressive elements that in visual reality are highlighted through light and shadow.

The project focuses on case studies of pictorial art; for example, works by Caravaggio, where chiaroscuro creates a true 'tactile map' through:
- Anatomies highlighted by shadows (muscles, ribs, veins)
- Movement-defining fabric folds
- Depth of environments
- Textures of objects (fruit, fabrics, metals)

The objective of this thesis is to develop an AI system that overcomes current limitations in the automatic conversion of images into tactile representations, focusing on:
- Recognition of elements that in visual reality are highlighted by light and shadow
- Translation of these elements into appropriate depth maps
- Preservation of key narrative elements
- Optimisation of high relief rendering
The project is being developed in parallel with tactile image production experts who will test the effectiveness of the developed solutions.

The project offers the opportunity to:
- Working on a concrete problem of cultural accessibility
- Developing an innovative image interpretation system
- Collaborating with experts in the field of tactile representations
- Contributing to innovation in art accessibility