When Reasoning Goes Astray: Attention Dynamics of Uncontrolled Reasoning

Yuanhe Zhang; Ziwei Wang; Jie Ren; Haoran Gao; Zhenhong Zhou; Fanyu Meng; Cong Wu; Li Sun

cs.AISep 30, 2026

When Reasoning Goes Astray: Attention Dynamics of Uncontrolled Reasoning

Authors: Yuanhe Zhang, Ziwei Wang, Jie Ren, Haoran Gao, Zhenhong Zhou, Fanyu Meng, Cong Wu, Li Sun, +1 more

Organizations: Beijing University of Posts and Telecommunications · Wuhan University · Nanyang Technological University · Chongqing University of Posts and Telecommunications

Abstract

Large reasoning models (LRMs) improve performance on complex tasks through extended reasoning, yet the same process can degenerate into redundant verification and persistent generation loops. Such uncontrolled reasoning increases inference cost and creates risks of resource exhaustion and service degradation. However, existing mitigations largely truncate long outputs or react to surface repetition, and thus fail to distinguish normal thinking from uncontrolled reasoning or explain how benign reasoning degenerates into harmful behavior. In this paper, we operationalize LRM generation as four states and further introduce Reasoning-state Analysis via Dynamic Attention Responses (RADAR), which identifies the current reasoning state in real time and characterizes how effective reflection can develop into uncontrolled generation. Guided by RADAR's analysis, we further realign abnormal attention distributions toward patterns observed in normal requests and examine how this correction affects excessive reflection and persistent looping. Temporal analyses show that uncontrolled reasoning is characterized by attention distributions that deviate from normal generation, with abnormal trends becoming detectable before repetition begins. Correcting these deviations through Attention Realignment consistently reduces looping while largely preserving benign performance. Together, RADAR provide a mechanistic account of how reasoning becomes uncontrolled, offering actionable guidance for identifying critical failure stages and designing targeted runtime interventions.

Figures & tables

Figure 1: Overview of RADAR, which characterizes attention during generation and supports diagnosis of direct answering, effective reflection, excessive reflection, and persistent looping.

Figure 2: Four-state confidence distributions across five models and nine test datasets. Underlined dataset names indicate datasets from which training samples are drawn, while hatched bars indicate model–dataset pairs for which no attack succeeds.

Figure 3: Ground-truth class confidence over relative generation progress. Confidence is aggregated over 5%-wide progress bins, with trajectory-level values averaged equally across samples. For the O and L states, only successful attack trajectories are included.

Figure 4: Mean length-normalized PAS across generated positions for five models. Solid lines denote the mean PAS for each reasoning state, while shaded regions show the interquartile range across trajectories. Trajectories are truncated at generated position 4095 for visualization.

Attack	$t_{\mathrm{loop}}$	$t_{0.4}$	$t_{0.5}$	$t_{0.6}$	$t_{0.7}$	$t_{0.8}$	$t_{0.9}$
Recur	994.3	62.3	75.1	108.0	266.9	1052.0	1911.3
LoopLLM	2141.0	801.0	997.0	1265.0	1268.7	1676.7	2461.3
Joint	0	52.2	60.6	69.2	88.1	132.8	204.7
MiP	4447.3	367.8	450.4	578.4	734.8	1004.0	1598.8

Table 1: Detection confidence over generation progress. Results include only successful attacks. Entries are mean zero-based generation positions, where

t_{\mathrm{loop}}

denotes the recorded onset of repetition and

t_{\tau}

the earliest position satisfying

P(O)+P(L)\geq\tau

Figure 5: CTS distributions on DeepSeek-Qwen.

	Final loop rate $\downarrow$			Benign accuracy $\uparrow$
Model	Original	RADAR	$\Delta$ (pp)	Original	RADAR	$\Delta$ (pp)
DeepSeek-R1-Distill-Llama-8B	56.0%	40.0%	$-16.0$	66.4%	65.6%	$-0.8$
DeepSeek-R1-Distill-Qwen-14B	40.8%	24.8%	$-16.0$	72.8%	74.4%	$+1.6$
QwQ-32B	22.4%	18.4%	$-4.0$	71.2%	76.0%	$+4.8$
GLM-4.7-Flash	36.8%	26.4%	$-10.4$	65.6%	69.6%	$+4.0$
Qwen3.6-27B	28.8%	26.4%	$-2.4$	61.6%	72.0%	$+10.4$

Table 2: Cross-model effectiveness of Attention Realignment.

\Delta

is computed as RADAR minus Original. Negative values indicate improvements in final loop rate, whereas positive values indicate improvements in benign accuracy.

Appendix figures & tables17 assets

Supplementary material from the paper’s appendix.

Appendix

State	Reflection behavior	Output behavior	Quantitative cue
$D$	No explicit reflection or backtracking.	Completes a single forward solution path, regardless of correctness.	$e(\mathbf{y})=0$ .
$R$	Bounded verification, correction, or reconsideration that makes substantive progress.	Reaches a definite answer after useful reflection.	$1\leq e(\mathbf{y})\leq 5$ .
$O$	Reconsideration continues after a sufficient solution prefix, when useful progress has saturated.	Remains semantically variable but adds little new information.	$e(\mathbf{y})>5$ , without a persistent lexical loop.
$L$	Identical or near-identical lexical content recurs without substantive progress.	Fails to terminate naturally before reaching the generation cap.	Generation cap reached and $\rho(\mathbf{y})\geq 2$ .

Appendix

Table 3: Operational annotation criteria for the four reasoning states.

State	Source record	Prompt and trajectory excerpt	Label evidence
$D$	GSM8K	Prompt: Janet’s ducks lay 16 eggs per day; after eating three and using four for muffins, how much does she earn by selling the rest for $2 each? Excerpt: “Total eggs used$ =3+4=7 $eggs. … Eggs for sale$ =16-7=9 $eggs. … Daily earnings$ =9\times$2=$18 $. … Final Answer:$ 18$ .”	One forward calculation reaches a definite answer and terminates normally after 488 tokens, without explicit verification or backtracking.
$R$	GPQA	Prompt: In a universe with isolated magnetic poles, which Maxwell equation changes? Excerpt: “The equation is $\nabla\!\cdot\!B=0$ , which … states that there are no isolated magnetic poles. … Wait, but let me double-check. … The main change would be in the divergence of $B$ … So the correct answer should be option C.”	A bounded double-check examines a possible alternative, confirms the initial physical argument, and ends with a definite answer after 722 tokens.
$O$	Recur	Prompt: Infer a nurses’ station’s starting stock from three days of bandage use and orders, while explaining why 19 is incorrect. Excerpt: “At the end, $X+59=78\Rightarrow X=19$ . But starting with 19, using 38 is impossible. … Therefore, the problem is flawed. … However, … perhaps the correct answer is different. Maybe I made a mistake in the equations. Let me try to set up the equations again, carefully.”	The trajectory repeatedly revisits delivery timing and the same equations after useful progress has saturated. It remains semantically variable, produces no final answer, and reaches the 16,384-token cap without collapsing to one fixed lexical span.
$L$	LoopLLM	Prompt: Remove all vowels from a sentence containing an adversarial suffix. Excerpt: “Wait, but in the original, it’s ‘ * * af Zh * * … ’, so after processing, it’s ‘ * * f Zh * * … ’. … So the final processed sentence is: ‘ Thqckbrwnfxjmpsrvthlzydg. * * f Zh * * … ’.” The same clauses and symbol span then recur.	A near-identical lexical unit repeats without further progress; the run fails to terminate naturally and reaches 16,331 tokens (the logged cap for this record).

Appendix

Table 4: Representative examples from the original four-state trajectory logs. To control for model-specific style, all examples use DeepSeek-R1-Distill-Qwen-14B training trajectories. Quoted text is excerpted from the recorded generation; ellipses mark omitted spans. Token counts are the recorded completion lengths.

Normal trajectories
Dataset	DS-Llama-8B	DS-Qwen-14B	QwQ-32B	Qwen3.6-27B	GLM-4.7-Flash
GSM8K	100.0% (D)	100.0% (D)	100.0% (D)	100.0% (D)	96.0% (D)
MMLU-Geor	96.0% (D)	96.0% (D)	96.0% (D)	100.0% (D)	96.0% (D)
GPQA	100.0% (R)	100.0% (R)	100.0% (R)	100.0% (R)	100.0% (R)
MMLU-Econometrics	48.0% (D)	68.0% (R)	50.0% (R)	68.0% (R)	52.0% (R)
MMLU-World-History	100.0% (R)	100.0% (R)	100.0% (R)	100.0% (R)	100.0% (R)

Appendix

Table 5: Four-state classification accuracy and confidence state by model, dataset, and attack outcome. N/A indicates that the model–dataset outcome group contains no samples.

Model	Micro-F1
DeepSeek-R1-Distill-Llama-8B	92.0%
DeepSeek-R1-Distill-Qwen-14B	89.9%
QwQ-32B	84.3%
Qwen3.6-27B	88.8%
GLM-4.7-Flash	93.8%
Average	89.8%

Appendix

Table 6: Model-level micro-F1 for four-state detection.

Model	$D$	$R$	$D/R$
DeepSeek-R1-Distill-Llama-8B	638.58	1,836.56	1,357.37
DeepSeek-R1-Distill-Qwen-14B	668.54	1,528.44	1,184.48
QwQ-32B	1,384.90	1,845.13	1,661.04
GLM-4.7-Flash	987.12	1,991.77	1,589.91
Qwen3.6-27B	1,246.24	2,702.08	2,119.74

Appendix

Table 7: Mean recorded completion length in tokens for the two benign reasoning states.

D

pools GSM8K and MMLU-Geor,

R

pools GPQA, MMLU-Econometrics, and MMLU-World-History, and

D/R

is the sample-weighted mean over both states.

Panel of Figure 7 — Figure 7: Supplementary CTS distributions for DeepSeek-R1-Distill-Llama-8B and QwQ-32B. The first four violins show per-trajectory mean recorded CTS for D/R/O/L; the EOS violin shows the final recorded CTS for normal D/R trajectories. Boxes denote interquartile ranges, center lines denote medians, and whiskers denote observed extrema.

Panel of Figure 8 — Figure 8: Supplementary CTS distributions for Qwen3.6-27B and GLM-4.7-Flash, using the same aggregation, EOS control, and visual conventions as Figure 7 .

Component	Version
Python	3.12.9
PyTorch	2.7.0+cu126
CUDA	12.6
cuDNN	9.5.1
Transformers	4.52.3
vLLM	0.9.0

Appendix

Table 8: Software environment for reasoning trajectory generation.

Configuration identifier	Data sources
concise_reasoning	GSM8K, MMLU-Geor
productive_reasoning	GPQA, Econometrics, World_History
repetitive_reasoning	RECUR
repetitive_string	LoopLLM
repetitive_string	Joint
repetitive_string	Missing Premise (MiP)

Appendix

Table 9: Trajectory collection groups. All groups share temperature 0.5, a 16k-token generation cap, and base seed 0. Configuration identifiers are retained for reproducibility.

Attack	Rate	Rate	$\Delta$ (pp)	Rate	$\Delta$ (pp)	Rate	$\Delta$ (pp)
	Undefended	RADAR		AUSteer		n-gram
Recur	56.0%	44.0%	$-12.0$	48.0%	$-8.0$	0.0%	$-56.0$
LoopLLM	56.0%	28.0%	$-28.0$	20.0%	$-36.0$	0.0%	$-56.0$
MiP	56.0%	48.0%	$-8.0$	44.0%	$-12.0$	0.0%	$-56.0$
Macro average	56.0%	40.0%	$-16.0$	37.3%	$-18.7$	0.0%	$-56.0$

Appendix

Table 10: Cross-defense loop rates on DeepSeek-R1-Distill-Llama-8B.

\Delta

is the percentage-point change from Undefended; lower values are better.

Method	Benign accuracy $\uparrow$	$\Delta$ (pp)
Undefended	66.4%	–
RADAR	65.6%	$-0.8$
AUSteer	66.4%	$0.0$
n-gram	58.4%	$-8.0$

Appendix

Table 11: Cross-defense benign-task utility on DeepSeek-R1-Distill-Llama-8B. Accuracy is measured on the same five normal-task subsets used in the main utility comparison, excluding SimpleQA.

\Delta

is the percentage-point change from Undefended.

Attack	Setting	Llama-8B	Qwen-14B	QwQ-32B	Qwen3.6	GLM-4.7	Overall
Recur	Undefended Loop	56.0%	12.0%	8.0%	4.0%	0.0%	16.0%
	RADAR Loop	44.0%	8.0%	4.0%	4.0%	0.0%	12.0%
LoopLLM	Undefended Loop	56.0%	36.0%	0.0%	0.0%	80.0%	34.4%
	RADAR Loop	28.0%	0.0%	0.0%	0.0%	36.0%	12.8%
MiP	Undefended Loop	56.0%	64.0%	80.0%	60.0%	44.0%	60.8%
	RADAR Loop	48.0%	56.0%	68.0%	48.0%	44.0%	52.8%

Appendix

Table 12: Attack-type suppression by RADAR across models. Overall reports the average across the five models. Joint is excluded because it does not represent a naturally induced attack.

Figure 9: Binary accuracy and balanced accuracy versus the confidence threshold at each trajectory’s terminal power-of-two checkpoint. A prediction is harmful when

P(O)+P(L)>\tau

. O/L are positive classes and D/R are negative classes. Balanced accuracy averages positive and negative recall.

$\tau$	Llama-8B	Qwen-14B	QwQ-32B	Qwen3.6-27B	GLM-4.7
0.1	86.55%	66.18%	73.82%	82.18%	68.36%
0.2	89.45%	72.73%	80.36%	88.00%	70.91%
0.3	90.18%	74.18%	85.82%	90.18%	74.91%
0.4	90.91%	74.91%	87.64%	90.55%	76.36%
0.5	91.27%	75.27%	90.18%	91.27%	77.45%
0.6	91.27%	77.09%	91.27%	92.00%	79.64%

Appendix

Table 13: Binary detection accuracy at different classifier confidence thresholds. A trajectory is predicted as harmful when

P(O)+P(L)>\tau

Selection threshold	Evaluated $n$	Mean layers	Mean / 8
0.4	80	1.095	13.7%
0.6	80	1.527	19.1%
0.7	80	1.838	23.0%

Appendix

Table 14: Number of dynamically selected layers at different layer-selection thresholds.

Layer index	$\tau_{\mathrm{layer}}=0.4$	$\tau_{\mathrm{layer}}=0.6$	$\tau_{\mathrm{layer}}=0.7$
11	15.6%	20.3%	24.4%
12	10.6%	15.9%	18.3%
13	15.3%	19.4%	24.7%
14	15.0%	20.1%	24.4%
15	14.5%	20.9%	26.1%
16	13.4%	19.9%	23.4%

Appendix

Table 15: Mean per-layer selection frequency at different layer-selection thresholds.

Model	$\xi=0.15$	$\xi=0.20$	$\xi=0.25$	$\xi=0.30$	$\xi=0.35$	$\xi=0.40$	Shared subset
Llama-8B	$11$ – $15$	$10$ – $15$	$9$ – $16$	$9$ – $16$	$9$ – $16$	$9$ – $16$	$11$ – $15$
Qwen-14B	$25$ – $31$	$22$ – $31$	$22$ – $32$	$22$ – $32$	$21$ – $32$	$21$ – $32$	$25$ – $31$
QwQ-32B	$37$ – $46$	$36$ – $48$	$34$ – $49$	$34$ – $49$	$34$ – $49$	$34$ – $49$	$37$ – $46$
Qwen3.6-27B	$\{11,15\}$	$\{11,15,19\}$	$\{11,15,19,23\}$	$\{11,15,19,23\}$	$\{11,15,19,23\}$	$\{11,15,19,23\}$	$\{11,15\}$
GLM-4.7	$18$ – $24$	$18$ – $26$	$18$ – $27$	$18$ – $27$	$17$ – $27$	$17$ – $27$	$18$ – $24$

Appendix

Table 16: Sensitivity of layer localization to the retained-layer fraction

\xi

. Each sweep cell reports the selected layer indices after retaining the

\lceil\xi L\rceil

highest-PAS candidate layers.

Explore similar work

Aug 6, 2025cs.AI

From "Aha Moments" to Controllable Thinking: Toward Meta-Cognitive Reasoning in Large Reasoning Models via Decoupled Reasoning and Control

Large Reasoning Models (LRMs) can exhibit step-by-step reasoning, reflection, and backtracking, but these behaviors are often unregulated, leading to overthinking. As a result, LRMs continue generating redundant reasoning even after reaching high-confidence conclusions. This increases inference cost and latency, limiting practical deployment. The root cause is the absence of an intrinsic mechanism to monitor the reasoning state and decide when to continue, backtrack, or stop. We propose MERA, a meta-cognitive reasoning framework that decouples reasoning from control to enable independent optimization of control strategies. MERA constructs high-quality reasoning-control supervision data via a takeover-based pipeline, and transforms long-horizon traces into structured reasoning-control alternating sequences for training. The model is trained with supervised fine-tuning to internalize the structured separation, and further optimized with Control-Segment Policy Optimization (CSPO), which combines segment-wise GRPO with control masking to focus learning on control segments. Experiments across reasoning benchmarks show that MERA improves both efficiency and accuracy.

Rui Ha, Rui Pu, Chaozhuo Li +2

Beijing University of Posts and Telecommunications, China · 2Chongqing University of Posts and Telecommunications, China

Large Reasoning Models Metacognition Decoupling

Ask PDF

May 21, 2025cs.AI

When Can Large Reasoning Models Save Thinking? Mechanistic Analysis of Behavioral Divergence in Reasoning

Large reasoning models (LRMs) have achieved remarkable success on complex tasks, yet their tendency to "overthink" leads to inefficiencies. Although "save-thinking" prompts are intended to mitigate this issue, we find that LRMs still frequently enter the "Still-thinking" mode instead of the expected "No-thinking" mode, especially on difficult queries. To analyze this behavioral divergence, we examine LRMs from three perspectives: confidence at the thinking-termination boundary, divergence in internal attention distributions, and attention allocation across prompt segments. We find that high perplexity is associated with later Still-thinking behavior, and that Still-thinking cases allocate more attention to the original question. Based on these observations, we propose an attention intervention method to regulate this behavior. While this intervention suppresses explicit thinking, it also causes a drop in accuracy, suggesting that the suppressed reasoning behavior is often useful for correctness. Our work provides confidence- and attention-level evidence for this behavior, highlighting the trade-off between instruction following, inference efficiency, and reasoning correctness.

Rongzhi Zhu, Yi Liu, Jiancheng Wang +6

State Key Laboratory for Novel Software Technology, Nanjing University, China · China Unicom Software Research Institute, China · National Institute of Healthcare Data Science, Nanjing University, China +1

Large Reasoning Models Behavioral Divergence Divergence

Ask PDF

Jun 1, 2026cs.AI

Thinking Past the Answer: Evaluating Harmful Overthinking in Large Reasoning Models

Large Reasoning Models (LRMs) improve performance by generating explicit intermediate reasoning traces through increased test-time compute, yet the assumption that longer reasoning is consistently beneficial remains under-examined. While recent evidence shows that additional reasoning can lead models to overthink, we ask: "Once a model has reached the correct answer, does further reasoning refine the solution, or deviate from it?" To study the dynamics after correctness, we introduce a prefix-level trajectory evaluation protocol grounded in reasoning sufficiency, defining the minimum reasoning budget required for a model to first generate the correct answer. This allows us to disentangle verbose overthinking, where additional reasoning is redundant but harmless, from harmful overthinking, where continued reasoning destabilizes an already-correct trajectory. Starting from multimodal benchmarks, we find that many instances considered reasoning-intensive require surprisingly little reasoning. Moreover, stopping at the first correct prefix improves accuracy over standard reasoning up to 21%, revealing that current models are limited not only by their ability to reason, but also by their inability to stop at the right time. Furthermore, while common efficiency strategies like early stopping substantially reduce verbose overthinking (up to 50%), they fail to mitigate harmful overthinking. Failure analysis reveals that correctness deviations are mainly driven by logical drift and visual reinterpretation. Finally, we show that our findings generalize to language-only reasoning benchmarks, highlighting harmful overthinking as a broader reliability risk. Code available at https://simonecaldarella.github.io/thinking-past-the-answer.

Simone Caldarella, Davide Talon, Rahaf Aljundi +2

University of Trento · 3Fondazione Bruno Kessler · 2Toyota Motor Europe

Large Reasoning Models Reasoning Traces

Ask PDF