Thinking cannot be switched off, and the think tags come back inside the text: litert-lm 0.18.0 (measured 2026-10-13)

README.md

@@ -32,6 +32,8 @@ base_model_relation: quantized
 
 Graded with `--max-tokens 2048`: this is a reasoning model and emits a `<think>` block before answering, so a smaller budget scores it as wrong when it has simply not finished thinking.
 
+**Thinking cannot be switched off, and the think tags come back inside the text.** On litert-lm 0.18.0 `ThinkingConfig(enable_thinking=False)` has no effect on `TwIL-LM3_int4.litertlm`: with a 400-token cap the model's `<think>…</think>` reasoning fills the cap and the answer is cut (measured 2026-10-13). The bundle declares no thought channel, so the tags appear in the message text. Set `thinking_token_budget` (300 answered 17 × 23 under a 700-token cap) and a cap above it.
+
 ### The formal-logic claim: measured against the base, and no difference detected
 
 Upstream describes TwIL-LM3 as specialised for formal logic. The eight-question gate is general knowledge and cannot see that claim in either direction, so it was tested separately: **12 items with one defensible answer each** (syllogisms, modus tollens, affirming the consequent, contrapositive, De Morgan, quantifier negation, disjunctive syllogism, transitivity, necessary vs sufficient, biconditional, vacuous truth), run against **this model and its base [SmolLM3-3B](https://huggingface.co/HuggingFaceTB/SmolLM3-3B) on identical items, identical flags, identical runtime, and identically-converted bundles** (same rail, same recipe, both 2.00 GB, both duplication-free).