←
laxman3,369 posts
laxman avatar

laxman

@llmluthor

20 • Scaling RL envs thru AWMs • It is our moral obligation to make an impact in the world proportionate to our potential • shipping bandr.ai

◷ Joined September 2022▧ 183 Photos and videos
laxman avatar
laxman@llmluthor·22h

We might’ve killed retraining for harness changes. Our model broke on a new harness. We didn’t retrain it. We edited 5,064 weights. Out of 4 billion. 20.1% -> 88.2% No gradient descent. Old-harness accuracy didn’t move. DAWE(Data-Attributed Weight-Editing) is fully open 🔥

1:10
Jack Lu ✈️ NeurIPS 26 avatar
⟳ laxman retweeted
Jack Lu ✈️ NeurIPS 26@Jacklu_me·21h

Interesting work! “edits to one or two weight rows" Let the crusade of sparse edits continue.

We might’ve killed retraining for harness changes. Our model broke on a new harness. We didn’t retrain it. We edited 5,064 weights. Out of 4 billion. 20.1% -> 88.2% No gradient descent. Old-harness accuracy didn’t move. DAWE(Data-Attributed Weight-Editing) is fully open 🔥

1:10
Pierfrancesco Beneventano avatar
⟳ laxman retweeted
Pierfrancesco Beneventano@PierBeneventano·21h

As a gradient-based optimization person, this new wave of work (rocked by @rikiyatakehi and @lateinteraction) is mindblowing. Can’t wait to see what people do with it. Lowkey WSP twitter is my new goto tv-series with new episode getting released every day

We might’ve killed retraining for harness changes. Our model broke on a new harness. We didn’t retrain it. We edited 5,064 weights. Out of 4 billion. 20.1% -> 88.2% No gradient descent. Old-harness accuracy didn’t move. DAWE(Data-Attributed Weight-Editing) is fully open 🔥

1:10
Siva avatar
⟳ laxman retweeted
Siva@ergodicthought·10h

Very clever. Another manifestation of how you use a "world model" to attribute outcomes to specific control inputs, and solve counterfactual reasoning.

We might’ve killed retraining for harness changes. Our model broke on a new harness. We didn’t retrain it. We edited 5,064 weights. Out of 4 billion. 20.1% -> 88.2% No gradient descent. Old-harness accuracy didn’t move. DAWE(Data-Attributed Weight-Editing) is fully open 🔥

1:10
Rikiya Takehi avatar
⟳ laxman retweeted
Rikiya Takehi@rikiyatakehi·21h

Really cool use of weight space programming!

We might’ve killed retraining for harness changes. Our model broke on a new harness. We didn’t retrain it. We edited 5,064 weights. Out of 4 billion. 20.1% -> 88.2% No gradient descent. Old-harness accuracy didn’t move. DAWE(Data-Attributed Weight-Editing) is fully open 🔥

1:10
laxman avatar
laxman@llmluthor·22h

Setup: Qwen3.5-4B, LoRA SFT (rank 32, merged) on 595 BFCL items under harness v1. The new harness requires plan-first replies. Eval: 204 held-out BFCL items (simple, multiple and parallel calls). Single runs, our own scorer.

Post attachment
laxman avatar
laxman@llmluthor·22h

You bring your data and your model. An agent(astra/fable) reads the failing traces, finds where in the model the failure lives and fits a closed-form edit to just those weights. It then scores the patch on both the new and old harness before accepting it.

Post attachment
laxman avatar
laxman@llmluthor·22h

How DAWE works: run the model on the new harness and capture its internal states at the exact decisions where it fails(the data part). Then solve for the weight edit in closed form (regression). No training loop. 5,064 weights changed.

Post attachment
laxman avatar
laxman@llmluthor·Nov 14

the goal is to make this a continuous loop. agents run in production → traces become verifiable tasks → models get evaluated and post-trained. we're getting closer to agents building their own gyms :) Mint is open source! github.com/bandr-ai/bandits

Post attachment
laxman avatar
laxman@llmluthor·Nov 14

we removed the task IDs and answers from 60 τ³-bench traces across 20 tasks and 11 models. Mint recovered all 20 tasks with no incorrect merges or splits. the interesting part was that its grader agreed with the original τ³ grader on 17/18 verdicts in our earlier 6-task test. it even caught an agent saying "I'll transfer you" without actually calling the transfer tool lol.

Post attachment
laxman avatar
laxman@llmluthor·Nov 14

Mint reads agent traces, figures out the underlying tasks, reconstructs them using the available environment and builds outcome verifiers. each task is validated in Harbor before it can be used for evals, rejection sampling, OPD or RL.

Post attachment
laxman avatar
laxman@llmluthor·Nov 14

most production traces are used for debugging and observability. but they contain actual user requests, tool calls, failures and corrections. we wanted to turn these into reusable tasks that agents can be evaluated and trained on, instead of manually creating everything from scratch. open source: github.com/bandr-ai/bandits Free to use in: app.bandr.ai

Post attachment