The Halt Vector: Internalizing a Causal Steering Intervention for Efficient Reasoning

Dylan Jayabahu; Tinuade Adeleke

Source-linked AI summary

The Halt Vector: Internalizing a Causal Steering Intervention for Efficient Reasoning

Dylan Jayabahu, Tinuade Adeleke

arXiv:2608.28859v1Aug 28, 2026cs.LGcs.AIcs.CL

TL;DR

Reasoning models can continue thinking after their answer is known, while global shortening objectives cannot adapt the removable slack to each problem. This paper internalizes a causal layer-18 halt vector into the weights using full-vector reconstruction, yielding shorter thinking at held accuracy without reinforcement learning. The method also addresses non-termination, but the paper claims a mechanism and deployment advantage rather than superiority on the raw compression-accuracy frontier.

Problem
Reasoning models overthink after their answer is known, and outcome-level shortening objectives do not provide a per-problem adaptive halt.
Method
The paper constructs a difference-of-means halt vector, identifies its causal stopping effect, and internalizes it at layer 18 by reconstructing the full steered activation with off-axis dimensions pinned.
Results
About 24% less thinking is achieved at held accuracy across five unseen benchmarks after fitting from 24 problems without reinforcement learning.
Takeaways & Limitations
The resulting hook-free halt is per-problem adaptive, tracks removable slack, and closes a difficulty-growing non-termination pathology.
Takeaways & Limitations
The study is confined to DeepSeek-R1-Distill-Qwen-7B, and its offline forced-answer training target requires the gold answer and a base-model probe.

Abstract

from arXiv · show

Reasoning models do not stop when they know the answer. On DeepSeek-R1-Distill-Qwen-7B the chain of thought runs about twice as long as the model's own answer probability takes to settle, and how much of that excess is removable varies from problem to problem, so a global length penalty cannot take it out. We take it out by internalizing a causal interpretability finding into the weights. The mechanism is a halt vector: a difference-of-means direction at layer 18 of this model whose steering strength controls how long it thinks, while a replicated value axis does nothing. Installing that intervention in the weights is harder than it looks. Maximizing the scalar projection onto the direction corrupts the off-axis dimensions a frozen downstream reader depends on, and generation gets longer instead of shorter; what works is reconstructing the whole steered activation with those dimensions pinned to their natural values. Fit from 24 problems and no reinforcement learning, the halt removes about a quarter of the thinking at held accuracy across five unseen benchmarks, and the cut tracks each problem's own removable slack at 0.70. It also closes a non-termination pathology that grows with difficulty and that a decoding-time confidence hook makes worse. We do not claim to beat a well-tuned length penalty or decoding-time early exit on the raw trade-off; the contribution is how the halt is obtained.

1 Introduction

Reasoning models often continue thinking after their answer is effectively known, and global or outcome-level shortening objectives do not provide a per-problem halt. The paper internalizes a causal halt direction into the model weights, while preserving the distinction between stopping and correctness.

Motivation: A matched comparison finds answer confidence settling at about 1,650 think-tokens versus about 3,700 generated tokens, a factor of 2.24.Thirteen percent of rollouts never saturate before the 16k cap.
Why existing objectives fail: Shortness rewards, short-trace imitation, and per-problem budget annealing fail to install an adaptive halt at each problem's answerable point.A global length penalty can match or beat the tested per-problem budgets because one scalar reward cannot localize where to stop.
Method: The method separates how to stop from when to stop by using a causal halt vector and a replicated value axis as a control.The value axis is near-orthogonal to the halt vector and does not govern math halting.
Contribution: A layer-18 difference-of-means halt vector is installed with full-vector reconstruction and off-axis dimensions pinned, rather than scalar projection alone.The resulting hook-free self-halt is fit from 24 problems without reinforcement learning and removes about 24% of thinking at held accuracy across five unseen benchmarks.
Scope: The paper does not claim to beat well-tuned length-penalty reinforcement learning or decoding-time early exit on the raw compression-accuracy trade-off.Its stated contribution is the mechanism and a deployment profile without serving-time overhead or a reinforcement-learning pipeline.

2 Related work

Prior approaches shorten reasoning through outcome-level rewards, budget limits, decoding-time exits, or compression, whereas this work learns a per-problem halt from a causal internal signal. Related steering and merging work provides the methodological context, but the paper distinguishes its weight-internalized halt from these alternatives.

Efficient reasoning: Efficient-reasoning methods include length-penalty reinforcement learning, test-time budgets, dynamic early exit, and supervised or prompt-based compression.The paper positions its approach as orthogonal because it is per-problem and learned from a causal internal signal.
Reading reasoning state: A probe-based verifier has reported roughly 24% token reduction at no accuracy cost, while this paper installs its halt in the weights so inference needs no auxiliary model.Its target is a ground-truth forced-answer measurement rather than a learned probe.
Activation steering: The work extends difference-of-means activation steering and prior attempts to move steering directions into model weights.The related methods include refusal ablation by weight orthogonalization and fine-tuning a steering vector into the residual stream.
Value axis: The replicated value axis modulates confidence, backtracking, and verbosity in prior work but is near-orthogonal to this model's halt vector and does not govern math halting.Here it serves as a control rather than the stopping mechanism.
Model merging: The paper's safe-scaling result draws on weight-space merging, including averaged LoRA adapters, model soups, TIES, and task arithmetic.These methods frame fine-tuned-minus-base changes as composable task vectors.

3 Background: measuring when the model knows

The paper measures when the model knows by forcing an answer from intermediate reasoning prefixes and identifying the earliest prefix that already yields a correct answer. This offline boundary defines removable slack for training, rather than serving as a deployable stopping signal.

Setup: The study uses DeepSeek-R1-Distill-Qwen-7B and evaluates zero-shot on MATH500, AMC, and AIME 2024/2025.The model has 7B parameters and 28 layers; training problems and offline pools come from DeepScaleR.
Forced-answer measurement: For each reasoning prefix, the method appends a final-answer scaffold, greedily generates the answer, and grades it with math_verify.This replaces an unsafe probe-based confidence cutoff with a ground-truth forced-answer measurement.
Forced-answer point: Bend is the earliest prefix whose forced answer is already correct, so reasoning after Bend is in-principle cuttable slack.Problems with substantial slack are compressible goldilocks cases, while anchors have none.
Scope of the target: Bend is measured offline on the base model and requires the gold answer, making it a training target rather than a deployable signal.The work therefore trains the model to choose to stop where it can be forced to answer correctly.
Direction construction: The paper constructs layerwise difference-of-means directions from positional contrasts and reads them through cosine projection.It builds both a replicated value axis and a halt vector from distinct contrast sets.

4 The halt vector is a causal stopping knob

Steering the halt vector provides a causal control over reasoning length, whereas a magnitude-matched value-axis control does not. Layer 18 offers the best balance between strong length control and retained accuracy for internalization.

Intervention: Adding a scaled halt direction to the residual stream lets the paper sweep steering strength as a percentage of the mean activation norm.The intervention is tested during generation across layers and steering strengths.
Causal knob: Steering strength correlates with length at ρ = −0.60 at layer 6, with positive steering prompting rapid wrap-up and negative steering preventing termination.An independent layer-6 measure gives ρ = −0.63.
Control: The magnitude-matched value-axis control has ρ = +0.09, showing no comparable length effect.This supports direction specificity rather than a generic consequence of residual-stream perturbation.
Layer selection: Layer 18 is selected because layer 24 preserves accuracy better but provides a weaker length lever, while layer 6 suffers an early-layer accuracy collapse.All subsequent internalization therefore uses layer 18.
Layer selection: A layer sweep finds that retained accuracy at +10 rises from 14% at layer 6 to 56% at layer 18 and 72% at layer 24.Layer 18 also gives the strongest clean length control, with ρ = −0.82 and 31% shorter generation at +10.

5 Internalizing the knob

Internalizing the halt requires reproducing the full steered layer-18 activation while preserving off-axis dimensions for the frozen reader. Scalar projection targets fail because unconstrained off-axis drift damages behavior, whereas reconstruction-based writer adaptation preserves the intended intervention.

Writer-reader split: LoRA is placed only on the writer through layer 18 so the reader remains frozen while the writer learns to reproduce the steered activation.The writer comprises layers 0 to 18, and the reader comprises layers 19 to 27.
Failure of scalar projection: Scalar-projection training lengthens and degrades generation because it leaves off-axis activation dimensions unconstrained.The strongest arm generated 43% longer, reached 0.57 accuracy, and stopped emitting </think> altogether.
Whole-vector reconstruction: Whole-vector reconstruction regresses layer-18 activations toward the steered target instead of optimizing only its projection onto the halt direction.The target uses the base activation for off-axis dimensions while applying the scheduled steer along the direction.
Preserving the reader: Pinning off-axis dimensions to natural values gives the frozen reader the vector it responds to, reducing disturbance relative to scalar arms.Reconstruction arms reached at most 3.66 units of off-axis drift per achieved on-axis movement, versus a scalar-arm minimum of 5.22, with no family overlap.
Ablations: A sustained steer after Bend is necessary, while ramp shape is irrelevant and attention-only adaptation outperforms adding MLP capacity.Holding the steer after Bend installs the halt; releasing it does no better than base, and extra capacity increases opportunities for off-axis corruption.

6 Results

Reconstruction installs a direction-specific, per-problem halt that generalizes across five held-out benchmarks, reduces thinking at held accuracy, and addresses non-termination. The resulting efficiency is competitive rather than state of the art, while adaptivity tracks each problem’s removable slack.

6 Results: 25% fewer think-tokens at 0.792 accuracy demonstrates that reconstruction compresses reasoning while staying within held accuracy on the 24-problem fit.Imitation runs 15% longer, while gated steering-distillation cuts 21% but loses 6.2 accuracy points.
6 Results: 11%, 25%, 27%, and 31% think reductions occur at strengths 10, 25, 40, and 50, while point-estimate accuracy declines from 0.833 to 0.740.Strength 25 is selected as the headline operating point because compression largely saturates while accuracy remains within about 2 points of base.
6 Results: At strength 10, anchor-problem length changes by only +2% with exact accuracy, while compressible problems are cut 11%, showing that the halt responds to early answerability.Higher strengths also trim anchors, but anchor accuracy never declines across the reported sweep.
6 Results: Only the halt vector installs a halt when reconstruction magnitude is fixed; the value axis barely changes generation and a random direction runs longer.These controls distinguish direction-specific behavior from effects caused by perturbing layer 18 at a matched magnitude.
6 Results: 20 to 24% less thinking at held accuracy transfers zero-shot across five held-out benchmarks, with α = 10 giving 11 to 16% reductions.The cut also holds on easy MATH500, where reasoning is already about 3,100 think-tokens.
6 Results: 0.70 Pearson correlation links each problem’s think-cut to its own compressible slack at strength 25, whereas a constant-cut control reaches 0.11 with an interval spanning zero.The strength-10 correlation is 0.59; both halt intervals exclude zero.
6 Results: 31 to 38% non-termination on the two AIME sets falls to 6 to 9% with reconstruction, addressing a difficulty-growing termination failure distinct from mean token reduction.The base often reaches the 16k-token cap, while reconstruction closes this gap on every benchmark; the non-terminating tail dominates worst-case latency and cost.
6 Results: 20 to 24% thinking reduction with at most 2 accuracy points lost places reconstruction competitively against DEER and published length-penalty reinforcement learning, not ahead of them.On AIME24, reconstruction loses 2.0 points at −24% compression versus DEER’s 6.7 points at −26%.

7 Scaling and robustness

A single training run loses the halt as exposure increases, but averaging independently trained 24-problem LoRA bursts preserves the reduction across larger datasets.

7 Scaling and robustness: 24% thinking reduction at held accuracy persisted from 24 to 144 problems when independently trained 24-problem LoRA bursts were averaged.Termination remained at or above 0.97, supporting robustness rather than a scaling law.
7 Scaling and robustness: 150 problems trained in one run reduced thinking by only 0.2%, versus 24% for the 24-problem setup.The full-trace language-model loss out-competed the reconstruction target with exposure.
7 Scaling and robustness: Trace truncation bought shorter generations only by reducing accuracy, so it did not provide a clean remedy for single-run erosion.

8 Comparison to prior methods

The paper compares its learned halt with decoding-time DEER and trained length penalties, but positions its contribution as an internalized, deployable mechanism rather than raw frontier dominance.

DEER is a strong decoding-time compressor, but the paper does not claim to beat it on raw compression.DEER forces </think> when hesitation-token confidence exceeds a threshold.
Well-tuned length penalties are at least as efficient on the raw trade-off among trained methods.
The paper’s claim concerns mechanism and deployment: the learned halt avoids runtime machinery while addressing DEER’s hard-set accuracy-dip failure mode.

9 Discussion and limitations

The discussion frames the halt as an interpretability finding converted into installed behavior, while bounding the evidence to one model and identifying concrete extensions and training-pipeline dependencies.

Discussion: The halt converts a read-out direction shown to be causal into installed weight behavior that fires without runtime machinery.
Limitations: The study is confined to DeepSeek-R1-Distill-Qwen-7B, making extension to other model families and sizes the most important next step.
Limitations: The 24-problem fit leaves training-set accuracy noisy, so evaluation relies on think-tokens and larger held-out sets.
Limitations: Although deployment is hook-free, the training pipeline still requires a base-model probe because the forced-answer point is measured offline using the gold answer.
Future work: Future work includes a reader-side mirror, re-deriving Bend on the trained model with correctness gating, evaluation beyond mathematics, and per-problem steering strength.

10 Conclusion

The paper concludes that its main contribution is internalized halting, while its outcome-reward experiments show why simpler global penalties outperform elaborate saturation curricula. The negative results motivate a mechanism-based approach rather than reinforcement-learning budget shaping.

Conclusion: Reasoning models overthink by a stable margin, and the layer-18 halt vector internalizes causal stopping into the weights without reinforcement learning.The resulting halt is hook-free and tunable per problem.
Conclusion: The appendix evaluates outcome-reward signals offline and trains halt behavior through supervised and preference objectives, with GRPO used for reinforcement-learning arms.
Conclusion: A single global length budget is ineffective across mixed-difficulty problems, where correct-answer lengths rise from about 2,342 tokens on easy problems to about 7,365 on hard ones.
Conclusion: A per-problem 30th-percentile budget reduces dead problems to roughly 1% and makes about 48% useful, with advantage-length correlation about −0.60 among correct rollouts.
Conclusion: Strict forced-answer correctness is necessary because first-token matching fires before reasoning; on AMC, saturation-point correctness is 86% versus 16% with zero reasoning.The passage attributes about 70 points of the difference to reasoning and estimates AMC saturation as about 94% earned rather than a format prior.
Conclusion: Outcome-RL preserves accuracy between 0.79 and 0.83 but achieves only 0–6% compression, while the naive global penalty beats every per-problem saturation budget.
Conclusion: Across outcome-RL arms, the internal saturation point stays near 1,650 tokens, while the length penalty’s savings are confined to rollouts that continue after saturation and become wrong.
Conclusion: Removing mechanical explanations with a constant-slope reward, reachable 20th-percentile target, and correctness floor leaves the negative result unchanged.

A.5 Imitation and preference optimization fail too

Imitation, preference optimization, and outcome-level objectives fail to install a per-problem stopping rule. Objectives that directly weight stopping instead impose a global cut, trading accuracy for compression.

Supervised imitation: Directly supervised stop examples use 1,300 rows, including 900 halt examples and 400 anchors, by cutting correct traces at their forced-answer points plus a margin.Anchor problems remain full length so the model learns when not to stop.
Supervised imitation: 0.34% of the token loss comes from stop-and-commit tokens, so plain fine-tuning leaves length unchanged and can even increase it.On training problems, the fine-tuned adapter reaches 7,174 tokens versus 6,505.
Conclusion: Outcome objectives either ignore the stopping decision or install a single global cut, because scalar rewards cannot localize where each problem should stop.This provides the evaluation gate for requiring an adaptive halt rather than merely shorter outputs.
Evaluation: The evaluation uses math_verify for accuracy and reports a grader-independent length signal alongside comparator agreement.Training and evaluation use H200 GPUs, a 16k-token cap, and no reinforcement-learning infrastructure for the headline result.

C How the halt vector is constructed

The halt vector is a within-rollout difference-of-means contrast between reasoning positions after and before the model locks onto the correct answer. Its layer-18 readout is less legible than the value-axis control, but steering—not AUROC—distinguishes the causal halt mechanism.

Contrast construction: The lock-on point is the first chunk where the forced-answer probability reaches 0.9 and stays there, and the contrast compares positions after versus before that point.Only correctly answered rollouts are used, so the landmark is defined for traces that reach the right answer.
Contrast construction: The construction uses 457 usable contrasts from 480 correct rollouts over 83 problems, with 11 lacking lock-on and 12 already locked at the first chunk.The split is by rollout, not by problem.
Interpretation and limits: The AUROC measures token-position separation between late and early reasoning, not rollout-level discrimination of correct versus incorrect answers.The reported held-out sample counts are 457 rollouts for the halt vector and 94 conversations for the value axis.
Readout and layer choice: The value-axis control reads 0.780 at layer 18 versus 0.758 for the halt vector, showing that the halt claim concerns steering rather than superior readout legibility.The value axis is held out by conversation, whereas the halt vector is held out by rollout.
Contrast validation: The rejected between-item contrast reaches 0.955 at layer 0 and remains near 0.95 across layers, unlike the depth-dependent lock-on contrast used for the halt vector.Its early separation is treated as a surface confound rather than a computed feature.

D.1 The reconstruction objective: schedule, strength, and capacity

The reconstruction objective succeeds by preserving the activation’s off-axis dimensions while steering its halt direction. Schedule and capacity matter: hold and gate install the halt, attention-only is safer than extra MLP capacity, and scalar projection fails through off-axis corruption.

Selection grid: The reconstruction grid selects schedule, steering strength, and adapter capacity on 24 compressible problems against a base of 0.812 accuracy and 6,516 think-tokens.The grid compares hold, gated, and tent schedules plus attention-only and attention-plus-MLP adapters.
Schedule and capacity: Hold and gate schedules install the halt, while attention-only beats attention-plus-MLP because extra capacity distorts more off-axis dimensions.Tent barely moves length.
Off-axis preservation: Reconstruction and scalar-projection arms separate fourfold across off-axis drift measures, including reader KL, and drift correlates negatively with training accuracy at ρ = −0.799.The reported permutation p-value is 0.0079, or 0.032 with fourfold Bonferroni correction.
Off-axis penalty: The off-axis penalty reduces raw drift from 0.769 to 0.041 but also reduces on-axis movement from +0.1474 to +0.0008, suppressing the edit rather than selectively cleaning it.No arm enters the reconstruction band or proceeds to held-out evaluation.
Interpretation: The crossed scalar-with-off-axis-constraint design remains uninformative because the penalty trades away on-axis achievement, leaving no evidence that it can preserve the edit while cleaning drift.The authors explicitly avoid treating the result as evidence for either side.
Truncation controls: Hard and soft truncation trace the same accuracy–compression trade-off: λ = 0 barely cuts, while λ = 1 removes 23.1% at an 11-point accuracy cost on the holdout.No soft holdout cell is both held-accuracy and materially compressed, so the headline remains the 24-problem adapter.