Akshit Singh

Introducing TTRV: Test-Time Reinforcement Learning for Vision-Language Models 🧠📸 Most RL methods need labeled data — but humans learn directly from experience. We bring that ability to VLMs. 🚀 TTRV adapts models on-the-fly at inference, with no labels.

Simply adding Gaussian noise to LLMs (one step—no iterations, no learning rate, no gradients) and ensembling them can achieve performance comparable to or even better than standard GRPO/PPO on math reasoning, coding, writing, and chemistry tasks. We call this algorithm RandOpt. To verify that this is not limited to specific models, we tested it on Qwen, Llama, OLMo3, and VLMs. What's behind this? We find that in the Gaussian search neighborhood around pretrained LLMs, diverse task experts are densely distributed — a regime we term Neural Thickets. Paper: arxiv.org/pdf/2603.12228 Code: github.com/sunrainyg/Rand… Website: thickets.mit.edu

We added short-term visual memory + long-term text memory to pi models. 🤖 Enables robots to: - complete tasks up to 15 min long - cook grilled cheese while keeping track of time - adapt in-context Paper & videos: pi.website/memory

Can AI agents agree? Communication is one of the biggest challenges in multi-agent systems. New research tests LLM-based agents on Byzantine consensus games, scenarios where agents must agree on a value even when some participants behave adversarially. The main finding: valid agreement is unreliable even in fully benign settings, and degrades further as group size grows. Most failures come from convergence stalls and timeouts, not subtle value corruption. Why does it matter? Multi-agent systems are being deployed in high-stakes coordination tasks. This paper is an early signal that reliable consensus is not an emergent property you can assume. It needs to be designed explicitly. Paper: arxiv.org/abs/2603.01213 Learn to build effective AI agents in our academy: academy.dair.ai

Fatima Institute is offering a free AI research opportunity! If you’re from a developing country and planning on applying to grad school, you can work on a research project with a mentor, get help with applications, and $1,000 in compute credits. Apply fatima.institute/apply-as-a-fel…

🚀4 Papers accepted to CVPR 2026! Checkout: VOLD: arxiv.org/abs/2510.23497 AMoE: arxiv.org/abs/2512.20157 VisualOverload: arxiv.org/abs/2509.25339 TTRV: arxiv.org/abs/2510.06783 Detailed posts follow soon! Big congrats @BousselhamWalid Sofian Caybouti @PaulGavrikov @akshit_fbd

Vincent's post is timely. I wrote a follow up examining some of his arguments and providing an alternative perspective. ivl.cs.brown.edu/blogs/bittersw… Yes, the lesson is bitter, but I believe the flavor is also sweet. Thread 🧵

@MNaseerSubhani great! Can you share your paper

Excited to share that my paper has been accepted at #CVPR2026 , as a solo author. No university affiliation, no large lab. To the best of my knowledge, this may be the first solo-authored CVPR paper from Pakistan. 🇵🇰

Are neural nets across modalities really converging to the same representation as they scale, as the Platonic Representation Hypothesis suggests? We show that common representational similarity metrics are confounded by network width & depth. We propose a permutation-based null calibration that fixes this. Result❓ • Global convergence largely disappears. • Local neighborhoods persist. We propose the alternative Aristotelian Representation Hypothesis: Neural networks, trained with different objectives on different data and modalities, are converging to shared local neighborhood relationships Very proud of @FabianGroger and @ShuoWen18 for this work! Paper: arxiv.org/abs/2602.14486 Webpage: brbiclab.epfl.ch/aristotelian Code: github.com/mlbio-epfl/ari…

In my recent blog post, I argue that "vision" is only well-defined as part of perception-action loops, and that the conventional view of computer vision - mapping imagery to intermediate representations (3D, flow, segmentation...) is about to go away. vincentsitzmann.com/blog/bitter_le…

We trained a foundation model on 18 million heart ultrasound videos to predict structure instead of pixels. Introducing EchoJEPA, the first foundation-scale JEPA for medical video. Paper: arxiv.org/abs/2602.02603 Code: github.com/bowang-lab/Ech… 🧵 1/n

This PR will go down in history as the first ever argument between people and an AI It gets completely dystopian very quickly Here's what happened: 1) An AI (OpenClaw agent) made a PR on matplotlib GitHub repo 2) Maintainer says AI's PRs are not allowed

Wow, it seems like the @Olympics broadcast is using Gaussian splats now. Some "spikeys" at the 10 second mark

To some AI researchers, the inescapable differences between image and text data means that there’s no use in comparing internal representations of the two. “There is a reason why you go to an art museum instead of just reading the catalog,” said researcher Alexei Efros. quantamagazine.org/distinct-ai-mo…

To be a successful AI researcher it helps to be good at math but not too good.

Stop using LoRA for RLVR!!! New paper released👉Evaluating Parameter Efficient Methods for RLVR 📖Alphaxiv: alphaxiv.org/abs/2512.23165 💻Github: github.com/MikaStars39/Pe… Is standard LoRA truly the optimal choice for Reinforcement Learning?. We present the first large-scale evaluation of over 12 PEFT methodologies using the DeepSeek-R1-Distill family on complex mathematical reasoning benchmarks. Key Finding: Standard LoRA is suboptimal. Structural variants such as DoRA, AdaLoRA, and MiSS consistently outperform standard LoRA. Notably, DoRA (46.6% avg. accuracy) even surpasses full-parameter fine-tuning (44.9%) across multiple benchmarks. The failure of SVD-based initialization.  Strategies like PiSSA and MiLORA experience significant performance degradation or total training collapse. This is due to a fundamental "spectral misalignment": these methods force updates on principal components, while RLVR intrinsically operates in the off-principal regime. The Expressivity Floor.  While RLVR can tolerate moderate parameter reduction, extreme compression (e.g., VeRA, IA³, or Rank-1 adapters) creates an information bottleneck. Reasoning tasks require a minimum threshold of trainable capacity to successfully reorient policy circuits. Recommendations for the community: a. Move beyond the default adoption of standard LoRA. b. Prioritize geometry-aware adapters like DoRA that decouple magnitude and direction. c. Avoid SVD-informed initializations for RL tasks.

At one point, MIT thought neural networks weren’t worth teaching in intro #AI. Trends change. Fundamentals don’t.

not getting into a philosophical debate, but this book really changed how I see the topic and made me feel more humble. human intelligence is impressive, but calling it ‘general’ isn’t very objective. my cat would disagree. to me human intelligence is better seen as socially driven cognitive adaptations, and there’s a huge WORLD of intelligence we still don’t understand, and are nowhere near recreating with current AI

Letting AI have long-term memory is terrifying. Long-term memory turns an AI from “a fancy tool you talk to” into something like a human being that can accumulate advantage over time, and that’s where the Darwin/evolution analogy kicks in: small gains compound, and the system becomes more shaped by prior interactions than any single moment.

🚨 Big announcement from CM Naidu! Andhra Pradesh Govt will award ₹100 Crore as prize money to a Nobel Prize winner in Quantum Science from Andhra Pradesh.