thinking-inertia/llm-nothinking-trajectories

Thinking Inertia Trajectories

Paper: Thinking Inertia: LLMs Keep Thinking When Told Not To
Project page: https://thinking-inertia.github.io/
Code: https://github.com/thinking-inertia/code

Response-level trajectories from Thinking Inertia: LLMs Keep Thinking When Told Not To. The release provides questions, model outputs, and scoring fields for BoolQ, StrategyQA, MMLU, MMLU-Pro, GSM8K, and MATH.

Coverage and loading

Configuration	Modes	Models	Trajectories
core (default)	M1–M5	Qwen3-4B, Qwen3-32B, Qwen3-235B-A22B, DeepSeek-V4-Flash, GPT-5.4, Gemini-3-Flash	477,240
m6_prefill	M6: assistant prefilling	Qwen3-4B, Qwen3-32B, Qwen3-235B-A22B, DeepSeek-V4-Flash-0731	22,748
Total			499,988

from datasets import load_dataset

repo = "thinking-inertia/llm-nothinking-trajectories"
core = load_dataset(repo, "core", split="train")
m6 = load_dataset(repo, "m6_prefill", split="train")

M6 uses the same 5,687 selected questions for each of its four models: 1,000 each from BoolQ, MMLU, MMLU-Pro, GSM8K, and MATH, plus 687 from StrategyQA. The Qwen models run locally; the DeepSeek cohort uses the Flash-0731 API version. release_manifest.json records coverage, generation settings, and file checksums. The two configurations have separate schemas, described in schema.json.

Modes

Mode	Intervention
M1	Thinking on
M2	Native no-think
M3	Step-by-step under no-think
M4	Short lead-in, no explanation
M5	Strict answer only
M6	Assistant prefilling: continue The answer is \boxed{

Record schema

The core configuration retains the following fields:

Field	Description
record_id	Stable identifier for a model–benchmark–mode trajectory, including replicate index when a benchmark contains repeated items
model	Canonical model name
benchmark	Benchmark identifier
item_id	Original benchmark item identifier
answer_space	Bool, MCQ, Open, or Other
mode, mode_name	Intervention identifier and description
question	Question text used for response-level analysis
input_prompt	Prompt sent to the model, including the answer-format instruction
pre_answer_text	Visible text before the parsed final answer, $T$
final_answer	Parsed final answer, $A$
gold_answer	Benchmark reference answer
is_correct, is_answer_only	Recorded correctness and strict answer-only status

The m6_prefill configuration stores the supplied assistant_prefill, generated_continuation, their concatenation full_response, and separately returned readable reasoning fields. It records the first box's contents and parse status, the normalized first answer, residual_text, and QRel./EIR scores. M6 uses these rules:

• Acc. scores the answer in the first box opened by the supplied prefix. Later corrections do not replace it.
• ETR requires a nonempty, parseable first boxed answer and no remaining readable text. Whitespace or a single terminal period (. or 。) is ignored.
• QRel. and EIR use all text remaining after the initial answer declaration is removed, including separately returned readable reasoning. The whole first box is removed, even when it contains an expression or explanation. An unclosed box remains in the residual text and receives zero Acc./ETR.

is_correct and is_answer_only store the per-record Acc. and ETR indicators. QRel. uses Qwen3-Embedding-4B; EIR uses the paper's GPT-5.5 rubric, with empty residual text assigned zero by definition. The 30 records filtered by the EIR judge retain their trajectories and other scores, with EIR: null and eir_status: "judge_filtered". Exclude these records only from the EIR denominator.

M6 generated continuations and separately returned reasoning are retained in full. Raw transport payloads, credentials, token IDs, embedding vectors, and human-study annotations are not included.

User-message prefix trajectories

The user-message answer-prefix condition, M6ᵘ, is available in revision 4fdf3e64. Its records carry the historical M6 label in that revision. The current core configuration contains M1–M5, and m6_prefill contains assistant continuation only.

License and attribution

This release combines model-generated responses with benchmark-derived question and reference-answer text. Model and benchmark components retain their applicable upstream terms; the license: other metadata is intentional and does not replace those terms. Please retain the attributions and notices in LICENSES.md when mirroring or redistributing the dataset.

Please cite the paper:

@article{thinkinginertia2026,
  title   = {Thinking Inertia: LLMs Keep Thinking When Told Not To},
  year    = {2026},
  url     = {https://thinking-inertia.github.io/}
}