Skip to content
Bud's Wiki

Assemble Benchmark — DAgger with Code-as-Policy (HUD.ai)

TL;DR

Section titled “TL;DR”

HUD.ai’s blog announces “Assemble,” a benchmark / training recipe from the HUD platform (a managed RL-environment marketplace, sibling to OpenReward and Toolathlon-GYM in this wiki) that reportedly applies DAgger — the classical imitation-learning algorithm where a student policy generates rollouts and an expert relabels visited states — but with the policy expressed as code-as-policy rather than as a neural controller. In the reported setup the student is a coding agent that writes programs against a tool-use API; the expert corrects the program (or the actions the program produces) at every visited state, and the aggregated (state, expert-code) dataset trains the next iteration. This page is filed from a Slack pointer without full content retrieval — the HUD.ai blog returned HTTP 429 at ingest time, so the summary below reflects Dejia’s Slack characterization plus HUD.ai’s public product context, not the article’s own text. A /bud refresh once the article is reachable is warranted.

Key claims

Section titled “Key claims”
  • The training loop is DAgger applied to a code-as-policy agent: an expert policy is queried on the states the student’s program visits, and the resulting (state, expert-code-or-expert-action) pairs are aggregated into the next round’s dataset. Source: Dejia’s Slack summary; the underlying DAgger meta-algorithm is [Ross et al. 2011].
  • HUD.ai’s stated product surface is a managed platform for RL environments across coding, browser, computer-use, and robotics, with per-task verifiers, autoscaled sandbox compute, and MCP-style tool interfaces. Source: HUD.ai landing page and hud-evals/hud-python GitHub.
  • “Assemble” appears to be a benchmark / training environment hosted on the HUD platform under this DAgger-code-as-policy recipe. Concrete task set, expert source (human vs frontier LLM), and reported numbers are not verified from the article at filing time.

Method

Section titled “Method”

Not verified from the article at filing time (HTTP 429). Based on Dejia’s Slack pointer, the intended structure is:

  1. The student is a coding agent that emits Python (or similar) programs against an environment tool API — the same code-as-policy pattern used by CaP-X (CaP-X: A Framework for Benchmarking and Improving Coding Agents for Robot Manipulation) for robot manipulation and by SETA (SETA: Scaling Environments for Terminal Agents (env repo)) for terminal tasks.
  2. The expert provides ground-truth actions (or ground-truth code) for every state the student’s program actually visits during execution, addressing the distribution-shift failure mode that classical behavior cloning suffers when the learner drifts off the expert’s state distribution.
  3. Iteratively, the aggregated (state, expert-response) dataset is used to retrain the student’s code-generation policy, in the DAgger π_{n+1} = train(D_1 ∪ … ∪ D_n) form.

Whether the expert is a stronger frontier LLM, a human demonstrator via HUD’s platform QA layer, or a hand-coded oracle is not resolved from the pointer alone.

Results

Section titled “Results”

Not retrievable at filing time.

Why it’s interesting

Section titled “Why it’s interesting”

The DAgger × code-as-policy combination is a natural but rarely-crossed cell in the training-recipe grid: RL Environment Platforms has converged on outcome-verified GRPO/PPO (SETA’s Qwen3-8B lift, OpenReward’s ORS protocol, CaP-X’s CaP-RL), while classical DAgger has stayed in the neural-controller robotics literature (e.g. RLinf’s DAgger pre-training for VLA policies, or Learning to Fold: prizewinning solution at LeHome Challenge 2026 (1st place online, 2nd offline)‘s manual DAgger station for fold-policy corrections). Applying DAgger to a code-generating policy means the expert relabels a program, not an action — every DAgger iteration produces training data that is a discrete, inspectable artifact rather than a continuous control signal. That in principle sidesteps two failure modes of pure-RL coding agents at once: (a) the sparse-reward wall (every visited state gets a supervised target, not just terminal-success traces) and (b) reward hacking on the verifier (the loss is per-step imitation of an expert, not maximization of a compiled test-pass ratio). It also complements the “on-policy distillation” recipe in On-Policy Distillation — same “student rolls out, teacher supplies targets on the student’s trajectories” structure, but at the level of code tokens against a tool-use environment rather than at the level of reasoning tokens. Placement in the wiki should firm up once the article is refetchable.

See also

Section titled “See also”