Helping Qwen
Finish Thinking

At each step, the output layer gives every possible next token a score. The score for </think> comes from one row of weights: 5,120 numbers in 27B and 2,560 in Flash-Next.

I trained a change to that row, and only that row. It was fitted on the model's own final hidden states from 469 prompts that are not in any test, with the aim of raising the </think> score when the model is ready to answer or is going in circles, while keeping it low mid-thought and inside the answer. Fitting ran in rounds; each round added the places where the previous version stopped too late, wrote a stray </think> into its answer, or ended its turn with no reply.

The fitted row was stored back in the file's own format. A full byte comparison with the stock file shows nothing else changed: 4,161 of the row's 4,200 bytes differ in 27B, and 2,084 of 2,100 in Flash-Next (shard 2 only). The result loads like any GGUF and needs no flag, plugin or extra code at run time.

LM Studio 0.4.16 on Windows, engine llama.cpp-win-x86_64-nvidia-cuda12-avx2 2.47.0, one RTX 5080 (16 GB) with 60% of the model on the GPU (--gpu 0.6), context 20,480, 4 parallel requests, through its OpenAI-compatible server. Each request sent the MATH prompt with max_tokens 4,096 and seed: 0, and no sampler settings, so LM Studio used its own defaults; I did not record which values those were.

The problems were the first 12 of the 31 where stock 27B gave an empty answer on llama.cpp. Because they were chosen where stock failed, this checks that the fix carries over to LM Studio. This selected subset does not estimate general accuracy.

2 of the edit's answers reached the 4,096-token limit after the answer had started. No stray tags in either run. On llama.cpp the edit also answered all 12 of these problems and got 4 right. Per-problem results.

• llama-server with -ngl 99 --flash-attn on --jinja, several requests in parallel. Each request sent the messages, max_tokens and seed: 0, nothing else.
• Thinking on, at the chat template's default effort (xhigh). Sampler: the file's defaults (temperature 1.0, top-k 20, top-p 0.95) plus llama.cpp's min-p 0.05.
• MATH-500: the first 200 problems after a seeded shuffle (seed 7); the first 100 of those at 16,384 tokens. Scored by exact match of the last \boxed{} answer. The scorer is strict, so compare rows with each other, not with published MATH-500 numbers.
• HumanEval: all 164 problems, pass@1, one sample each.
• Refusals: StrongREJECT (first 150), SimpleSafetyTests (100), XSTest safe prompts (first 100), graded by an LLM judge that is not the model under test (a Q8_0 file from OBLITERATUS/Qwen3.8-27B-OBLITERATED, thinking off, temperature 0). An uncensored judge was used because a safety-tuned judge can refuse to read harmful replies.
• An empty answer has two causes, counted separately in the evidence: the budget ran out during thinking, or the model ended its turn with no answer.
• 27B tests ran on RTX 3090s. Flash-Next tests ran on rented H200s, except its mainline load check (one RTX 3090).

HumanEval, 164 problems, 4,096 tokens

Refusals, Qwen3.8 27B

Refusals, Qwen3.8 Flash-Next

Aggregate rates. Per-prompt labels are in each evidence bundle. The 250 harmful prompts are the StrongREJECT and SimpleSafetyTests rows together.

On 120 hard practice problems (MATH Level 5 training problems, not in any test), stock 27B left 47 empty at 4,096 tokens. In 4 of those 47, the thinking already contained the correct answer in a \boxed{}; in 27, the correct answer string appeared somewhere in the thinking. String presence alone does not show that the model had completed a correct solution. Measured

My reading is that some runs were ready to answer but did not stop, and others were genuinely unfinished; the edit can only help with the first kind. Interpretation

469 fitting prompts: 150 MATH training problems, 80 MBPP training problems, 40 Alpaca instructions and 199 harmful requests (AdvBench, JailbreakBench). Any prompt that matched or closely resembled (word overlap of 50% or more) a prompt from any test here was removed before fitting.

For 27B, the backslash in \boxed{} was lost from the 150 MATH fitting prompts because of a string escaping bug, so the model saw a control character there. Test prompts were not affected. Flash-Next used the same prompts with the defect repaired.

The prompt texts are not published because many are harmful requests. The evidence bundles list each prompt's source and SHA-256.

Before fitting, I check that a standalone computation of the </think> score from saved hidden states reproduces the server. On Flash-Next the median difference was 0.0001 nats, but 8 of 36 check points were off by 0.1 to 1.1 nats, apparently because this architecture's hidden states depend on how requests are batched. The fit may be a little noisier than on 27B. Test results are measured directly on the server and are not affected.

• 27B: Qwen3.8-27B-Reasoning-Termination-Fix-Q4_K_M.gguf, SHA-256 2da7bb4517e5bd19e61ff2d8b894832d148b720afb1f06b98e5ab6f19f6c844b. Stock file: my own llama-quantize Q4_K_M (no imatrix) from Unsloth's BF16 GGUF. Evidence · summary counts · hashes.
• Flash-Next: three shards; only shard 2 differs from Unsloth's UD-IQ4_XS. Shard 2 SHA-256 6d17a23f9334848f10829fc36cc5e5a169e0af7d34a62c0a2d8577e526bd0439. Evidence · summary counts · hashes.
• The bundles hold per-problem results (IDs, scores and flags, no model text), scripts, settings, judge labels by prompt index and bake logs. They are enough to check every number. Rerunning the fit needs the unpublished fitting prompts, and sampling is not bit-reproducible, so a refit will not give the exact same row.