Researchers report 16.9% lower reasoning-token use with a training-free controller across six matched model–task settings. The evaluation used open-weight Qwen-family checkpoints, and the reported math results paired those savings with accuracy changes of just -0.1 and +0.1 percentage points on two splits.
A controller aimed at reflection
The system works in the model’s residual stream, the running internal representation passed through its layers. It estimates a direction from reflective and non-reflective states, uses principal-component analysis (PCA) to denoise it, removes overlap with a pooled direction for general reasoning, identifies stable layers through calibration, and applies bounded projection removal during decoding. In plain terms, the intervention is designed to reduce a measured reflection-related component without treating every reasoning signal as disposable.
How the comparison was built
The document is an arXiv v1 preprint dated 26 Aug 2026. Its evaluation covered MATH-500 and GPQA-Diamond using three Qwen-family open-weight checkpoints.
For direction construction, the authors used 150 MATH-500 records, numbered 0 through 149, including 20 nested calibration inputs. They also reported a record-disjoint MATH-350 evaluation using records 150 through 499, with zero exact overlap with GPQA-Diamond.
The protocol measured mean thinking tokens and token reduction, strict task accuracy, how often generation collapsed into repetition loops, and unparsable rates. To assess accuracy equivalence, it used a cluster-level paired TOST, with a margin of plus or minus one percentage point.
Math savings were the clearest result
On MATH-500, the reported thinking-token reduction was 21.8%, with an accuracy change of -0.1 percentage points. On record-disjoint MATH-350, token reduction was 23.4% and accuracy changed by +0.1 points. The second result used records outside the 150 inputs used to construct the direction.
The paired equivalence test supported both math comparisons within the one-point margin. For MATH-500, the accuracy difference was -0.13 points, with a 90% confidence interval from -0.60 to +0.33 and p = 0.0012. For MATH-350, it was +0.10 points, with an interval from -0.51 to +0.70 and p = 0.0074.
Transfer brought a more cautious answer
GPQA-Diamond gave a more cautious result. The same controller transferred without refitting and showed a 21.0% token reduction, compared with 17.7% for the reproduced ReflCtrl comparator, while mean accuracy matched it. The separate equivalence test did not pass: the estimated accuracy difference was -1.52 points, with a reported interval from -4.38 to +1.18.
A prospective pilot on 500 fresh, overlap-filtered MATH-train problems produced another mixed result. Thinking tokens were 21.0% lower, with a one-sided 95% lower bound of 17.6%, above the pre-registered 10% target. Accuracy changed by -0.60 points, with a one-sided lower bound of -1.69, so the pilot did not meet the one-point accuracy criterion.
Small design choices changed the trade-off
Technical ablations changed the balance between savings and accuracy. On a fit-disjoint GPQA-Diamond ablation, the full method had a 20.9% token reduction. Without PCA, the reduction was 15.8%; without orthogonalization, accuracy was 67.2%.
Steering all candidate layers added only 1.2 percentage points to token reduction but lowered accuracy by 4.0 points. An additive steering update increased token use by 27.8%. The calibration step retained 14 of 44 candidate layers for the 30B checkpoint, 2 of 34 for the 8B checkpoint, and six layers for the 32B checkpoint.
The selected operating point was alpha = 0.7. The authors say it captured most of the token saving with zero or near-zero collapse and offered the best overall trade-off, treating alpha as a deployment-time control rather than a rule that stronger intervention is always better.
The gains varied by checkpoint
The savings varied across the evaluated checkpoints. With Qwen3-8B, reductions were 6.6% on MATH-500 and 9.6% on GPQA-Diamond; with QwQ-32B, they were 26.4% and 15.8%, respectively. In all four reported settings, the reductions exceeded the reproduced ReflCtrl comparator.
The overlap analysis offered a narrower technical check. Median overlap was 0.0332 for the raw direction and 0.0385 after PCA, compared with 0.00042 after orthogonalization and 0.00048 for the Random reference. This supports the claim that the measured overlap with the general-reasoning direction was reduced; it does not establish complete separation of reflection from all reasoning.
What the results do not settle
One important caveat is that the MATH-500 result was not fully record-disjoint: 150 of its records, 0 through 149, were used in direction construction. The separate MATH-350 test used records 150 through 499, but the study still evaluated only three Qwen-family checkpoints, so it does not establish transfer beyond that family.
Taken together, the preprint supports a model- and task-specific token-saving setting: the reported math results met the one-point equivalence criterion, while GPQA-Diamond and the prospective pilot did not. The result is therefore best read as evidence for a controllable token-accuracy trade-off in the tested Qwen-family settings, not as a general guarantee.
Paper data and sources
Original title: Reflection Steering: Disentangling Reflection from Reasoning in Activation Space for Token-Efficient Inference
Authors: Jiarui Hu, Zhiyuan Wen, Xiaoyun Liu et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text