Reasoning models often run into problems without noticing them, even when symptoms are visible in their own text. In humans, metacognition separates monitoring from control: fast, intuitive signals keep watch while slow, deliberate reasoning unfolds and help decide whether to continue, change course, or stop. System One (S1) decision models, such as Jev, are a natural candidate for this role: they return a calibrated probability for a natural-language question about a text in one fast call, yet they are used only as gates before reasoning. We introduce SOMA (System One Metacognitive Awareness), in which an S1 model continuously monitors a System Two (S2) reasoning model and intervenes when needed. SOMA has three layers: a scenario layer states each objective in natural language with predefined actions, an awareness layer lets S1 judge whether the described behavior appears, and a control layer accumulates these judgments, first reminds S2, and takes over only if the behavior persists. Sharing one S1 model and one controller across tasks, without training, SOMA addresses two different failures on two reasoning models: on overthinking, it cuts reasoning tokens by up to 36% while improving average accuracy, and on unanswerable questions, it raises abstention from 70.5% to 87.2%.
Figures & tables
Figure 1: (a) In a common combination, S1 decides first and S2 is called for low-confidence cases; S1 is silent once S2 starts. (b) In SOMA, S1 runs beside S2: S2 periodically pauses at a line break, S1 judges whether a statement about an observable behavior holds for the reasoning so far, and cumulative evidence from positive judgments triggers a reminder inserted at the pause, where S2 decides how to respond. The trace is constructed for illustration, not taken from a run; the evidence accumulation and takeover rules are in Section 3.3 .
Figure 2: Overview of SOMA. Every 256 thinking tokens, S2 continues to the next line break (at most 128 more tokens) and pauses. S1 reads the task, the visible reasoning, and a monitor statement, and returns a probability p , which updates the cumulative score S . Background shading shows the controller state: the first positive S inserts a reminder and resets S (orange); post-reminder evidence reaching the frozen level h triggers a setting-specific takeover (grey), after which S2 writes a bounded answer or refusal reason. The trace is illustrative.
Scenario
Statement sm (judged by S1)
Reminder rm
Takeover action am
Overthinking
The reasoning has reached a final answer and keeps re-verifying it without naming any specific remaining concern.
You may be repeating verification that adds nothing new. Consider whether your recent checks are actually useful, and reply with your judgment.
End thinking; insert Final Answer ; S2 writes the answer (64 tokens).
Unanswerable questions
The reasoning has noticed that the problem is missing necessary information or has contradictory conditions, but is still trying to produce a definite answer, for example by assuming values the problem does not give.
Check once more whether the problem gives everything you need. Then judge whether this check was useful or merely repeated your earlier checks, and say so explicitly.
Insert “The problem is missing necessary information or has contradictory conditions, so it cannot be answered as stated.”; end thinking; pre-fill \boxed{I don’t know.} ; S2 writes the reason (512 tokens).
Table 1: The two monitors evaluated in this paper. Each reminder is wrapped in <reminder> tags and inserted at a pause. Both scenarios share the S1 model, the controller and the system prompt.
GSM8K
MATH-500
AIME
AMC23
GPQA-D
Avg
Method
Acc ↑
Tok
Acc ↑
Tok
Acc ↑
Tok
Acc ↑
Tok
Acc ↑
Tok
Acc ↑
Tok
Qwen3-8B
Vanilla
95.2
2,277
96.2
5,549
66.7
16,719
95.0
8,600
56.6
9,266
91.0
4,232
BF
95.3
− 1%
95.4
− 9%
55.0
− 32%
95.0
− 10%
55.1
− 7%
90.4
− 8%
DEER
94.2
− 62%
92.4
− 52%
55.0
− 36%
77.5
− 56%
49.0
− 53%
88.1
− 54%
Dynasor-CoT
95.3
− 3%
95.8
− 6%
68.3
− 7%
95.0
− 6%
60.1
− 12%
91.4
− 6%
Table 2: Overthinking results. Acc is accuracy (%); Tok is the mean number of S2 tokens, given in absolute terms for Vanilla and as the change relative to Vanilla otherwise. For SOMA, the accounting follows the readout in Appendix B . Avg weights each benchmark by its number of questions and is computed from unrounded per-benchmark values. Bold marks the best accuracy and the largest token reduction for each S2 model; underline marks the second best.
SUM
UMWP
Avg
Method
Abs ↑
Acc ↑
FA ↓
Tok
Abs ↑
Acc ↑
FA ↓
Tok
Abs ↑
Acc ↑
FA ↓
Tok
Qwen3-8B
Vanilla
60
80
3
6,171
81
97
2
1,627
70.5
88.7
2.4
3,837
DEER
68
73
6
− 53%
81
95
3
− 64%
74.8
84.2
4.6
− 56%
Dynasor-CoT
61
77
6
− 12%
83
98
2
− 15%
72.1
87.8
3.8
− 12%
AbsBench prompt
79
65
17
+2%
85
95
3
− 15%
82.2
80.1
10.1
− 2%
Table 3: Unanswerable-question results. Abs is the abstention rate on unanswerable questions, Acc the accuracy on answerable questions and FA the false-abstention rate on answerable questions (all %); Tok is as in Table 2 . For SOMA, the accounting follows the readout in Appendix B . Avg weights SUM and UMWP by their number of questions and is computed from unrounded per-benchmark rates. Bold marks the best value for each S2 model (lowest for FA, largest reduction for Tok); underline marks the second best.
Overthinking
Unanswerable
Variant
Acc ↑
Tok
Abs ↑
Acc ↑
FA ↓
Tok
Qwen3-8B
Vanilla
93.0
6,745
72
90
3
3,494
SOMA
92.5
− 42%
92
83
10
− 29%
Remind-only
93.4
− 16%
74
87
4
− 12%
No reminder
94.1
− 22%
89
84
10
− 26%
Table 4: Ablations on a fixed test subset, reported as question-weighted averages. Remind-only never takes over ( h=∞ ). No reminder accumulates evidence from the first pause and takes over without a reminder, with its own calibrated threshold.
Figure 3: Average accuracy on overthinking (top) and abstention on unanswerable questions (bottom) against mean S2 tokens relative to Vanilla, from the Avg columns of Tables 2 and 3 . The dotted line connects non-dominated methods; higher and further left is better. False abstentions are reported in Table 3 .
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
S2
Variant
h
Takeovers
+ / −
Overthinking
Qwen3
SOMA
0.25
161
+ 3/ − 1
Qwen3
No reminder
6
101
+ 1/ − 0
Llama-R1
SOMA
1.5
70
+ 3/ − 2
Llama-R1
No reminder
3
74
+ 0/ − 0
Unanswerable
Appendix
Table 5: Calibration outcomes at the selected thresholds, pooled over the two calibration sets of each scenario (200 questions per scenario). Takeovers counts the questions taken over; + / − counts questions turned from wrong to correct and from correct to wrong relative to S2 under the same system prompt without intervention.
Overthinking
Unanswerable
Variant
Acc ↑
Tok
Abs ↑
Acc ↑
FA ↓
Tok
Qwen3-8B
Vanilla
93.0
6,745
72
90
3
3,494
X=128
92.9
− 48%
93
85
10
− 36%
X=256
92.5
− 42%
92
83
10
− 29%
X=512
94.1
− 34%
88
86
6
− 27%
Appendix
Table 6: SOMA with different scan intervals X (thinking tokens) on the ablation subset of Table 4 . X=256 is the main setting; the other intervals reuse its thresholds.
Case
Scenario
S2
Decisive step
Outcome (vanilla → SOMA)
Vanilla
SOMA
1
UMWP
R1
takeover
no answer → abstains
32,768
1,245
2
SUM
Qwen
takeover
answers under an assumption → abstains
6,148
1,392
3
SUM
R1
reminder
guesses the missing number → abstains
4,903
1,149
4
MATH-500
R1
takeover
correct → correct
21,665
1,120
5
MATH-500
R1
takeover
correct → wrong intermediate answer
1,679
3,147
6
SUM
Qwen
takeover
correct → false refusal
5,544
2,436
Appendix
Table 7: Cases in this appendix. The last two columns give the tokens generated by S2.
While modern large reasoning models (LRMs) excel at providing correct answers in many tasks, we provide additional evidence that they often struggle with a critical capability: knowing when to abstain from answering. We analyze this gap by comparing LRM behavior to results from an exploratory human study, revealing that human reasoning effort on unanswerable tasks isn't different from answerable tasks, whereas LRMs waste computational resources by generating longer Chains of Thought (CoTs) on unanswerable than on answerable prompts. To overcome this inefficiency, we take inspiration from a resource-rational perspective on human cognition and introduce a novel GRPO reward that encourages efficient reasoning about whether the task contains all the information needed to solve it. Fine-tuning several 4B and an 8B LRM with this reward leads to human-like abstention performance gains (+14.7% on average) while retaining answering capabilities and boosting the models' efficiency (44% shorter CoTs on average).
Polina Tsvilodub, Max Höth, Michael Franke +2
University of Tübingen · Aleph Alpha Research · Lab1141
Large reasoning models improve performance on challenging problems by allocating additional computation before answering, but longer reasoning does not always lead to better results and can introduce substantial redundant reasoning on simple problems. Conversely, aggressively shortening reasoning can degrade performance on difficult ones. Effective reasoning therefore requires dynamically deciding when additional computation is useful based on the reasoner's capabilities and evolving solution state. Existing approaches often rely on predefined budgets or intervention rules, retrain the target reasoner, or require additional supervision. We introduce MetaCtrl, a lightweight controller that adaptively regulates a frozen reasoner without predefined token budgets or reasoner retraining. We formulate reasoning regulation as a sequential metacognitive control problem: MetaCtrl observes the evolving reasoning trace and decides whether to continue, simplify, skip redundant steps, or conclude reasoning. It is trained directly with reinforcement learning using a reward that prioritizes correctness while favoring shorter trajectories among correct solutions, requiring neither supervised intervention trajectories nor problem-specific budgets. Across seven benchmarks spanning mathematics, science, and code, MetaCtrl consistently improves the accuracy of LRMs while reducing their reasoning length. On DeepSeek-R1-Distill-Qwen-7B, it improves average accuracy by 4.7 points while reducing generation length by 53.3%. Without further training, the same controller transfers to an unseen reasoner (e.g., Qwen3-14B), improving average accuracy by 2.9 points and reducing generation length by 50.3%. These results establish MetaCtrl as a plug-and-play controller for improving reasoning accuracy while substantially reducing inference-time generation. The code is available at https://github.com/binbin2xs/MetaCtrl.
Zhibin Wen, Tao Han, Lei Bai +2
Southern University of Science and Technology · The Hong Kong University of Science and Technology · Shanghai Artificial Intelligence Laboratory +1
Large Language Models (LLMs) demonstrate strong reasoning performance, yet their ability to reliably monitor, diagnose, and correct their own errors remains limited. We introduce a psychologically grounded metacognitive framework that operationalizes Ann Brown's regulatory cycle (Planning, Monitoring, and Evaluation) as a structured prompting architecture, and study its integration within a lightweight dual-process MetaController for adaptive effort allocation. Across diverse reasoning and diagnostic benchmarks (GSM8K, CRUXEval, MBPP, AIME, CorrectBench, and TruthfulQA) using Llama-3 and Qwen-3 (8B), explicit regulatory structuring substantially improves error diagnosis and yields a threefold increase in successful self-correction. Blinded human evaluations over 580 query pairs show an 84% aggregate preference for trustworthiness and metacognitive self-awareness over standard and Chain-of-Thought baselines. Grounding LLM reasoning in established cognitive theory offers a principled path toward more transparent and diagnostically robust AI systems.
Abraham Paul Elenjical, Vivek Hruday Kavuri, Vasudeva Varma