Prakhar Kaushik (@toshi2k2) header
Prakhar Kaushik (@toshi2k2) profile picture

-On the job market- CS PhD at JHU. I make AI robust, efficient and continual like humans (Cognitive AI). 3D parts based CV learning expert. Physics of AI.

Academic & Research Job Market
Joined December 2023
Photo by Prakhar Kaushik (@toshi2k2)Photo by Prakhar Kaushik (@toshi2k2)Photo by Prakhar Kaushik (@toshi2k2)Photo by Prakhar Kaushik (@toshi2k2)Photo by Prakhar Kaushik (@toshi2k2)Photo by Prakhar Kaushik (@toshi2k2)Photo by Prakhar Kaushik (@toshi2k2)Photo by Prakhar Kaushik (@toshi2k2)Photo by Prakhar Kaushik (@toshi2k2)Photo by Prakhar Kaushik (@toshi2k2)
1/6 We analyzed the intrinsic dimension of 1k+ Neural Models (ViT, LLaMA-3, Mistral). The verdict? High-dimensional weights are an illusion. Models trained on disjoint datasets collapse into a shared, low-rank subspace. 🤯 The paper: toshi2k2.github.io/unisub/ A breakdown 🧵👇
2
15
4
31
11,811
Can interp/AI safety ppl write a decent paper about something truly original without a damn PR storm? Like stop rehashing old ideas into a rambling word salad, and at least put the same effort in your research that you put into your PR videos.
81
🐝 Introducing Honeycomb-Surfels (Hexels) honeycomb-surfels.github.io/ High-quality 3D. Less memory. Faster rendering. We introduce Hexels, advancing state-of-the-art quality–efficiency for surface-aligned primitives (surfels) - with on-device training now running on iPad and iPhone. The world is structured. Many regions are locally flat, with smoothly varying color. Hexels capture both within each hexagonal surface patch, allowing fewer, larger primitives to represent more of a scene. Across 34 scenes from Mip-NeRF 360, Tanks & Temples, and DTU: • More quality for the same memory: Highest mean PSNR among evaluated methods on all three datasets at matched model memory. +4.53–7.38 dB over 2DGS, with lower peak training VRAM. • More from each primitive: At matched primitive budgets, up to +4.74 dB and 2.3× the rendering FPS of 2DGS, while using less model memory and training VRAM. • Accurate 3D surfaces: 28.9% lower mean Chamfer error than memory-matched 2DGS across 15 DTU scans. Highest mean F1 among compared methods on six Tanks & Temples scenes: 34.38%, versus 31.72% for 2DGS. • The same primitive works in 2D: On the 45-image Image-GS benchmark with 10K primitives, 35.82 vs. 33.16 dB PSNR, with 34% less training time. No quantization or hacks required. 📱 Meet Hexel Edge: Capture. Train. Explore. Our open-source app for iPad/iPhone is in active development, with the complete capture-to-3D workflow already running locally on iPad. Capture your scene, train its Hexel representation, and explore new viewpoints - locally, efficiently, and privately. Your data remains yours. 📄 Paper: arxiv.org/abs/2610.37668 💻 Hexel Edge (source & build): github.com/honeycomb-surfels…
5
6
1
61
5,494
Prakhar Kaushik retweeted
Honeycomb-Surfels: Compact Hexagonal Primitives for Efficient Scene Rendering and Reconstruction @toshi2k2, @__mvp18__, @YuilleAlan tl;dr: hexagonal support + first-order local color field arxiv.org/abs/2610.37668
4
26
1,256
A few days ago, I was feeling agonized because 2-3 LLM sentences made into my ICLR paper draft. Reviewing 5 papers in my batch have relieved me a bit. They are so terribly AI slop, that my works look like epitome of human research experience in comparison. Wtf
I hate LLM writing. In a recent ICLR submission, my coauthor and I were running on fumes towards the end, and one of us tried to fix a latex table errors to fix its format. It ended up changing a number of sentences in the entire paper, which I'd so painstakingly written and rewritten - and didn't catch until I reread the draft next day. They are not wrong, but damn it hurts me reading them after we put in so much effort in the work (over a year).
322
It's a bit ironic to see so many internships from industry nepotistic teams to pop up this season due to CPT issues. Usually, they hire juniors, bf/gf, bhai-bhateejas, to produce mediocre papers using the immense compute resources they have. Shitty ppl.
1
151
I hate LLM writing. In a recent ICLR submission, my coauthor and I were running on fumes towards the end, and one of us tried to fix a latex table errors to fix its format. It ended up changing a number of sentences in the entire paper, which I'd so painstakingly written and rewritten - and didn't catch until I reread the draft next day. They are not wrong, but damn it hurts me reading them after we put in so much effort in the work (over a year).
1
3
623
i've come to realize that all the uncreative and technically weak ppl who used to do the barely there data, benchmark, "better model because more data and compute" stuff - are now just publishing barely there agentic crap.
1
121
The amount of garbage papers I have seen from the @NeurIPSConf acceptance posts alone is incredible. Maybe it should be renamed #NaIsloPSConf.
111
Do ppl posting about 3,4, 5 paper acceptances at @NeurIPSConf realize that it's not the flex they think it is, and is simply an expression of unethical academic conduct...
127
Everybody's complaining about the AI slop at @iclr_conf . There are so many easy solutions: 1. For first-time submitters, even if they are co-authors on a paper, they are asked to first submit to a workshop at the said conference or similar conferences. Only after they have one paper at one of the workshops (it doesn't need to be archival, necessarily) can they submit to the main conference. This will revive workshops, and workshops often are better venues for most of the papers and have better reception and a better experience for researchers. 2. Desk rejection based on AI slop and reproducibility. If there is no code or not enough information available for reproducibility, and if there are AI slop errors, then we can test and reject it explicitly. 3. 1- to 3-year bans for all co-authors for an AI slop-detected paper which goes past the desk rejection phase. Obviously, the bar for this will be higher, as in oral interviews with an author, but still, this will still scare at least a few of the AI slopers. 4. Aggressively taxing a large number of paper submissions. If you have the money to have a big lab and 30 to 40 papers per conference, then you likely have the money to pay for the reviews as well. No paper limits with following fee structure: C(N) = Σ (i=2 to N) 100 × 2ⁱ⁻² 5. The above revenue collected can be used to hire or pay AI slop detectors whose job will be explicitly detecting papers with AI slop, and one or more senior expert reviewers per paper, who are actually paid for their review of the paper and have more weightage. This will also allow detection of AI-generated or poor-quality reviews and meta-reviews.
1
4
490
There is a pretty easy way of reducing AI slop submissions - yearly bans would be great. If a paper is unanimously detected to be largely AI generated (can be verified like TMLR's viva) - ban ALL authors from submitting any research papers to major confs for 1-3 years.
114
I suspect this @iclr_conf review cycle to be terrible. Already, there were issues regarding extremely poor reviewing, collusion & bias - given the no. of papers now, this has become astronomically expensive to even care about the earlier issues. We are going to have a plethora of stupid reviewers, and stupider ACs taking the limelight in an utterly broken system. The bias and collusion will worsen - with ppl submitting 20 papers or hiding behind big names getting low quality/AI slop/agentic slop papers accepted - while many novel, great ideas losing on the coin flip.
103
In last few job interview sessions, I've had this weird occurrence - with this 1 person trying their best to grandstand - often asking me questions clearly not in my research forte or trying to argue against my own works - which I clearly know better than them. Stop the crabism.
1
106
People whose research largely depends on data-driven methods, driven by associations with well-funded industry labs always have similar unfounded opinions. There are a lot of problems to solve - creative, original ideas will always be better than agentic slop of the many - academia will stay.
the reason is simple and obvious. full papers and conferences take way too long to publish and way too much extra effort than the work itself. papers are for humans but this era is for agents. with tiny compute resources and long delays, nothing academia makes is gonna come up relevant. and since this field is (at least should be) all about making stuff work rather than self-entertainment like manufactoring good paper stories, I don't understand what the point for all these is. It's sad for academia for sure, but we have to move on. the only way academia can stay relevant is with more agility than industry.
130