Add a minimal Verifiers agentic regeneration demo

[mgoin]

What this demonstrates

• run upstream agent environments through the stable Verifiers CLI
• keep the integration in a small UV project pinned to verifiers==0.2.0 and an exact research-environments commit
• provide Prolog, LiveCodeBench, and R2E-Gym example profiles
• convert native Verifiers traces into exact input_ids and loss_mask
• replay those exact IDs through vLLM for aligned hidden-state collection

The boundary is intentionally narrow: Verifiers owns environments, tools, Docker, turns, and scoring. Speculators does not add an environment wrapper or rollout loop.

Live demonstration

A Qwen3-8B Prolog run produced a six-turn trajectory with six assistant tool calls and five tool observations. The unsuccessful max_turns trace contained 5,708 exact tokens (4,914 sampled), and replay returned an exact token match with BF16 hidden states shaped (5708, 4, 4096).

The zero reward is retained intentionally because failed trajectories are still valid on-policy data.

Scope

This omits generated rollout artifacts and does not lock ordinary transitive Python dependencies. The compatibility boundary is expressed through the direct pins in pyproject.toml.

Validation

• 77 passed, 2 skipped
• Ruff passed
• all three Verifiers profiles pass CLI dry-run validation
• uv sync --project scripts/agentic_regeneration succeeds

[coderabbitai]

Important

Review skipped

Draft detected.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro

Run ID: 0e4cf18d-12ce-4e8d-8b62-369aed7f6106

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch feat/verifiers-agentic-regeneration

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[mergify]

The quality checks have failed. Please run make style and make quality under
the root directory to address the lint failures. You will need to install the
dev optional install to get the required linting packages:
https://github.com/vllm-project/speculators/blob/main/CONTRIBUTING.md

[mergify]

Merge Protections

🔴 1 of 1 protections blocking · waiting on 👀 reviews

	Protection	Waiting on
🔴	Require approval from approved reviewers list	👀 reviews

🔴 Require approval from approved reviewers list

Waiting for any of

• approved-reviews-by = dsikka
• approved-reviews-by = fynnsu
• approved-reviews-by = orestis-z
• approved-reviews-by = rahul-tuli
• approved-reviews-by = shanjiaz

This rule is failing.

All pull requests must have at least one approving review from a member of the approved reviewers list before merging.

• any of:
  • approved-reviews-by = dsikka
  • approved-reviews-by = fynnsu
  • approved-reviews-by = orestis-z
  • approved-reviews-by = rahul-tuli
  • approved-reviews-by = shanjiaz

[WindChimeRan]

I think this PR conflates several data pipeline scripts.

• scripts/launch_vllm.py this is for online training hidden state generation, so it’s very slow, and we should not use it for regen.
• regen actually uses scripts/response_regeneration/run_all.sh
• Prepare and replay exact IDs —> this is for off-policy data preparation… we shouldn’t include it for any regen training.

This is indeed a very confusing UX. 😅 But it's a pre-existing bad design.

I suggest we fix the UX first.

cc: @fynnsu @shanjiaz