All postsResearch skills

AI research topics for students: 30 open questions

APJ-RAISE team · · 15 min read

A good AI research topic for a student is a small, open question that you can test with public data and modest compute within one semester. "Large language models" is a field, not a topic. "Do small open models lose accuracy on maths word problems when only the numbers change?" is a topic you can finish.

A student writing equations on a large whiteboard, seen in silhouette
Photo: ThisIsEngineering, Pexels

The goal of this guide is practical: help you turn a topic into a question you can answer in about 16 weeks. It starts with one worked example of narrowing, then gives 30 open questions in machine learning, grouped by area. Each comes with a recent paper (2023 to 2026) showing why it is still open, the compute it needs and a public dataset or tool. We checked every paper and link on 18 November 2026.

Why most AI research topic lists do not help

Most lists you find online give areas, such as "explainable AI" or "federated learning in healthcare". An area is where a supervisor starts, not where a student should. It hides the three decisions that make a project finishable: what you change, what you measure, and on which data.

Lists also skip the cost. A topic that needs eight GPUs for a month is not a student topic, however interesting it sounds. So every question below states its compute level, and we counted them at the end.

How to narrow down a research topic: a worked example

Take one broad topic and narrow it in four passes. We use bias, because it is a common choice and often stays too vague.

  1. 1Area: bias in AI. Too wide; thousands of papers.
  2. 2Topic: caste bias in language models. Better, but you still cannot say when you are done.
  3. 3Question: do the caste stereotypes that Indic-Bias finds in English prompts appear more or less often when the same prompts are in Hindi?
  4. 4Testable plan: take a fixed sample of Indic-Bias templates, translate them, run two small open models in both languages and compare the rate of stereotyped answers.

The last version names one change (the language), one measure (stereotype rate) and the data (Indic-Bias). It also has a useful result either way. If bias drops in Hindi, that is a finding, and if it rises, that is one too. Our course on designing experiments covers how to make such a comparison fair.

Before you commit, run a quick check. Open the dataset link and download a small part on day one. Run the smallest model on free Colab or your laptop. Read one recent paper with the three-pass method. Then search arXiv and Google Scholar for your exact question; our list of AI tools for literature review helps. If you cannot write "I will measure X on dataset Y and compare it with Z", narrow again.

How long should a one-semester research question take?

We suggest this split for a 16-week semester. Weeks 1 to 3 go to reading and reproducing the cited paper's setup. Weeks 4 to 9 cover the one change you are testing. Weeks 10 to 12 are for repeats with different seeds and an error analysis. The last four weeks are for writing, with some slack for exams. If reproduction alone eats more than four weeks, the question is too big for this semester; cut a model or a language.

Compute levels in the list mean: laptop (any recent laptop, no GPU), free Colab (a free Google Colab notebook with a T4 GPU and run-time limits) and one GPU (a single 16 to 24 GB GPU for a few days, from your lab or a paid cloud). Each entry gives a short label, the question, why it is still open, the compute and the data.

LLM evaluation research topics

Evaluation questions are the easiest place to start because you do not train anything. You run models on a benchmark and study where they fail. Read our posts on LLM-as-a-judge and reasoning models first.

  • Number swaps: Do small open models lose accuracy on grade-school maths when only the names and numbers change? Open because: Apple's GSM-Symbolic paper (October 2024) found that every model it tested dropped when only the numbers changed, and one irrelevant extra sentence cut scores by up to 65%. Newer small models have not all been tested this way. Compute: free Colab with a 1B to 4B model. Data: GSM8K on Hugging Face.
  • Local knowledge: Are models weaker on state-level questions (local history, law, arts) than on science, and does that gap change by language? Open because: MILU from AI4Bharat (November 2024) covers 11 Indian languages and 41 subjects. The best model scored 74%, and models did worst in arts, humanities, law and governance. Compute: free Colab. Data: MILU benchmark.
  • Saying "not sure": If you score models with a penalty for wrong answers instead of a zero, which small models learn to say "I don't know"? Open because: OpenAI's "Why Language Models Hallucinate" (September 2025) argues that most benchmarks reward guessing. Few studies test this on small models. Compute: laptop with Ollama. Data: SimpleQA in OpenAI's simple-evals.
  • Ranking stability: Does the ranking of five small models change when you change only the prompt format or the few-shot examples? Open because: "The Leaderboard Illusion" (April 2025) showed how private testing skews Chatbot Arena, including 27 private variants tested by Meta before Llama 4. Small leaderboards have their own weak points. Compute: free Colab. Data: GSM8K or MILU, linked above.

Research questions on AI agents

An AI agent is a model that plans and calls tools in a loop. Agent research is newer, so simple measurements are still useful. Read what AI agents are and why multi-agent systems fail first.

  • Agent failures: Which failure types show up most when you build a two-agent coding team from small open models? Open because: the MAST paper (March 2025) found 14 failure modes in 3 groups after studying more than 1,600 traces from 7 frameworks. The study used models such as GPT-4, Claude 3, Qwen2.5 and CodeLlama. Compute: free Colab or a free API tier. Data: MAST taxonomy and traces.
  • Small tool users: How close does a 3B to 8B open model get to a big model on customer-service tasks with tools, and which tool calls break? Open because: an NVIDIA position paper (June 2025) argues that small models are enough for many agent steps, and invites critique. Compute: one GPU, plus API calls for the simulated user. Data: τ²-bench (the maintained version of τ-bench).

Mechanistic interpretability research topics

Mechanistic interpretability tries to find the parts inside a network that do each job. Free tools now let you study real models in a notebook. Read our mechanistic interpretability guide first.

  • Hindi features: Does Gemma 2 2B have sparse autoencoder (SAE) features that fire only for Hindi or Tamil text, and does switching them off hurt only that language? Open because: Deng and colleagues (May 2025, ACL 2025) found language-specific SAE features whose removal hurt mainly one language. You can test whether this holds for Indian languages. Compute: free Colab. Data: Gemma Scope SAEs and parallel sentences from FLORES+.
  • Seed stability: If you train the same SAE twice with different random seeds, which features survive? Open because: Paulo and Belrose (January 2025) found that only 30% of features were shared across seeds in one SAE on Llama 3 8B. Nobody knows yet which kinds of features are stable. Compute: one GPU for a small model. Tool: EleutherAI sparsify.
  • Truth direction: Does the "truth direction" found in English true or false statements still work when you translate the statements into Hindi? Open because: "The Geometry of Truth" (October 2023) found a simple linear direction for truth in model activations, using English datasets. Compute: free Colab. Data: geometry-of-truth datasets.

Small and efficient model research topics

Efficiency research asks how much you lose when you make a model smaller or cheaper. Much of it runs on a laptop. Read our post on small language models first.

  • Quantisation cost: How much does 4-bit quantisation hurt each Indian language, compared with the full model, on the same questions? Open because: a Cohere study (July 2024) found that non-Latin scripts were hit worst. A 1.7% drop in Japanese on automatic tests matched a 16.0% drop in human ratings. Compute: laptop with llama.cpp. Data: MILU, linked above. Tool: llama.cpp.
  • Tokenizer tax: How many tokens per word do current open-model tokenizers need for your language, and for code-mixed text? Open because: a July 2026 paper measured an average 8.0 times "tokenizer tax" for Indian languages under GPT-4's cl100k_base tokenizer, and 13.0 times for Malayalam. The MUTANT tokenizer (November 2025) shows the gap can shrink. Dialects and Hinglish are barely measured. Compute: laptop, no GPU. Data: FLORES+.
  • LoRA forgetting: After LoRA fine-tuning a 1B model on a Hindi task, how much English and maths skill does it lose, and does the LoRA rank change that? Open because: "LoRA Learns Less and Forgets Less" (May 2024) studied code and maths, not languages. Compute: free Colab with Unsloth. Data: GSM8K to measure forgetting, plus your Hindi task data.
  • 1-bit models: Does a 1.58-bit model keep up with a 4-bit quantised model of similar size on Indian-language questions? Open because: Microsoft's BitNet b1.58 2B4T report (April 2025) claims it matches full-precision models of similar size, but Indian languages are an open test. Compute: laptop CPU. Tool: bitnet.cpp. Data: MILU.

AI safety and security research topics

Safety research tests whether a model refuses harmful requests and resists attacks. Read our guides to AI guardrails and the OWASP Top 10 for LLMs first. Run these tests only on models you host or are allowed to test, and do not publish working attack prompts.

  • Language gap: Do small open models refuse harmful requests less often in Hindi, Bengali or Hinglish than in English? Open because: Deng and colleagues (October 2023, ICLR 2024) found that low-resource languages were about three times as likely to get harmful replies from ChatGPT and GPT-4. Newer small models and code-mixed text are less studied. Compute: free Colab. Data: JBB-Behaviors.
  • Injection defence: Which defence pattern cuts prompt-injection attacks most on agent tasks, and how much useful work does it cost? Open because: AgentDojo (June 2024) has 97 tasks and 629 security test cases. A June 2025 paper proposed design patterns with provable resistance, but few compare them on the same tasks. Compute: one GPU or API credits. Data: AgentDojo.

Computer vision research topics

Vision research for students is strongest where Indian images differ from the Western images most models learned from. Start by reading one benchmark paper below with the three-pass method.

  • Cultural VQA: Do vision-language models answer the same question about the same Indian image worse in Odia or Marathi than in English? Open because: IndicVisionBench (November 2025) has about 5,000 images and 37,000+ question-answer pairs in English and 10 Indian languages, with parallel answers, and found large gaps. Compute: free Colab with a small open model. Data: IndicVisionBench.
  • Step by step: Does asking a vision-language model to reason step by step help or hurt on Indian cultural questions in low-resource languages? Open because: DRISHTIKON (September 2025) covers 15 languages and 64,000+ image-text pairs, tests chain-of-thought prompts, and found models weakest on low-resource languages. Compute: free Colab. Data: see the DRISHTIKON paper.
  • Indian roads: How much does a segmentation model trained on European city streets lose on Indian roads, and which classes fail? Open because: the India Driving Dataset (IDD) has 10,004 images with 34 classes from 182 drives, and models scored much lower on it than on Cityscapes. Newer models trained on huge image sets have not closed the question. Compute: one GPU. Data: IDD from IIIT Hyderabad.

Speech and Indian language research topics

India has 22 scheduled languages and many dialects, so speech and translation work here is both needed and publishable. Read our post on Sarvam 105B for the state of Indian models.

  • Rural voices: How badly do open speech recognition (ASR) models do on rural speakers of your language, and does synthetic speech help? Open because: the SRUTI benchmark (June 2025) showed poor results for rural Bhojpuri women. Synthetic speech built from 25 to 30 seconds per speaker cut word error rate by 4.7 points. Other languages are untested. Compute: free Colab. Data: IndicVoices.
  • Fair scoring: How much of the measured error in Indian-language ASR is only spelling variation? Open because: a March 2026 AI4Bharat paper proposed a spelling-aware error rate that lowered error rates by 6.3 points on average and shrank one model gap from 18.1 to 11.5 points. Compute: laptop, since you only score outputs. Data: Lahaja Hindi accent benchmark.
  • Indian English: Which Indian English accents do open ASR models find hardest? Open because: Svarah (May 2023) has 9.6 hours of English from 117 speakers in 65 places across India and showed clear room to improve. Newer models have not been compared by region. Compute: free Colab. Data: Svarah.
  • Rare languages: For Santali, Bodo or Manipuri, does IndicTrans2 keep its quality on everyday text such as notices or chat messages? Open because: IndicTrans2 (May 2023) was the first model to cover all 22 scheduled languages, and its IN22 benchmark covers diverse domains. How it holds up on local notices and chat in low-resource languages is open. Compute: free Colab. Data: IN22-Gen.

AI for science research topics

Physics-informed neural networks (PINNs) train a network to obey an equation. They fail in known ways, which makes them good for careful student studies. Read our PINNs guide first.

  • PINN rescue: For the convection equation, which fix works best under the same time budget: curriculum training, sequence-to-sequence training or switching optimisers? Open because: Krishnapriyan and colleagues (2021) showed PINNs fail on convection and their fixes cut error by 10 to 100 times. A 2024 study by Pratik Rathore and colleagues found Adam followed by L-BFGS trains better than Adam alone. No one fixed budget compares them all. Compute: laptop CPU. Tool: DeepXDE.
  • Operators vs PINNs: On simple 1D problems, when does a neural operator (FNO) beat a PINN for the same training time? Open because: PDEBench (2022) gives FNO, U-Net and PINN baselines with data, but fair time-matched comparisons are rare. Compute: one GPU. Data: PDEBench.

AI research topics in health, farming and climate

Applied topics need care: report what the model does on real data, and do not claim it is ready for patients or farmers. Use public Indian data where you can.

  • Monsoon onset: How well do AI weather models predict monsoon onset over your state, compared with a simple average of past years? Open because: in 2025, AI-based onset forecasts reached 38 million Indian farmers by SMS, in a project with India's agriculture ministry, according to Google. The forecast blend included Google's NeuralGCM and ECMWF's AIFS. Independent checks by state are still few. Compute: free Colab. Data: IMD gridded rainfall (0.25 degree, 1901 to 2024), read with imdlib, and forecasts from WeatherBench 2.
  • Eye screening: Does a diabetic retinopathy model trained on the APTOS 2019 retina photos still work on IDRiD photos taken with different cameras? Open because: medical image models often drop when the camera or clinic changes. IDRiD has 516 graded retina images and calls itself the first set representative of an Indian population. Compute: free Colab. Data: IDRiD (free IEEE account) and the APTOS 2019 Kaggle set. This is research only, not diagnosis.

Fairness and bias research topics for India

Most bias benchmarks were built for the US and Europe. India-specific bias, such as caste, religion and region, is a field where Indian students have an advantage.

  • Bias by language: Do the caste and religion stereotypes found by Indic-Bias in English prompts appear more or less often in Hindi or Tamil? Open because: Indic-Bias (June 2025) covers 85 identity groups with 20,000 checked templates and found strong negative bias across 14 models. Its templates are in English. Compute: free Colab. Data: Indic-Bias.
  • Hidden bias: Does a model that refuses direct caste questions still show bias in indirect tasks, such as writing a story or choosing a job candidate? Open because: DECASTE (May 2025) measured both direct and indirect caste bias across social, economic, educational and political areas, and found models reinforce caste bias. Compute: free Colab. Data: Indic-Bias, linked above, which includes judgment and generation tasks.

Reinforcement learning research topics

Reinforcement learning (RL) trains a model with rewards. Since DeepSeek-R1 (January 2025), RL on maths and code has been a hot topic, and fast simulators now run on one GPU. Read our post on reasoning models first.

  • RL vs base: Does RL training on a 0.5B model teach new maths skills, or only make good answers more likely? Open because: Yue and colleagues (April 2025) found RL-trained models win when you sample one answer, but base models win when you sample many. Tests on very small models are still few. Compute: free Colab for a 0.5B model, slowly. Tool: TRL. Data: GSM8K.
  • Exploration: Which exploration method gets furthest in Craftax, a fast open-ended game benchmark? Open because: the Craftax paper (February 2024, ICML 2024) runs 1 billion steps of PPO in under an hour on one GPU in its simpler mode, and reports that existing exploration methods fail to make real progress on the full game. Compute: one GPU. Environment: Craftax.

Can you do AI research on free Colab?

Mostly yes. We counted the 30 questions above by compute level: 6 run on a laptop, 18 fit in a free Colab notebook and 6 need one GPU for a few days. So 24 of 30, or 80%, need no paid hardware at all.

Free Colab sessions end without warning, so plan for it. Test your code on 50 examples before the full run. Save results after every batch, so a timeout costs minutes rather than hours. Use the smallest model that shows the effect, and confirm on one larger model only if time allows.

From research question to paper

Treat the first month as a reproduction. Run the cited paper's setup and get close to its numbers. Then make your one change and measure the difference. That change is your contribution.

Report the details others need to rerun your work: model names and versions, prompts, random seeds and the number of examples. Share your code. Our guide on how to start research as a student covers the full path. If you need a build project for your final year instead, see our final year project ideas for AIML.

At APJ-RAISE, students pick questions like these in small teams and publish their results on our Open Research page. Joining is free for students aged 18 or older.

Questions people ask

What are good AI research topics for college students?
Good topics are narrow, open questions with public data, such as testing whether 4-bit quantisation hurts Indian languages more than English, or whether caste bias changes when prompts are in Hindi. Each topic in this guide comes with a recent paper and a free dataset.
What is the difference between a research topic and a research question?
A topic is an area, such as "bias in language models". A research question names what you change, what you measure and the data, such as "Does caste bias in Indic-Bias change when the prompts are in Hindi instead of English?"
What are good AI research topics for beginners?
Evaluation is the easiest start, because you run existing models on a benchmark and study where they fail. You need no training budget, and careful evaluation papers are useful to the field.

Sources

  1. 1.Mirzadeh et al. GSM-Symbolic: Understanding the Limitations of Mathematical Reasoning in Large Language Models. arXiv, October 2024.
  2. 2.Verma et al. MILU: A Multi-task Indic Language Understanding Benchmark. arXiv, November 2024.
  3. 3.Kalai et al. Why Language Models Hallucinate. arXiv, September 2025.
  4. 4.Wei et al. Measuring short-form factuality in large language models (SimpleQA). arXiv, November 2024.
  5. 5.Singh et al. The Leaderboard Illusion. arXiv, April 2025.
  6. 6.Cemri et al. Why Do Multi-Agent LLM Systems Fail? arXiv, March 2025.
  7. 7.Belcak et al. Small Language Models are the Future of Agentic AI. arXiv, June 2025.
  8. 8.Yao et al. τ-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains. arXiv, June 2024.
  9. 9.Deng et al. Unveiling Language-Specific Features in Large Language Models via Sparse Autoencoders. arXiv, May 2025.
  10. 10.Lieberum et al. Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 2. arXiv, August 2024.
  11. 11.Paulo and Belrose. Sparse Autoencoders Trained on the Same Data Learn Different Features. arXiv, January 2025.
  12. 12.Marks and Tegmark. The Geometry of Truth: Emergent Linear Structure in Large Language Model Representations of True/False Datasets. arXiv, October 2023.
  13. 13.Marchisio et al. How Does Quantization Affect Multilingual LLMs? arXiv, July 2024.
  14. 14.Srivastava. The Tokenizer Tax: Quantifying and Explaining the Cross-Lingual Cost of Subword Tokenization for Indian Languages. arXiv, July 2026.
  15. 15.Rana et al. MUTANT: A Recipe for Multilingual Tokenizer Design. arXiv, November 2025.
  16. 16.Biderman et al. LoRA Learns Less and Forgets Less. arXiv, May 2024.
  17. 17.Ma et al. BitNet b1.58 2B4T Technical Report. arXiv, April 2025.
  18. 18.Deng et al. Multilingual Jailbreak Challenges in Large Language Models. arXiv, October 2023.
  19. 19.Debenedetti et al. AgentDojo: A Dynamic Environment to Evaluate Prompt Injection Attacks and Defenses for LLM Agents. arXiv, June 2024.
  20. 20.Beurer-Kellner et al. Design Patterns for Securing LLM Agents against Prompt Injections. arXiv, June 2025.
  21. 21.Faraz et al. IndicVisionBench: Benchmarking Cultural and Multilingual Understanding in VLMs. arXiv, November 2025.
  22. 22.Maji et al. DRISHTIKON: A Multimodal Multilingual Benchmark for Testing Language Models' Understanding on Indian Culture. arXiv, September 2025.
  23. 23.Varma et al. IDD: A Dataset for Exploring Problems of Autonomous Navigation in Unconstrained Environments. arXiv, November 2018.
  24. 24.Joshi et al. Recognizing Every Voice: Towards Inclusive ASR for Rural Bhojpuri Women. arXiv, June 2025.
  25. 25.Bhogale et al. Towards Orthographically-Informed Evaluation of Speech Recognition Systems for Indian Languages. arXiv, March 2026.
  26. 26.Javed et al. Svarah: Evaluating English ASR Systems on Indian Accents. arXiv, May 2023.
  27. 27.Gala et al. IndicTrans2: Towards High-Quality and Accessible Machine Translation Models for all 22 Scheduled Indian Languages. arXiv, May 2023.
  28. 28.Krishnapriyan et al. Characterizing possible failure modes in physics-informed neural networks. arXiv, September 2021.
  29. 29.Takamoto et al. PDEBench: An Extensive Benchmark for Scientific Machine Learning. arXiv, October 2022.
  30. 30.Google. How AI is helping 38 million farmers with advance weather predictions. 15 September 2025.
  31. 31.Rasp et al. WeatherBench 2: A benchmark for the next generation of data-driven global weather models. arXiv, August 2023.
  32. 32.Nawale et al. FairI Tales: Evaluation of Fairness in Indian Contexts with a Focus on Bias and Stereotypes. arXiv, June 2025.
  33. 33.Vijayaraghavan et al. DECASTE: Unveiling Caste Stereotypes in Large Language Models through Multi-Dimensional Bias Analysis. arXiv, May 2025.
  34. 34.DeepSeek-AI. DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning. arXiv, January 2025.
  35. 35.Yue et al. Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model? arXiv, April 2025.
  36. 36.Matthews et al. Craftax: A Lightning-Fast Benchmark for Open-Ended Reinforcement Learning. arXiv, February 2024.
All posts