Qwen3.5-35B-A3B-Asis
Asis is the unmodified-baseline 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 an in-house corpus of web-search agent trajectories, used as-is — no filtering, deletion, or editing. The three sibling arms (Drop, Repair, Correct) apply different intervention classes to the same corpus, so this model serves as the common reference point.
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.
| 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) |
The four-arm design
All four arms start from the same base model and the same in-house web-search trajectory corpus; they differ only in what is done to the corpus before training. The interventions are organized around an A–G taxonomy of over-reflection behaviors in browsing agents (e.g., post-answer verification loops, redundant re-search after sufficient evidence has been gathered, ungrounded reasoning not supported by retrieved content), described in the associated research.
| Arm | Intervention class |
|---|---|
| Asis (this model) | 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 | correctness-filtered subset (deterministic answer matching) |
An RL continuation of this arm is released as Qwen3.5-35B-A3B-Asis-GRPO.
Intended use
Research artifact. Intended for studying over-reflection, stopping behavior, and search efficiency in tool-using web agents, and as a controlled SFT baseline / 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.
- As the unmodified baseline arm, this model inherits the over-reflection pathologies present in raw agent trajectories (e.g., continued searching after the answer has already been stated); that is by design — it is the control condition.
- No additional safety alignment was applied beyond what the base model provides.
- Trained primarily for English-language reasoning traces.