Garima

PhD student @IIIT Allahabad

[Garima @Garima68793847]

🚗 From Waterfall to Vibe Coding, which Team Are You On? 😅

This visual perfectly sums up how different development approaches feel in real life:

Which “car” best represents your current team’s approach?

[jules @juleslogs]

Some upcoming AI/CV conference deadlines I have on my radar:

• WACV 2027 (Disney World, Florida, USA) - R1: Jun 26, 2026

• ACCV 2026 (Osaka, Japan) - Jul 5, 2026

• AAAI 2027 (Montréal, Canada) - Jul 28, 2026

• WACV 2027 (Disney World, Florida, USA) - R2: Aug 28, 2026

• 3DV 2027 (Thessaloniki, Greece) - Aug 28, 2026

Links in the replies...

[Samuel Schmidgall @SRSchmidgall]

Our posting for joining Google DeepMind as a Research Scientist was down for a few days but now it is back up!

Apply here: https://t.co/Yk5iMbMQPu

And fill out this form: https://t.co/zdeqryH3hB

[Jiefeng Chen @jiefengchen1]

Our team at Google Cloud AI Research is looking for a Student Researcher Intern this Fall to dive deep into long-horizon coding agents.

We’re looking for a builder. Someone who doesn’t just read about agentic workflows, but actively creates them.

What we’re looking for:
🎓 Academic Rigor: Currently pursuing a Ph.D. with a strong publication record.
💻 Technical Chops: Excellent, hands-on coding skills are an absolute must.
🤖 Agent Experience: If you’ve built or experimented with long-horizon coding agents (like Claude Code, Gemini CLI, or similar frameworks), we want to talk to you.

Come help us push the boundaries of LLM-based software engineering. 🚀

If this sounds like you, please fill out the interest form below:
🔗 https://t.co/kx7h3uJ8Rg

Please share or tag anyone who might be a good fit!

#Google #AI #PhD #LLM #AIAgents #Internship

[Google Research @GoogleResearch]

Google’s PhD Fellowships Program supports graduate students doing exceptional and innovative research in computer science and related fields.

Applications close this Thursday, April 30. Apply today! → https://t.co/ySLv0Jet0X.

[Gabriel Agana🦍👨🏽‍⚕️ @gaagana_]

📍📍APPLY NOW

Applications are open for Google Deepmind AI Research Foundation Courses.

Deadline: May 17, 2026

Link: https://t.co/usBv3vvaBY

[Viktor Oddy @viktoroddy]

I recorded a 18-min tutorial on The New Google Stitch - Better than Figma (And It's Free)

[Google Labs @GoogleLabs]

Introducing the new @stitchbygoogle, Google’s vibe design platform that transforms natural language into high-fidelity designs in one seamless flow.

🎨Create with a smarter design agent: Describe a new business concept or app vision and see it take shape on an AI-native canvas.
⚡️ Iterate quickly: Stitch screens together into interactive prototypes and manage your brand with a portable design system.
🎤 Collaborate with voice: Use hands-free voice interactions to update layouts and explore new variations in real-time.

Try it now (Age 18+ only. Currently available in English and in countries where Gemini is supported.) → https://t.co/pmT9iHEpZa

[Peter Agboola @baba_Omoloro]

Anthropic has launched free courses to master AI with certificates for $0.00

https://t.co/hwJvuJHIdo

[Mathieu @miniapeur]

Mathematician training:

[RapperPandit @RapperPandit]

🚨This is what everyone should talk about AI India Summit not one off incident of Galgotia

Bharat has successfully made LLMs scanning Lacs of Ancient Manuscripts ❗️
•Ancient Medicine & Surgery
•Ancient Mathematics
•Ancient Sciences
•Ancient Ayurveda

[Dan Biderman @dan_biderman]

People think LoRA is a magic bullet for LLMs. Is it? Does it deliver the same quality as full finetuning but on consumer GPUs?

Though LoRA has the advantage of a lower memory footprint, we find that it often substantially underperforms full finetuning. However, it forgets less of the base model’s capabilities. In this work, we exhaustively explore this trade-off and provide practitioners a clear view of the difference between the methods.

https://t.co/DOWR8feRp3

[Cameron R. Wolfe, Ph.D. @cwolferesearch]

Continual learning is being positioned as a prerequisite for AGI (i.e., general systems must be adaptable).

I spent a large part of graduate school working on continual learning. Over the last month I've revisited the literature, considered its connections / relevance to LLMs, and captured everything in a long-form blog (scheduled for release tomorrow morning).

Most research on the topic of continual learning is very different from the LLM research we see today. The continual learning problem for LLMs is unique due to scaling. Prior knowledge and data are nearly infinite, which creates a lot of new considerations.

One could argue that the generality of LLMs makes continual learning easier, but it also increases the risk of catastrophic forgetting. Plus, pure model and data scale make the application of even basic techniques more difficult.

As an illustrative example, replay buffers are a common and simple technique to retain a model's knowledge over prior data. But, efficiently maintaining a replay buffer over a multi-trillion-token training corpus is an extremely complex systems problem. We might also not even have access to any of the model's prior data.

Despite these difficulties, it's possible the tools we are currently using in LLM research may naturally lend themselves toward solving continual learning:

- Large-scale multi-task training is fundamental to how LLMs are trained.
- RL training naturally avoids forgetting of prior knowledge.

There are very few "free" wins at LLM scale, but RL might be one of them. Continual learning is not a completely disjoint problem from what we are already trying to solve in LLM research. Rather, continuing on the current trajectory (with modifications to consider different styles of learning) will yield natural progress.

[Cameron R. Wolfe, Ph.D. @cwolferesearch]

Continual learning is a popular topic in LLM research, but it might not be as far away as we think. Standard RL is naturally good at continual learning…

Continual learning refers to the ability of an AI model to learn from new tasks and data over time. General intelligence should be adaptable, which has led some to believe that continual learning is a prerequisite for AGI. The reasoning for this argument is clear—adapting to new tasks (i.e., “on-the-job” learning) is a common trait of humans—but rigorously studying this concept is hard. Real-world continual learning is unstructured, noisy, and open-ended.

Why is this hard? If incoming data were sampled uniformly from the LLM’s existing training distribution, continual learning would be less difficult. However, the problem occurs when our incoming data is non-IID (i.e., taken from previously unseen domains / tasks). In this case, our LLM may undergo catastrophic forgetting, where it performs poorly on old tasks or knowledge.

How do we test this? There are two common experimental setups for continual learning:

1. Batch-incremental: The model receives a full batch (e.g., a new task or dataset) at a time and performs offline training on each batch.
2. Streaming: The model receives one example at a time and performs real-time updates (one or a few forward/backward passes).

Both of these setups receive non-IID data and are sequential in nature, meaning when new data is received we cannot return to the prior data (unless explicitly stored in a buffer). We can interpolate between them by varying batch size and number of model updates per batch.

RL for continual learning. Although continual learning is quite difficult, there are many papers showing that:

- SFT had a tendency to catastrophically forget,
- But RL tends to avoid catastrophic forgetting naturally.

[1] tests LLMs with SFT and RL on a sequential learning task with seven vision-language datasets (i.e., each dataset is a batch). SFT improves in-domain performance but catastrophically forgets, while RL both i) maintains prior task performance and ii) keeps / improves performance on general benchmarks. Similar findings are shown in [2, 3, 4].

Why is this the case? Several works provide useful analysis as to why RL works well for continual learning naturally. In particular, the following sequence of findings provide insight into the mechanics of not forgetting in RL:

- [2] shows that on-policy data is pivotal for RL’s lack of forgetting.
- [3] shows that, due to the use of on-policy data, RL is biased towards solutions with a low KL divergence w.r.t. the original model’s distribution, which reduces forgetting.
- [4] explains that on-policy data allows us to avoid “confident conflict” tokens that disagree with the LLM’s internal knowledge and cause detrimental policy updates.

[spidey @lochan_twt]

single research paper that completely changed the future of LLMs

[Amit LeVi ✈️ Tokyo @AmitLeViAI]

Continual learning in 10 minutes!

DeepMind researchers recently predicted: “2026 will be the year of continual learning.”

So, do you want to know what this is all about? Read this post carefully 🙂

Continual learning in LLMs means a deployed model can keep improving learning new domains, products, policies, and user needs without full retraining and without losing old skills.

In practice, the moment you try to teach a model something new via fine-tuning, you hit the core problem: catastrophic forgetting. New updates modify the same parameters that supported older abilities, causing performance on earlier tasks to drop.

What’s changed in recent research is the shift from “it forgets because it overwrites” to “what about optimization makes it overwrite.”

1. Replay (rehearsal) The first solution is replay. While Atlas learns new domain for instance Orion, the training set includes examples of older tasks as well for (e.g emails, FAQs, reasoning prompts).

Replay works because you cannot forget what you keep practicing.
It remains the strongest baseline because it directly attacks the forgetting mechanism. The modern twist is small, strategic replay and streaming replay, where only a limited memory budget is kept and updated over time (https://t.co/YrrfQaZKnR).

For Atlas, this means keeping a small set of “anchor behaviors” that must never be lost, instead of storing everything.

2. Parameter-efficient continual learning The second family. Instead of updating all weights, adapters or LoRA modules are added and only those are trained.

These act like plug-ins: new skills live in small components while the base model stays mostly frozen. Forgetting is reduced by restricting plasticity.
A popular hybrid combines replay + LoRA, keeping learning lightweight while protecting old behavior. One concrete method is ERI-LoRA for dialogue and NLU tasks (https://t.co/yej7JCwRLg).

A recent idea is spurious forgetting. Sometimes the model hasn’t lost the knowledge it has lost alignment with the task format or instruction style.

Atlas may still “know” old FAQs, but after learning Orion it interprets prompts differently, changes formatting, or shifts decision boundaries, causing benchmarks to mark answers wrong.

An ICLR 2025 paper shows much early performance collapse is task-alignment disruption, not true knowledge loss, and that freezing lower layers reduces this effect (https://t.co/CKn0Qdo4k7).
The knowledge is still there; the behavioral contract has shifted.

3. Test-time training and adaptation
The third family is Test-time training, which performs small, temporary learning steps while the model is being used, often updating only adapters. The motivation is simple: repeated fine-tuning is one of the fastest ways to destroy old behavior.

A 2025 method proposes verifier-driven test-time training: Atlas generates multiple answers, a verifier scores them, and only high-confidence cases are used for tiny, reversible updates (https://t.co/xvffgB5Sz8).
The key insight: continual learning doesn’t require a permanently evolving model. It can be a stable core plus short-lived micro-learning.

Closely related is test-time scaling. Instead of updating weights, the model uses more computation, self-checking, step-by-step verification to improve reliability.

Test-time scaling changes how hard the model thinks, not what it knows. A 2025 EMNLP paper explores verifier-guided, training-free test-time scaling for reasoning reliability (https://t.co/j35oG6mgOJ).
For Atlas, this means improving math accuracy by double-checking reasoning, not retraining.

4. (RAG). Instead of storing knowledge in weights, documents are retrieved externally and injected at inference. So now Atlas doesn’t retrain when Orion changes, the docs are retrieved.

RAG avoids forgetting, but it updates facts more than behavior. This is why hybrid systems are emerging: retrieval for knowledge, adapters for skills.

[Massimo @Rainmaker1973]

Computer engineering for babies

[Nijol @Nijol71]

BREAKING: Don't copy and paste answers from ChatGPT.

ChatGPT writing is easily detectable.

Use these prompts instead and see the magic:

[The Math Flow @TheMathFlow]

Notations of derivative. :D