Qwen3.5-35B-A3B-Repair

Repair is the surgical-repair 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 same in-house web-search agent trajectory corpus as the Asis baseline, with trajectories edited rather than merely selected: post-evidence over-search tails are truncated away, and trajectories whose reasoning is ungrounded in retrieved content are removed.

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: taxonomy-guided surgical repair

Guided by the study's A–G taxonomy of over-reflection behaviors in browsing agents (e.g., post-answer verification loops, redundant re-search after sufficient evidence, ungrounded reasoning), this arm applies two operations:

1. Truncation of post-evidence over-search tails. When a trajectory has already gathered sufficient evidence for its final answer but continues searching, the redundant tail is cut and the trajectory ends at the natural stopping point. The grounded prefix — the useful demonstration — is retained.
2. Removal of ungrounded trajectories. Trajectories whose reasoning or answer is not supported by the content actually retrieved during the episode are dropped, since no truncation point can make them sound.

This is the finest-grained intervention class in the study: unlike the Drop arm it salvages the healthy portion of pathological trajectories, and unlike the Correct arm it conditions on process quality rather than outcome correctness.

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

Intended use

Research artifact. Intended for studying whether editing demonstrations at the point of pathology (rather than deleting or outcome-filtering them) transfers a better stopping policy to the trained agent. 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.
• Truncated demonstrations end at an imposed stopping point; the model may under-verify in settings where additional confirmation is genuinely warranted.
• No additional safety alignment was applied beyond what the base model provides.
• Trained primarily for English-language reasoning traces.