Qwen3.5-35B-A3B-Correct

Correct is the correctness-filtered arm of a four-arm controlled study of training-data interventions against over-reflection in web-search agents. It is a supervised fine-tune of Qwen/Qwen3.5-35B-A3B on the subset of the in-house web-search agent trajectory corpus (shared with the Asis baseline) whose final answers match the reference under deterministic answer matching. No trajectory editing is performed — selection is purely by outcome.

This repository contains the full serving weights (13 safetensors shards, ~65 GB, bf16) plus tokenizer, config, and chat template.

Model description

• Architecture: Qwen3.5-35B-A3B — a mixture-of-experts (MoE) decoder-only transformer with ~35B total parameters and ~3B active per token. Architecture, tokenizer, and chat/tool-calling template are unchanged from the base model.
• Training context length: 131,072 tokens (full agent trajectories, multi-turn with interleaved tool calls and tool results).
• Behavior domain: multi-turn web-browsing agentic tool use (search / page-open / in-page find style tool surface) with explicit reasoning.

Training procedure

Full-parameter supervised fine-tuning on a Megatron-based distributed training stack. Identical recipe to the sibling arms; only the data intervention differs.

Hyperparameter	Value
Epochs	2
Global batch size	128
Learning rate	5e-6, cosine decay to 5e-7
Sequence length	131,072
Objective	token-level cross-entropy on assistant turns (trajectory imitation)

Intervention: correctness filtering (outcome-conditioned selection)

Each trajectory's final answer is checked against its reference answer with a deterministic matching rule (no model-based judging), and only trajectories that pass are kept. This is the classic rejection-sampling-style selection baseline of the study: it conditions on outcome only, and is agnostic to how the agent behaved along the way. In particular, a correct-but-wasteful trajectory (one that reaches the right answer and then keeps searching) survives this filter — which is exactly the contrast the four-arm design probes against the process-conditioned Drop and Repair arms and their A–G over-reflection taxonomy (post-answer verification loops, redundant re-search after sufficient evidence, ungrounded reasoning).

Arm	Intervention class
Asis	none — unmodified trajectory imitation baseline
Drop	typed whole-trajectory deletion of pathological over-reflection types
Repair	taxonomy-guided surgical repair (truncate post-evidence over-search tails; remove ungrounded trajectories)
Correct (this model)	correctness-filtered subset (deterministic answer matching)

An RL continuation of this arm is released as Qwen3.5-35B-A3B-Correct-GRPO.

Intended use

Research artifact. Intended for studying outcome-conditioned data selection as a baseline against process-conditioned interventions for over-reflection, stopping behavior, and search efficiency in tool-using web agents, and as an RL initialization. It expects a browsing-agent tool surface compatible with its training format (function-calling tools for web search, page opening, and in-page lookup).

Limitations

• Behavior is tuned for the browsing-agent tool surface described above; it is not a general-purpose assistant fine-tune.
• Outcome filtering keeps correct-but-inefficient demonstrations, so over-reflection pathologies are not removed by construction.
• Correctness filtering shrinks the training set relative to the Asis baseline; arm-to-arm differences confound data quality with quantity by construction (part of what the study measures).
• No additional safety alignment was applied beyond what the base model provides.
• Trained primarily for English-language reasoning traces.