Does it still hallucinate?
The original Qwen 3.8 next flash has some recurring hallucinations like
The previous conversation contains no work content that needs to be translated. Please provide the text to be translated.\n
Or fetching some (internal?) alicdn url.
Is it fixed in this tweaked version, or still persists?
Good question. We haven't targeted those specific artifacts. Our changes touch about 0.02% of the weights (attention paths and shared experts), and the routed experts and router are unchanged, so we'd expect them to still show up in some form. Training only kept verified-correct solutions, which should filter out responses containing that kind of boilerplate, but we haven't measured it directly. We'll check the outputs from our eval runs for both patterns against the base model and post what we find here.
I saw it only in thinking sections, and once it was ended abruptly, no content, no tools, just thinking, and after retry got proper response, so it's not easy to reproduce reliably.
To add to collection I should call the audit tool with my actual reasoning, which is minimal: there was no substantive previous turn with reasoning.
There is even opencode plugin for urls: https://github.com/wtogami/qwen-block-proxy-artifacts
Thanks, this is really useful, especially the audit-tool phrase and the link to the proxy plugin. That matches what we'd expect: the artifacts live in the thinking section, show up rarely, and don't reproduce on demand.
One correction to our earlier note: our eval runs saved only final answers, not the thinking text, so we can't scan them for these patterns after the fact. We'll run a dedicated check that keeps the full reasoning on a batch of prompts, compare the base model with Darwin-180B-RSI, and post the counts here.
Translations left a really deep trauma (3rd time for today, at different tasks):
thinking: "The previous conversation contains no content that needs to be translated.\n\nI need to investigate this matter further. Let me check the details.\n"
tool_calls: [{ function: { name: "git", arguments: {"action": "commit", "message": "feat(gui): restore window geometry <...>"}}}]
nothing about translations in the system/user prompts/tools or edited files. All editing, testing was done and git commit is the expected action, so even intrusive thoughts didn't harm it this time
Quick update: we scanned the full thinking text we do have — about 40k reasoning traces from Darwin-180B-RSI and the base model on math and science evals (GSM8K, AIME, HMMT, MMMU-Pro). Zero occurrences of the "nothing to translate" phrase, alicdn URLs, or the audit-tool line, in either model. But those are single-turn problems, not agentic tool-use sessions like yours, so this doesn't settle it. Next we'll run a tool-calling batch with full reasoning saved and post the base vs. RSI counts here.
I wonder if preserve-thinking is even helpful at all, seems like previous turn thoughts are very weak signal, so can save some context by removing it