Nope, it was our Israel team.

AstaBench with abstract measurement icons

Agent benchmarks don't measure true *AI* advances

We built one that's hard & trustworthy:
👉 AstaBench tests agents w/ *standardized tools* on 2400+ scientific research problems
👉 SOTA results across 22 agent *classes*
👉 AgentBaselines agents suite

🆕 arxiv.org/abs/2510.21652

🧵👇

Screenshot of the Ai2 Paper Finder interface

Meet Ai2 Paper Finder, an LLM-powered literature search system.

Searching for relevant work is a multi-step process that requires iteration. Paper Finder mimics this workflow — and helps researchers find more papers than ever 🔍

Ai2 ScholarQA logo, with a red sign that says "Updated!"

Hope you’re enjoying Ai2 ScholarQA as your literature review helper 🥳 We’re excited to share some updates:

🗂️ You can now sign in via Google to save your query history across devices and browsers.
📚 We added 108M+ paper abstracts to our corpus - expect to get even better responses!

More below…

Ai2 ScholarQA logo

Can AI really help with literature reviews? 🧐
Meet Ai2 ScholarQA, an experimental solution that allows you to ask questions that require multiple scientific papers to answer. It gives more in-depth and contextual answers with table comparisons and expandable sections 💡
Try it now: scholarqa.allen.ai

The CEO using AI to fight insurance-claim denials says he wants to remove the 'fearfulness' around getting sick Claimable has helped patients file hundreds of health-insurance appeals. Its CEO says its success rate of overturning denials is about 85%.

Building off the story I shared yesterday about fighting potential Insurance Company AI with AI: Claimable uses AI to tackle insurance claim denials. With an 85% success rate, it generates tailored appeals via clinical research and policy analysis. 🩺 #HealthPolicy

(3) they also studied multiple rounds of the above. iterative self improvement. saturation happens after 2 or 3 rounds. I'm surprised it's not 1!

(4) Ensemble Heuristic: Simple verification ensemble heuristics can improve performance

6/6

(2) CoT Verification is More Stable than MC: "Some MC verification incurs non-positive gap even for medium-sized models such as Qwen-1.5 14/32B, while CoT verification always has a positive gap for medium/large-sized models"

5/n

Results
(1) Small Models can not Self-improve. Models such as Qwen-1.5, 0.5B, Qwen-2 0.5B and Llama-2 7B, gap(f ) is non-positive for nearly all verification methods, even though the models have non-trivial generation accuracy

4/n

(3) Then they compute the gap which is the average accuracy diff between the filtered generations (those that are correct after step 2 according to self-verification) and the original 128 responses.

3/n

(2) For each of the 128, they sample one verification for each response of one of 3 styles: (a) correct vs incorrect, (b) CoT + score 1 to 10, or (c) "Tournament" style, which you can find in the paper.

2/n

Super awesome paper that directly addresses questions I've had for a while: arxiv.org/abs/2412.02674

Their experiments:

(1) They get 128 responses from a LLM for some prompt. p = 0.9, t = 0.7, max length of 512 and 4-shot in-context samples

1/n

Check out our #NeurIPS2024 poster (presented by my collaborators Jacob Chen and Rohit Bhattacharya) about “Proximal Causal Inference With Text Data” at 5:30pm tomorrow (Weds)!

neurips.cc/virtual/2024...

Windows has issue:

Person: fuck this I'm going to Linux

Narrator: and they quickly learned to hate two operating systems.

Thanks!

If you know papers or blog posts that address these, I'd be happy to have the links. Thanks!

(7) Others found a good recipe for distilling: first fine-tune the biggest model on small gold data, then use that fine-tuned model to make silver data. Does that work for IR distilling? If we fine-tune a 405b before using it as the silver data source, what should we use as gold? How much do I need?

(6) You can get better LLM labels if you do all pair comparisons on the passage set (citation needed, but I read a few papers showing this). Obviously much more expensive. Should I spend my fixed computer/money budget on all-pairs O(few_queries * passages^2) or pointwise O(more_queries * passages)?

(5) Does the type of base model to be distilled matter much? Should I distill roberta-large or some modern 0.5b LM?

(4) From our experience at AI2, LLM-generated search queries are weirdly out of distribution and non-human in various ways. Does this matter? Do we have to get human queries?

(3) Can we do better than human labeled data because we have no gaps in the labels? And can get more data at will?

(2) How to distill well? Do we use the same loss functions that we used when obtaining gold data from human labelers?

(1) Say I have 10000 queries and 100 passages/docs for each query, labeled or ranked by the best LLM (with optimized prompt or fine-tuning), how close can we get to the LLM's performance? Result is a plot with number of distilled model parameters on the x-axis and NDCG vs LLM on y-axis.

Here are some research questions I'd like to get answers to. We are using LLMs to make training data for smaller, portable search or retrieval relevance models. (thread)

Image of an email from a student asking if sources "from the late 1900s" are acceptable.

I will never recover from this student email.

#mlsky

www.semanticscholar.org/paper/Ground...

I really like this paper. They study whether LLMs do reasonable things like ask follow-up questions and acknowledge what the users are saying. The answer is "not really".

bsky.app/profile/did:...