Two months on an open problem in reinforcement learning or computer vision, with a Vizuara research team working on your paper alongside you. The goal is a manuscript ready to submit by 20 January 2027.
The robot is Rollo, named after a rollout: one run of an agent through its environment. Scroll and it explores the bootcamp with you. Drag to look around.
In most mentorship programs you do the work by yourself and meet a mentor once a week. Here a Vizuara research team works on your paper with you, and two mentors guide it from the first week to the final review.
You choose the problem, run experiments and write.
We draft, run experiments of our own and review every result you send.
They set the direction, approve the method and give the final review.
They help our team search papers, log runs and check numbers.
You get a list of open questions, each small enough for two months, and choose one.
Our team maps the related work and writes the first draft of your literature review.
We draft the method and the experiment plan. The mentors approve them before any experiment starts.
You run yours. We run our own on Vizuara's GPUs.
Numbers and plots are verified against the run logs before they go into the paper.
In the format of the venue we choose together. The mentors give the final review.
Each question is open, starts from a paper accepted at a major venue in 2025 or 2026, and is small enough for two months on modest compute.
Very deep networks help self-supervised RL when the agent can keep collecting experience. Train the same contrastive critic at depths 4 to 64 on offline goal-conditioned datasets and find where the benefit stops.
A possible result: success rate against depth on fixed data, and the depth where it stops rising.
At large k, base models can score higher on pass@k than their RL-trained versions. Check whether that crossover appears for a small model trained on one narrow task, and whether an exploration bonus slows the narrowing.
A possible result: pass@k curves for a base model and its RL-trained version on one task, with and without the bonus.
LeWorldModel trains a JEPA world model of about 15M parameters from pixels on one GPU. Measure how planning success falls with horizon and with goals outside the training data, and whether a little data collected by the planner repairs it.
A possible result: a map of where planning breaks, by horizon and by distance from the training data.
Ordered action tokens let a policy decode only a prefix and trade inference cost against precision. Choose the prefix length at each timestep, for example from token entropy, and measure success rate and latency.
A possible result: a rule that picks the prefix length at each step, with its cost in success and its gain in speed.
A recent audit found that a small probe with no language encoder scores near the state of the art on LIBERO. Apply the same four diagnostics to a benchmark the paper did not audit.
A possible result: an audit of one more benchmark, showing which reported gains survive the four checks.
FlashSAC bounds weight, feature and gradient norms so that off-policy training stays stable with larger models. Remove the three bounds one at a time on two tasks with different action dimensions.
A possible result: an ablation that shows which of the three bounds matters, and at which scale.
In CLIP and DINOv2, a few neurons push high-norm activations onto outlier tokens, and moving them into an extra token at test time cleans up the feature maps. Check whether the same neurons exist in DINOv3 or in a backbone trained by reconstruction.
A possible result: a yes or no for two more model families, with feature maps before and after.
Linear probes show that DINOv2 lines up geometric signals in a low-rank subspace, while MAE spreads them out. Repeat the analysis on I-JEPA and V-JEPA encoders to see which pattern latent prediction follows.
A possible result: the same subspace analysis for JEPA encoders, placed next to DINOv2 and MAE.
Theory says LeJEPA recovers the world's latent variables up to a linear map, at the population optimum and for Gaussian latents. Measure how that degrades with sample size and with latents that are not Gaussian.
A possible result: how recovery of the latent variables degrades with fewer samples and with non-Gaussian latents.
Depth Anything 3 predicts geometry from any number of views with one plain transformer. Probe the layers of its small model, then measure how accuracy falls as the overlap between views shrinks.
A possible result: a layer-by-layer map of pose and depth, and accuracy against view overlap.
When the image is swapped for a similar one in the middle of reasoning, open VLMs mostly miss it. On a 7B to 8B model, try to repair this without a user instruction, for example by encoding the image again when the model says it is checking.
A possible result: a fix that makes the model notice the swap, measured on the released image pairs.
Training has two timescales. Good samples appear early, and memorisation emerges later, at a time that grows with the size of the dataset. Test whether that holds for a small diffusion transformer, or with flow matching.
A possible result: the two timescales measured again for a transformer, or for flow matching.
The open problems start from 21 papers accepted at NeurIPS, ICML, ICLR, CVPR, ECCV, CoRL and RSS in 2025 and 2026. Four of them won a best paper award.
We checked each acceptance in November 2026 against the venue's program or awards page, or the authors' own record.
This year, ten papers from Vizuara's research program were accepted at NeurIPS 2026 workshops. They came out of the same way of working this bootcamp uses. One of them studies how RL post-training of language models is evaluated.
Two of the ten accepted papers. Read the announcement ›
You pick a problem. We draft the literature review and the method with you, and the mentors approve the plan.
You run experiments and send results. We review each one, suggest the next step, and run experiments of our own.
We write the paper with you. The mentors give the final review, and the manuscript is ready to submit.
There are no scheduled calls. Work passes back and forth between you and our team on the Vizuara Research Mentorship Portal, so feedback begins when your result arrives.
An illustration of one week, Monday to Friday.
Both mentors set the direction of each project, agree the method before the experiments start, and give the final review of the paper.
Sreedath works on vision-language models, robotics and transformer architectures. He holds a PhD from MIT and a BTech from IIT Madras.
Rajat works on reinforcement learning from human feedback, policy optimisation and agentic systems. He holds a PhD from Purdue and a BTech from IIT Madras.
Every paper deadline below falls after 20 January 2027, so a finished manuscript can go straight to one of them. We choose the target with you.
Main conference
Workshops
Conference, through a journal
Main conference
Workshops
Main conference
Main conference
JournalDeadlines marked expected follow each venue's last cycle and will be confirmed when the 2027 calls are published. ICML 2027 also asks for an abstract by 14 January, before the bootcamp ends, so that route needs an early start. CoRL and NeurIPS have not opened their 2027 pages yet, and both reviewed on OpenReview in 2026.
No one can guarantee that NeurIPS, ICML, CVPR or any other venue will accept a paper, and no credible program will promise it. Peer review exists so that an independent community decides whether a piece of research meets its standard.
What you pay for is the research effort itself: the attempt to turn a difficult open problem into publishable scientific work, with a team that designs experiments, develops methods, evaluates them rigorously and keeps iterating until the paper is ready.
The experiments our team runs are paid for from Vizuara's own budget.
We mostly provide tools such as Google Colab. Anything beyond that is yours to arrange.
We only take on topics whose experiments fit modest compute. If a problem needs more than that, for example training a large vision-language-action model from scratch, we cannot take it on, however interesting it is.
The group is small, because both mentors work with every team personally.
The early-bird price holds until 16 November 2026, 11:59 PM IST. From 17 November, the price is ₹1,25,000.
Compute for your own experiments beyond tools such as Google Colab is not included.
Batch 1 runs from 20 November 2026 to 20 January 2027 · Entirely asynchronous · Questions? Write to research@vizuara.com
Batch 1 begins on 20 November 2026. The early-bird price holds until 16 November.
Enroll for ₹95,000
Vizuara Research