Steering vectors offer a training-free mechanism for controlling reasoning behaviors in large language models, but constructing effective vectors requires identifying genuine behavioral signals in the model's hidden states. For behaviors that can be toggled via prompts, this is straightforward. However, many reasoning behaviors -- such as self-reflection -- emerge spontaneously and resist prompt-level control. Current methods detect these behaviors through keyword matching in chain-of-thought traces, implicitly assuming that every detected boundary encodes a genuine behavioral signal. We show that this assumption is overwhelmingly wrong: across 541 keyword-detected boundaries, 93.3% are behaviorally unstable, failing to reproduce the detected behavior under re-generation from the same prefix. We develop a probabilistic model that formalizes intrinsic reasoning behaviors as stochastic events with context-dependent trigger probabilities, and show that unstable boundaries dilute the steering signal. Guided by this analysis, we propose stability filtering, which retains only boundaries where the model consistently reproduces the target behavior. Combined with a content-subspace projection that removes residual question-specific noise, our method achieves 0.784 accuracy on MATH-500 (+5.0 over the strongest baseline). The resulting steering vectors transfer across models in the same architecture family without re-extraction, improving Nemotron-Research-Reasoning-1.5B (+5.0) and DeepScaleR-1.5B-Preview (+6.0). Code is available at https://github.com/zhmzm/stability-steering.
Figures & tables
Figure 1: Extrinsic behaviors are steered via prompt contrast (left). Intrinsic behaviors emerge spontaneously and resist prompting (middle-left). Keyword matching detects them from traces (middle-right), but most detections are unstable under re-generation (right; data from § 4.3 ).
Figure 2: Overview of our method. Left : SEAL mixes all keyword-detected boundaries, including unstable ones that dilute the signal. Middle : Our content-aware extraction computes per-question steering directions and projects out question-specific information via SVD. Right : Stability probing re-generates from each boundary’s prefix and retains only those that reliably reproduce the target behavior.
Model
Method
Acc
Δ
Avg Tok
Avg R
Source model (vectors extracted here)
Qwen-1.5B
Baseline
0.608
—
1547
39.1
Prompt text
0.672
+6.4
1116
24.3
Prompt vector
0.610
+0.2
1604
39.4
SEAL
0.734
+12.6
1103
10.3
Ours
0.784
+17.6
1008
10.7
Table 1: Main results on MATH-500 (500 questions, max_tokens=4096, layer 20). Steering vectors extracted from DeepSeek-R1-Distill-Qwen-1.5B. Δ = accuracy change vs. each model’s own baseline. Avg Tok = average word count; Avg R = average reflection keyword count.
Figure 3: (a) Distribution of stability scores across 541 R+T boundaries. 93.3% are unstable ( s(b)<0.8 ); only 6.7% pass the threshold. (b) Accuracy peaks at τ=0.8 (+7.0 over SEAL), then drops at τ=0.9 as sample size shrinks. Shaded band shows random baseline ±1σ .
Figure 4: Behavior probe confidence by stability bin (balanced R+T vs. E, GroupKFold by question). Stable boundaries ( s≥0.8 ) yield the highest confidence (0.942).
Table 6
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
α
SEAL
Content Removal ( K=4 )
−100
0.426
0.442
−50
0.348
0.342
−20
0.258
0.256
−1
0.192
0.192
Baseline
0.182
0.182
1
0.190
0.190
Appendix
Table 4: Full coefficient sweep on MATH-500 (Qwen-1.5B, layer 20, max_tokens=1024 , 500 questions).
α
vstable ( τ=0.8 )
A+B Combined
−100
0.496
0.504
−50
0.364
0.366
−20
0.254
0.260
Appendix
Table 5: Coefficient sweep for vstable ( τ=0.8 ) and A+B combined on Qwen-1.5B (MATH-500, max_tokens=1024 , 500 questions).
Deployed large reasoning models (LRMs) often behave unexpectedly. Test-time steering controls LRM outputs by intervening on their hidden representations, but it can degrade output quality. We argue that prior steering work implicitly relies on internal features that detect behavior in already generated text. We show that these detection features are poor predictors of future behavioral outcomes, and thus not the natural intervention target. Instead, we train activation probes to predict future behavior likelihoods from intermediate reasoning steps. These probes predict the most likely behavior with 64%-91% accuracy, revealing a separate type of internal prediction features. Building on these prediction features, we introduce a text-level steering method, Future Probe Controlled Generation. FPCG samples multiple candidate sentences and chooses the best one according to a probe predicting the future behavior likelihood. This enables steering with almost no output quality degradation. FPCG also enables steering in several evaluations where activation steering fails. These results show that distinguishing detection and prediction features enables a more nuanced approach to controlling LRM behaviors.
Evgenii Kortukov, Piotr Komorowski, Florian Klein +5
1Fraunhofer HHI · 2Northeastern University · 3KAIST
Extended reasoning has become standard for frontier Large Language Models (LLMs), yet the trajectories these models produce remain largely uncontrollable. Existing methods for shaping how a model reasons are prompt based approaches and operate at the input level, offering no fine-grained control over the reasoning process itself. Related work analyzes and discovers latent transition dynamics in the reasoning traces from Large Language Models. Building on this, we statistically characterize these states, and show that failure trajectories get stuck in self-loops, exhausting the token budget without progress toward the final answer. To intervene on these failures, We propose SOPHIA: Steering Of reasoning Processes via Hidden-state Intervention and Activations. We treat each reasoning trace as a sequence of latent states rather than an unstructured texts, and investigate whether inference time interventions can provide fine-grained control over the self-looping reasoning process. We classify every prefix to a latent state, record step level transitions, and use them to construct a bank of steering vectors indexed by state pairs. At inference time, a controller infers the current state and, given a target state, retrieves the corresponding vector and can also detect self-loops online from the transition structure to prevent the model from sinking into a reasoning black hole. Through extensive experiments, our method reliably intervenes on self-loop failures, with steering vectors that generalize to different state pairs. End task accuracy and token efficiency indicate that fine-grained controllability results in better reasoning quality.
Sheldon Yu, Tong Yu, Xunyi Jiang +6
UC San Diego · Adobe Research · University of New South Wales
Reasoning in Large Language Models (LLMs) poses a challenge for oversight as many misaligned behaviors do not surface until reasoning concludes. To address this, we introduce Behavior Cue Reasoning for making LLM reasoning more controllable and monitorable. Behavior Cues are special token sequences that a model is trained to emit immediately before specific implicit and explicit behaviors, acting as dual purpose signal and control levers. When fine-tuning a weaker external monitor with Reinforcement Learning for reasoning oversight, a compressed view of only information surfaced by Behavior Cues is sufficient signal for the monitor to prune up to 50% of otherwise wasted reasoning tokens in complex math problem solving. When leveraged by an almost optimal rule-based monitor in an environment where excessive constraint violations results in failure, Behavior Cues allows for the recovery of safe actions from 80% of reasoning traces that would otherwise end with the proposal of an unsafe action, more than doubling the success rate from 46% to 96%. Through evaluation across two model families and three domains, we show that Behavior Cue Reasoning improves reasoning monitorability and controllability with no cost to performance. More broadly, our work progresses scalable oversight by demonstrating how the monitored model itself can be trained to reason more tractably to oversight. Code: https://github.com/christopherzc/behavior-cues
Christopher Z. Cui, Taylor W. Killian, Prithviraj Ammanabrolu
University of California, San Diego · 2Brigham Young University