Pinned Tweet
Life Update: Last month I started as a Postdoctoral Researcher at Microsoft Research India!
If you are curious on why I chose to return to India after my PhD from Canada; I've written some personal thoughts about it! x.com/i/article/2072…
If you are curious on why I chose to return to India after my PhD from Canada; I've written some personal thoughts about it! x.com/i/article/2072…
63
63
898
314,076
Nandan Thakur retweeted
Is there a way to scale a reasoning model’s judgment for searching a massive corpus?
We introduce the Amortized Reasoning Index (ARI), a new kind of retriever that updates its weights during search, using gradients derived at inference time from the ranking preferences of a session, current agents spend lots of reasoning deciding which search results matter, and then throw all of that judgment away upon compaction. ARI is our attempt to “amortize” that judgment, so that reasoning spent on past queries keeps improving search on future ones.
With no relevance labels, ARI takes a 400M cross-encoder that starts out weak and grows its Recall@100 on held-out queries by up to 7×, ending above Qwen3-Embedding-8B on various oblique/reasoning-intensive search tasks.
Read our blogpost: dianetc.com/musings/ari/
Preliminary results and discussions are below!
We introduce the Amortized Reasoning Index (ARI), a new kind of retriever that updates its weights during search, using gradients derived at inference time from the ranking preferences of a session, current agents spend lots of reasoning deciding which search results matter, and then throw all of that judgment away upon compaction. ARI is our attempt to “amortize” that judgment, so that reasoning spent on past queries keeps improving search on future ones.
With no relevance labels, ARI takes a 400M cross-encoder that starts out weak and grows its Recall@100 on held-out queries by up to 7×, ending above Qwen3-Embedding-8B on various oblique/reasoning-intensive search tasks.
Read our blogpost: dianetc.com/musings/ari/
Preliminary results and discussions are below!
17
54
346
41,481
Nandan Thakur retweeted
Introducing Inkling 2 Small. Scoring 31 on the AA Intelligence Index, it's a big jump over its predecessor.
Inkling 2 Small is available for fine-tuning and inference on Tinker today, with full weight release planned soon.
thinkingmachines.ai/news/inkling-2…
Inkling 2 Small is available for fine-tuning and inference on Tinker today, with full weight release planned soon.
thinkingmachines.ai/news/inkling-2…
285
581
6,229
1,918,798
Nandan Thakur retweeted
Official now! 🚀
Excited to share the outcome of a project I'm particularly excited about. We're introducing a SOTA int8-quantized contextual embedding model, a public leaderboard for chunk retrieval evaluations, and exploring our model's potential for visual document retrieval.
Excited to share the outcome of a project I'm particularly excited about. We're introducing a SOTA int8-quantized contextual embedding model, a public leaderboard for chunk retrieval evaluations, and exploring our model's potential for visual document retrieval.
Introducing pplx-embed-v2-context: open-weight vision-language embedding models in 9B and 0.6B.
Both return one int8 vector per document, or one per chunk that sees the whole document.
Read the paper, written with @turbopuffer: arxiv.org/abs/2610.38854
3
4
32
1,408
With agents, one thing I'm seeing nowadays is reproduction happens effortlessly (ofcourse with supervision)!
but back in the days, manually understanding someone's code + making sure you are faithful to the setup + lots of pain and effort.
but back in the days, manually understanding someone's code + making sure you are faithful to the setup + lots of pain and effort.
i've reproduced the results from the MIT blog post today on weight-space programming. but one additional result i have to add: this is a new, step-change emerging capability that's appeared in just the latest set of models.
0
1
14
814
Nandan Thakur retweeted
you thought LLMs can only do text optimization and discovery for prompts, harnesses, and code?
turns out they’re starting to be good at natively reasoning about and writing out neural network weights directly too
in the limit, where does this leave symbolic local optimization heuristics like gradient descent…??
turns out they’re starting to be good at natively reasoning about and writing out neural network weights directly too
in the limit, where does this leave symbolic local optimization heuristics like gradient descent…??
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: fronely read/write neural network weights! 🧵
Blog: rikiyat.github.io/blog/weight-sp…Community noteLLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math.
rikiyat.github.io/blog/weight-sp…
github.com/RikiyaT/lm_und…
huggingface.co/Qwen/Qwen3.5-0…
huggingface.co/RikiyaT/Qwen3.…
49
47
535
48,056
Nandan Thakur retweeted
If this result holds up, it's a huge moment for computer architecture. We're already exploring machines that aren't programmed declaratively, but are trained. Their programming are their weights. Having weights as a direct output of knowledge is a HUGE shift in representation!
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: fronely read/write neural network weights! 🧵
Blog: rikiyat.github.io/blog/weight-sp…Community noteLLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math.
rikiyat.github.io/blog/weight-sp…
github.com/RikiyaT/lm_und…
huggingface.co/Qwen/Qwen3.5-0…
huggingface.co/RikiyaT/Qwen3.…
20
13
200
15,573
Wow, woke up to see this amazing direction of work from @rikiyatakehi jayden, @a1zhang, @lateinteraction!
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: fronely read/write neural network weights! 🧵
Blog: rikiyat.github.io/blog/weight-sp…Community noteLLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math.
rikiyat.github.io/blog/weight-sp…
github.com/RikiyaT/lm_und…
huggingface.co/Qwen/Qwen3.5-0…
huggingface.co/RikiyaT/Qwen3.…
1
0
11
750
Nandan Thakur retweeted
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: fronely read/write neural network weights! 🧵
Blog: rikiyat.github.io/blog/weight-sp…
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: fronely read/write neural network weights! 🧵
Blog: rikiyat.github.io/blog/weight-sp…
Community note
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math.
rikiyat.github.io/blog/weight-sp…
github.com/RikiyaT/lm_und…
huggingface.co/Qwen/Qwen3.5-0…
huggingface.co/RikiyaT/Qwen3.…
rikiyat.github.io/blog/weight-sp…
github.com/RikiyaT/lm_und…
huggingface.co/Qwen/Qwen3.5-0…
huggingface.co/RikiyaT/Qwen3.…
2,364
7,276
55,949
9,930,171
Nandan Thakur retweeted
This is a great post! A must read!
thomasjiralerspong.substack.com/p/chasing-the-…
thomasjiralerspong.substack.com/p/chasing-the-…
11
107
762
68,814
Nandan Thakur retweeted
Good slides are extremely intentional. When someone presents a paper, the paper already has all the information. But the speaker would make deliberate choices about things like, what to say and what not to say, what to put on a slide and what to say only verbally.
What to frame as a "side note" and when to say "this is the part I found particularly exciting."
Even individual word choices, like the subject, verb, and adjective, and when to pause, are intentional.
In fact, a good speaker would never really give a talk about a paper, even when that's seemingly what they're doing.
They would talk about what they want to say, like the story they want to tell, the kind of research they like, the vision they have, etc. Then they use the paper to support / back up their message.
Some tools are perfectly fine to use. But some skills take so much time and efforts, like years, and some tools proactively prevents learning them.
What to frame as a "side note" and when to say "this is the part I found particularly exciting."
Even individual word choices, like the subject, verb, and adjective, and when to pause, are intentional.
In fact, a good speaker would never really give a talk about a paper, even when that's seemingly what they're doing.
They would talk about what they want to say, like the story they want to tell, the kind of research they like, the vision they have, etc. Then they use the paper to support / back up their message.
Some tools are perfectly fine to use. But some skills take so much time and efforts, like years, and some tools proactively prevents learning them.
my hypothesis for why AI-generated slides for technical presentations are so awful:
people who don't have a lot of experience presenting think that to make a talk you just put all of the stuff you want to talk about on the slides, then you go through each slide and say the stuff the slide says.
how you're actually supposed to prepare a presentation: you decide what to say, write it out as a rough script, then put maybe 3% of the text on the slide, plus a bunch of visual aides (e.g. plots that only include data you're planning to discuss).
AI-generated presentations are mostly the former. they have a ton of stuff the presenter never mentions, and the stuff that is most useful for the presenter is absent or not highlighted. they include enough content for a presentation 10 times longer than the actual presentation is supposed to be.
maybe try doing it the other way: first write the script for what you actually want to say, and make sure you can say all of it in the time you have. then prompt the model to create the slides based on the script. ask it to use as little text a possible, and a minimalist visual style rather than the claude house style. you can thank me later (I accept both venmo and zelle).
14
75
694
51,515
Nandan Thakur retweeted
Very proud on our Series E.
When joining Cohere 4+ years ago, AI was still a niche product people barely cared about, and thought LLMs are just for auto-complete like your phone does it.
Like AI, search drastically changed in the time window. My org build the best retrieval systems on the market. We first build them for human users, but now AI agents are our largest users: We had to re-think what makes search great for AI.
Search / finding knowledge / managing knowledge remains still an open and challenging problem, and we are looking forward to the next generation of models & products in that space.
When joining Cohere 4+ years ago, AI was still a niche product people barely cared about, and thought LLMs are just for auto-complete like your phone does it.
Like AI, search drastically changed in the time window. My org build the best retrieval systems on the market. We first build them for human users, but now AI agents are our largest users: We had to re-think what makes search great for AI.
Search / finding knowledge / managing knowledge remains still an open and challenging problem, and we are looking forward to the next generation of models & products in that space.
We're proud to announce the initial close of our Series E, targeting up to CAD $7 billion at a CAD $35.7 billion valuation; marking the largest private financing round in Canadian startup history.
11
7
118
5,013
Nandan Thakur retweeted
Love BrowseComp-Plus? Meet 𝐂𝐌𝐀𝐒𝐒! 🚀
cmass.tech
🔥Your next challenge for agentic search: 423 question–answer pairs spanning BrowseComp-Plus, DeepSearchQA, Total Recall QA, and HotpotQA—all projected onto ClimbMix, a realistic corpus of over 550M documents.
cmass.tech
🔥Your next challenge for agentic search: 423 question–answer pairs spanning BrowseComp-Plus, DeepSearchQA, Total Recall QA, and HotpotQA—all projected onto ClimbMix, a realistic corpus of over 550M documents.
2
8
25
1,010
Unusual food take: Chhena poda by Omm sweets in Koramangala is literally the best cheesecake I've had in Bangalore, much better than actual western style cheesecakes here.
3
0
12
1,062
One more take: Hokkaido japanese restaurant in Indiranagar is very underrated: very authentic sushi, rolls and butter squid!
Much much better than Shiro in UB City (please avoid) and other sushi places in Indiranagar and around.
Much much better than Shiro in UB City (please avoid) and other sushi places in Indiranagar and around.
1
0
2
168
Nandan Thakur retweeted
If you're not properly checking the evaluations you are running and try to improve on them, there's a very large chance you're just fitting the bench biases/noise
1
2
11
923
I'm unfortunately missing #EMNLP2026! But our recent work we observe that Benchmarks such as FreshStack remain "fresh" overtime!
Do give the camera ready version a read: aclanthology.org/2026.findings-…
Do give the camera ready version a read: aclanthology.org/2026.findings-…
Technical documentation evolves rapidly with repository changes within weeks, but can IR benchmarks remain “fresh” over time?
In our new preprint, we stress-test retrieval models on FreshStack (LangChain) across two temporal snapshots: Oct 2024 vs Oct 2025.
Findings below 🧵👇
0
2
26
1,297
Nandan Thakur retweeted
@beirmug consistently bringing the "Fresh" into FreshStack 🔥
Recently revamped the FreshStack website and leaderboard! 🚀
Added newer models, including recent releases: Cohere Embed v5, Google's EmbeddingGemma2 & Perplexity late-interaction models.
Use the benchmark to pick robust retrieval models and findings! 👀
0
1
1
251
Nandan Thakur retweeted
I think I spy something new on there
Recently revamped the FreshStack website and leaderboard! 🚀
Added newer models, including recent releases: Cohere Embed v5, Google's EmbeddingGemma2 & Perplexity late-interaction models.
Use the benchmark to pick robust retrieval models and findings! 👀
4
4
64
3,942
Some important trends:
1️⃣ Nemotron-8B/Octen-8B are great open-weight robust dense models to use for retrieval usecases.
2️⃣ Cohere Embed v5 pro is a great closed-source embedding model.
3️⃣ Perplexity late-interaction models (some internal, some open) are great choices to pick.
1️⃣ Nemotron-8B/Octen-8B are great open-weight robust dense models to use for retrieval usecases.
2️⃣ Cohere Embed v5 pro is a great closed-source embedding model.
3️⃣ Perplexity late-interaction models (some internal, some open) are great choices to pick.
Recently revamped the FreshStack website and leaderboard! 🚀
Added newer models, including recent releases: Cohere Embed v5, Google's EmbeddingGemma2 & Perplexity late-interaction models.
Use the benchmark to pick robust retrieval models and findings! 👀
2
0
21
1,214