[feat] Add multi-turn format checking reward managers with refactored…

[albertimff]

… utilities

Files Modified:

• verl/experimental/agent_loop/tool_agent_loop.py: Enhanced to record full message history for format reward computation
• verl/experimental/reward/reward_loop/init.py: Added exports for new format check reward managers
• verl/workers/reward_manager/init.py: Added exports for new format check reward managers

Files Added:

• verl/experimental/reward/reward_loop/format_check_dapo.py: DAPO reward manager with format checking
• verl/experimental/reward/reward_loop/format_check_naive.py: Naive reward manager with format checking
• verl/utils/format_reward.py: Core format reward computation utilities with helper functions
• verl/workers/reward_manager/format_check_dapo.py: Worker implementation for DAPO format checking
• verl/workers/reward_manager/format_check_naive.py: Worker implementation for naive format checking

Purpose of Changes:

The format reward check manager was rewritten to address multi-turn conversation scenarios where the previous implementation would concatenate all subsequent assistant/tool messages, causing single-turn format checks to become ineffective.

The new solution extracts all assistant messages from the complete conversation history and performs individual format validation on each turn, ensuring proper format checking regardless of conversation length.

Implementation Details:

• tool_agent_loop.py: Modified to always record assistant messages and pass full message context via extra_fields
• format_reward.py: Provides extensible format validation with regex patterns for thinking/tool/answer tags, plus refactored utility functions:
  • compute_format_reward(): Core format validation logic
  • apply_format_reward_to_score(): Helper for scalar reward scores
  • apply_format_reward_to_tensor(): Helper for tensor reward processing
• Reward Managers: Refactored to use shared utility functions, reducing code duplication
• Both DAPO and naive reward manager architectures supported
• Full message history preservation enables per-turn format analysis

Key Features:

• Full message history preservation in agent loop for per-turn analysis
• Configurable format checking logic that users can customize
• Refactored utility functions for better code reuse
• Support for both DAPO and naive reward manager architectures
• No breaking API changes - fully backward compatible

What does this PR do?

Add concise overview of what this PR aims to achieve or accomplish. Reference related GitHub issues and PRs that help with the review.

Checklist Before Starting

• Search for similar PRs. Paste at least one query link here: ...
• Format the PR title as [{modules}] {type}: {description} (This will be checked by the CI)
  • {modules} include fsdp, megatron, sglang, vllm, rollout, trainer, ci, training_utils, recipe, hardware, deployment, ray, worker, single_controller, misc, perf, model, algo, env, tool, ckpt, doc, data, cfg, reward
  • If this PR involves multiple modules, separate them with , like [megatron, fsdp, doc]
  • {type} is in feat, fix, refactor, chore, test
  • If this PR breaks any API (CLI arguments, config, function signature, etc.), add [BREAKING] to the beginning of the title.
  • Example: [BREAKING][fsdp, megatron] feat: dynamic batching

Test

For changes that can not be tested by CI (e.g., algorithm implementation, new model support), validate by experiment(s) and show results like training curve plots, evaluation results, etc.

API and Usage Example

Demonstrate how the API changes if any, and provide usage example(s) if possible.

# Add code snippet or script demonstrating how to use this

Design & Code Changes

Demonstrate the high-level design if this PR is complex, and list the specific changes.

Checklist Before Submitting

Important

Please check all the following items before requesting a review, otherwise the reviewer might deprioritize this PR for review.

• Read the Contribute Guide.
• Apply pre-commit checks: pre-commit install && pre-commit run --all-files --show-diff-on-failure --color=always
• Add / Update the documentation.
• Add unit or end-to-end test(s) to the CI workflow to cover all the code. If not feasible, explain why: ...
• Once your PR is ready for CI, send a message in the ci-request channel in the verl Slack workspace. (If not accessible, please try the Feishu group (飞书群).)

[gemini-code-assist]

Code Review

This pull request introduces a robust mechanism for format checking in multi-turn conversations by refactoring the reward managers. The core logic is encapsulated in a new format_reward.py utility, which is then used by new DAPO and naive reward managers. The agent loop is also updated to ensure the full message history is available for this per-turn analysis. My review focuses on the correctness and robustness of the new format checking logic. I've identified a potential issue where leading or trailing whitespace in assistant messages could lead to incorrect format validation, and I've provided a suggestion to improve robustness.

[gemini-code-assist]

The current implementation for format validation using expected_pattern.match(content) is sensitive to any leading or trailing whitespace in the assistant's message content. Language models can sometimes generate extraneous whitespace, which would cause an otherwise correctly formatted message to fail validation and receive an incorrect penalty. This could negatively impact the training process by providing wrong reward signals.

To improve the robustness of the format check, I recommend stripping whitespace from the content before applying the regular expression match.

Suggested change

	if not expected_pattern.match(content):
	if not expected_pattern.match(content.strip()):