AI and Machine Learning Research Paper Topics

What makes an AI topic workable for a student

Competing with large labs on model size is not realistic, and it is not necessary. Student-scale AI research works best when you evaluate, adapt or constrain existing models: testing them on under-served languages, measuring how they fail, shrinking them for cheap devices or improving a pipeline around them. All of the ideas below can start with pretrained models and free notebook GPUs, although you should confirm the compute needed before committing.

Large language models and retrieval

1. Chunking strategies for RAG over technical documents — LangChain, pgvector, sentence-transformers
   Retrieval quality depends heavily on how documents are split. Compare fixed, semantic and structure-aware chunking on manuals containing tables, and score retrieval and answer accuracy separately.
2. Detecting unsupported claims in generated answers — NLI models, Python, Hugging Face
   RAG systems can still state things their sources do not support. Build a checker that verifies each claim against retrieved passages and test it on labelled answers.
3. Hybrid search for code-mixed Hinglish queries — BM25, multilingual embeddings, Elasticsearch
   Users mix scripts and languages, which can confuse embeddings. Evaluate keyword, dense and hybrid retrieval on Hinglish questions, using a test set you annotate yourself.
4. Prompt injection defences for document assistants — Open-weight LLMs, Python
   Instructions hidden in uploaded files can hijack an assistant. Create an attack suite, test defences such as input isolation and output filtering, and report success rates and lost usefulness.
5. Small language models for domain-specific tasks — PyTorch, Transformers, llama.cpp
   A fine-tuned small model may match a large one on a narrow task at far lower cost. Compare them on a defined task, covering accuracy, latency and memory.
6. Measuring bias in LLM responses for Indian contexts — Open-weight LLMs, statistical testing
   Many fairness benchmarks reflect Western social categories. Design test prompts around Indian names, regions and occupations, then measure how outputs differ, with an annotation protocol others could repeat.

AI agents and automation

1. Evaluating reliability of multi-step AI agents — LangGraph, Python
   A small error early in a workflow can compound across steps. Define repeatable tasks, record where agents fail, and compare single-agent and multi-agent designs on completion and cost.
2. Safe tool use with human approval checkpoints — Python, function calling, n8n
   Agents that send emails or change records need limits. Test approval policies by action risk and measure the balance between prevented mistakes and interruptions to the user.
3. Memory design for long-running assistants — Vector databases, Python
   Keeping everything raises cost and can bring in stale facts. Compare summary, vector and structured memory on multi-session conversations, using questions that test recall and outdated information.

Edge AI, efficiency and privacy

1. Quantisation and pruning on low-cost boards — TensorFlow Lite, ONNX Runtime
   Compression can lower accuracy unevenly across classes. Measure this for a vision model on a Raspberry Pi or similar board, reporting per-class impact as well as overall figures.
2. Federated learning with uneven client data — Flower, PyTorch
   Real clients rarely hold balanced data, which hurts the shared model. Compare aggregation methods under controlled imbalance and report both accuracy and communication overhead.
3. Machine unlearning for removal requests — PyTorch, membership inference tests
   Deleting a record from a dataset does not remove its influence on a trained model. Compare approximate unlearning with full retraining on removal quality and time saved.
4. Energy use of training versus inference choices — CodeCarbon, PyTorch
   Model size is only part of the energy picture. Log energy for different architectures and batch sizes, then explore which settings save most for an acceptable accuracy loss.

Vision, speech and Indian-language AI

1. Crop disease models that work in the field — PyTorch, EfficientNet, ViT
   Models trained on clean lab-style leaf images often struggle on farm photographs. Measure this drop on images you collect, then test augmentation and domain adaptation to reduce it.
2. Medical image classification with limited labels — PyTorch, MONAI
   Expert labels are slow and expensive. Compare self-supervised pretraining with standard transfer learning at different label counts on a public dataset, with attention to patient-level data splits.
3. Speech recognition for Marathi dialects — Whisper, Hugging Face
   Accuracy can differ sharply between standard speech and regional variation. Evaluate pretrained models on recordings grouped by region, then test whether small-scale fine-tuning narrows the gap.
4. OCR for handwritten Devanagari forms — OpenCV, TrOCR, PaddleOCR
   Printed text works far better than handwriting in structured forms. Build a pipeline for field detection and recognition, and score it at field level, since one wrong digit matters.
5. Detecting AI-generated images after compression — PyTorch, diffusion model outputs
   Detectors can weaken after images are resized and shared on messaging apps. Measure the fall under repeated compression, then test training methods that make detection more robust.

Trustworthy and applied machine learning

1. Detecting concept drift in deployed models — River, Evidently, Python
   Performance decays quietly as real-world data changes. Compare drift detectors on a time-ordered dataset, measuring detection delay and false alarms, and test when retraining actually helps.
2. Are explanations stable enough to trust? — SHAP, LIME, XGBoost
   SHAP and LIME can give different reasons after small input changes. Quantify the stability of their explanations for a credit or healthcare tabular model and discuss what that means in practice.
3. Gradient boosting versus deep learning on small tabular data — XGBoost, TabNet, Optuna
   Deep models are not always the best choice for tables. Run a fair comparison with proper tuning across several small datasets from Indian public data portals.
4. Synthetic data for rare-event detection — SDV, CTGAN, scikit-learn
   Fraud and equipment failures supply few positive examples. Compare synthetic data generators with simple resampling, always evaluating on real held-out data rather than generated records.

What reviewers and examiners look for

Many student papers fail for reasons unrelated to the idea. Accuracy above 99 percent on a random split, for instance, is more likely to prompt questions about data leakage than praise. Your experimental design deserves as much attention as your model.

• A baseline: compare with a simple method as well as the latest architecture.
• Clean splits: separate by patient, user or time wherever the same source could appear on both sides.
• Suitable metrics: use F1, AUC or field-level accuracy where classes are imbalanced, and report per-class results.
• Repeatability: report results over several random seeds, and record versions and hyperparameters.
• Honest limits: state where your method fails and what your dataset does not represent.
• Disclosure: follow the target venue policy on AI writing tools and confirm that every citation is real, since LLMs can invent references.

Mentoring for AI research work

Road2Tech in Pune mentors students on topic scoping, reference papers, algorithm details, testing reports, runnable source code with explanation and IEEE paper support. Publication decisions rest with journals and conferences, so the focus stays on sound work that you understand and can defend.