Arnesh Batra
I study what neural networks know, where they keep it, and whether we can check.
I'm a final-year B.Tech student in Computer Science & AI at IIIT Delhi. I work on mechanistic interpretability, representation learning and AI safety: reading what is stored inside a model's layers, testing whether it is actually used, and building simple tools that catch models when they go wrong.
Right now I'm a research fellow at SPAR and a research scholar at MARS V (Cambridge AI Safety Hub) with Yossi Gandelsman, where I study whether LLMs can report on their own computation. Before this I worked with Randall Balestriero (Brown / Meta FAIR) and Anubha Gupta (SBI Lab). I also founded StarkVision, a student research group of 20+ people.
Now · Fall 2026
- Research fellow at SPAR.
- MARS V scholar: probing and steering LLMs for self-knowledge about their own behaviour.
- AI-safety research with Shiba AI on interpretability for agentic and multi-agent systems.
- What Do Audio Models Really Hear? accepted at NeurIPS 2026.
I'm looking for pre-doctoral and PhD positions starting 2027. If you think we'd work well together, please email me.
arnesh23129@iiitd.ac.inResearch
The questions I'm working on right now.
Mechanistic interpretability: better tools, better interpretations
Methods for finding what a model computes, and checks that what we read off the activations is actually what the model uses. Probing, steering, causal interventions and residual-stream analysis.
Shiba AI · Whisper RSC (ICML-W '26)Model psychology: introspection and chain of thought
Can a model know things about its own computation? When does its chain of thought reflect what it is really doing, and when does it not?
MARS V · SPARRepresentation learning and geometry
How information is arranged across layers and how that shape relates to what a model can do. Self-supervised and JEPA-style objectives.
ICML '26 Spotlight · UniReps '25 · Lens-JEPAInterpretability beyond language models
Applying the same tools to audio, vision and world models, and seeing what carries over.
NeurIPS '26 · Whisper RSC
-
NeurIPS 2026
What Do Audio Models Really Hear? Layer Selection and Mechanistic Structure in Sound Representations
Neural Information Processing SystemsWhich layers of audio models hold the useful information, and what internal structure sound representations actually have.
paper coming soon -
ICML 2026Spotlight
Uncovering the Latent Potential of Deep Intermediate Representations
International Conference on Machine LearningTask information is spread non-monotonically across depth. Intermediate layers often hold more than the final one, and it can be recovered with the right readout.
-
EMNLP 2026Industry
Catching AI-Generated Code at Scale: A Lightweight and Interpretable Detector for the LLM Era
EMNLP Industry TrackA small, interpretable detector for LLM-written code that is cheap enough to run at scale.
-
TMLR
TOAM-YOLO: A Tiny Object-Aware Multi-Expert YOLO Framework for Diverse Domains
Transactions on Machine Learning ResearchA multi-expert YOLO built to find very small objects across different domains.
-
ICML 2026 Workshop
Residual Stream Contrast: A Training-Free Counterfactual Listening Test for Whisper Hallucinations
ICML Workshop on Machine Learning for AudioFlag Whisper hallucinations without training anything, by contrasting the residual stream on real audio against a counterfactual input.
-
TMLR
Melody or Machine: Detecting Synthetic Music with Dual-Stream Contrastive Learning
Transactions on Machine Learning ResearchA benchmark and a dual-stream contrastive model for detecting AI-generated music.
-
NeurIPS 2025 Workshop
Relational Representation Learning
UniReps: Unifying Representations in Neural ModelsRepresentations that capture the relations between samples, not just each sample on its own.
-
CVPR 2025 Demo
M-SCAN: A Multistage Framework for Lumbar Spinal Canal Stenosis Grading Using Multi-View Cross Attention
CVPR Demonstrations TrackMulti-view cross-attention over spinal MRI to grade lumbar stenosis.
-
NeurIPS 2025 Workshop
CLAM: Safeguarding Authenticity and Addressing Implications for the Music Industry
AI for Music WorkshopDetecting AI-generated music, and what it means for artists and the industry.
-
NeurIPS 2025 Workshop
DentalNet: Geometric Aware Multi-View Transformer for Occlusion Grade Prediction in Dental 3D Scans
Imageomics WorkshopA geometry-aware multi-view transformer that grades dental occlusion from 3D scans.
-
ACM ICMR 2025 Workshop
SocialDF: Benchmark Dataset and Detection Model for Mitigating Harmful Deepfake Content on Social Media
MAD '25: Multimedia AI against DisinformationA dataset of real-world social-media deepfakes and a model to detect them.
-
INTERSPEECH 2024
ASGIR: Audio Spectrogram Transformer Guided Classification and Information Retrieval for Birds
Conference of the International Speech Communication AssociationAn audio spectrogram transformer for bird species classification and retrieval.
Timeline
research · industry · leadership · education-
Fall 2026 — now● active
SPAR
Research Fellow · Remote -
Jul 2026 — now● active
MARS V · Cambridge AI Safety Hub
Research Scholar, Computational Self-Awareness stream, led by Yossi Gandelsman · Remote- Do instruction-tuned LLMs encode knowledge about their own computations and behaviour? Studied with residual-stream analysis, introspective probing, steering and causal interventions.
- Testing where this information becomes readable, whether it generalises across prompts and tasks, and whether intervening on it changes behaviour instead of just predicting it.
-
May 2026 — now● active
Shiba AI
Research Collaborator, AI Safety · Remote- Mechanistic interpretability for agentic and multi-agent systems: internal representations tied to coordination, emergent behaviour, moral reasoning and multi-turn behaviour shifts.
- Lightweight monitors and interventions for jailbreaks, adversarial context and multi-turn safety failures in agent workflows.
-
May 2026 — Aug 2026
Google Summer of Code · ML4SCI
Open-source developer, advised by Michael Toomey (MIT CTP) and Sergei Gleyzer (Alabama) · Remote- Built Lens-JEPA, a self-supervised Joint Embedding Predictive Architecture for gravitational-lensing images when labels are scarce.
- Wrote reusable PyTorch infrastructure for training, evaluation, ablations, checkpointing and downstream transfer.
-
May 2026 — Jul 2026
Lexsi Labs
AI Engineering Intern · Remote- Built a variance- and probability-calibrated distillation method that keeps 98% of TabPFN v3's performance on TabArena.
- Analysed how distillation changes internal representations and predictive uncertainty in tabular foundation models.
-
May 2025 — now● active
StarkVision · NivionAI
Founder & Lead Researcher · Delhi- Started and run a 20+ member student research group. More below ↓
-
May 2025 — Dec 2025
Brown University · Meta FAIR
Research Collaborator with Prof. Randall Balestriero · Remote- Self-supervised learning theory, representation robustness and generalisation. Co-authored Relational Representation Learning (NeurIPS 2025 UniReps).
- Studied the relational structure that deep representations encode, beyond pointwise embedding quality.
-
May 2025 — Jul 2025
Mythyaverse
AI Engineering Intern · Remote- Explainable deep-learning models for ECG interpretation.
- Real-time sign-language translation with transformers, 95%+ accuracy on Arabic Sign Language. Shipped LLM and multimodal pipelines to production.
-
Feb 2025 — May 2025
UC San Diego
Research Intern with Prof. Pengtao Xie · Remote- Built an LLM framework for DNA–RNA interaction prediction. Curated a 100k+ sequence dataset and designed lightweight genomics architectures that scale efficiently.
-
May 2024 — May 2026
SBI Lab · IIIT Delhi
Research Intern with Prof. Anubha Gupta · Delhi- Led Uncovering the Latent Potential of Deep Intermediate Representations (ICML 2026 Spotlight): task-relevant information is spread non-monotonically across depth and can be recovered from intermediate layers.
- Designed controlled layer-wise probing, representation-geometry analysis, ablations and matched readouts, to separate where information is stored from how well a head uses it.
- Medical imaging across 2D and 3D: blood-cell, dental and spinal scans, with multi-view and volumetric transformers.
-
May 2024 — Oct 2024
HMI Lab · IIIT Delhi
Research Intern with Prof. Jainendra Shukla · Delhi- Multimodal LLMs for live sports commentary and fine-grained video captioning. Designed and benchmarked video encoders and attention-based fusion.
-
Dec 2023 — now
Cyborg · IIIT Delhi
Coordinator, AI/ML club · Delhi- Ran hands-on ML and deep-learning workshops for 50+ undergraduates; mentored first-time model implementations, debugging and reproducible experiments.
-
Aug 2023 — May 2027
Indraprastha Institute of Information Technology, Delhi
B.Tech in Computer Science & Artificial Intelligence · CGPA 8.81 / 10
StarkVision
founded 2025I started StarkVision, the research arm of NivionAI, so that undergraduates could do real research and publish it as first authors. It has 20+ members. I mentor them from the first question through experiment design, ablations, reproducibility and writing. Work from the group:
- Uncovering the Latent Potential of Deep Intermediate RepresentationsICML '26 Spotlight
- Catching AI-Generated Code at ScaleEMNLP '26
- Melody or MachineTMLR
- M-SCANCVPR '25 Demo
- CLAM · SocialDFNeurIPS-W '25 · ICMR-W '25
DroidPlanner
A JEPA world model for offline goal-conditioned planning. Adds a K=3 multi-step latent-consistency loss and gated pyramid aggregation: 85.0 ± 1.15 on DMControl Reacher, 4.6× lower variance, about 3% extra parameters.
Lens-JEPA
Self-supervised predictive pretraining for gravitational-lensing images, built for ML4SCI during Google Summer of Code. Read the blog post →
Market-JEPA
A multi-horizon JEPA-style representation model for financial markets, with leakage-aware evaluation, as-of targets, sector residualisation, EMA target encoders and event-aware memory over 1–72 bars.
QuantLangGraph
Weekly Indian-equity ranking with seven expert rankers plus macro, news and fundamentals agents. 22.51% cumulative return, Sharpe 1.45, beat the Nifty 50 in 28 of 41 weeks.
Earlier
- RE-DACTTransformer-based redaction for text, PDFs, images and video, 99%+ accuracy. 2nd place at Smart India Hackathon 2024.2024
- Super-resolution for gravitational lensingGANs, SRResNet and diffusion denoising. AUROC 0.9978, SSIM 0.94.2024
- Focus FlowAdaptive learning platform for neurodiverse students with real-time attention tracking.2024
Blog
Lens-LeJEPA: Self-Supervised Learning for Dark Matter Substructure
My Google Summer of Code 2026 project with ML4SCI DeepLense: pretraining one encoder on unlabelled lens images with two lensing-aware loss terms, what it gains on classification, where it loses, and the ideas that did not work.
Honors & Awards
- Grand AI Challenge PS-03 · NCIIPC / IndiaAI, Government of IndiaStage 2 winner (~$26K research grant + $5K cash prize); Stage 3 finalistTop 6 / 100+
- ICML 2026 SpotlightUncovering the Latent Potential of Deep Intermediate RepresentationsSpotlight
- Amazon ML Challenge 2025Multimodal product-price prediction from image + textRank 13 / 20,000
- Smart India Hackathon 2024RE-DACT, grand finale at IIT Kharagpur2nd place
- F1nalyze · Formula 1 DatathonKaggle competition1st place
- ML Hackathon · IIT Jodhpur4th place
- FILL THE VOID() · IIT Jammu5th place
- JEE MainOut of ~1.2 million candidates99.72 %ile · AIR 3334
Service, Teaching & Toolkit
Reviewing
NeurIPS 2026 · ICML 2026 Workshops · NeurIPS 2025 Workshops · ACM ICMR 2025
Teaching & mentoring
Mentor to 20+ student researchers at StarkVision. ML and deep-learning workshops for 50+ undergraduates at Cyborg, IIIT Delhi, covering neural networks, NumPy and Pandas.
Methods
Probing, activation steering, causal interventions, transformers, JEPA / world models, contrastive learning, distillation, calibration, time-series modelling, learning-to-rank, robustness evaluation.
Tools
Python, C++, SQL, PyTorch, TensorFlow, Hugging Face, scikit-learn, XGBoost, CatBoost, Docker, Linux, multi-GPU distributed training.
Off the clock
Photography
Street, low-light and long exposures. Composition is just attention with a camera. @arnesh_photography_
Badminton
Played since school. Doubles, preferably. The one place I don't overthink.
Travel
New cities, same notebook. Most of my good ideas show up on a train.
Contact
Looking for pre-doc and PhD opportunities
I'm applying for pre-doctoral and PhD positions starting in 2027, and I'm always happy to talk about interpretability, introspection or representations. Email is the best way to reach me.
Email mearnesh23129@iiitd.ac.in
- Google Scholar
- GitHub@arnesh2212
- LinkedInarnesh-batra
- alphaXiv@arnesh-batra
- Instagram@arnesh_photography_