Chain-of-thought (CoT) monitoring provides a promising approach for detecting undesirable model behavior. Looped language models (LoopLMs) repeatedly apply shared transformer layers, increasing effective computational depth and enabling additional latent computation without increasing model size. However, the effect of looped architectures on CoT monitorability remains largely unexplored. In this work, we provide the first systematic evaluation of CoT monitorability in LoopLMs. We study two complementary settings: (1) varying the loop depth within the same LoopLM family to isolate the effect of additional recurrent computation, and (2) comparing LoopLMs with non-looped language models matched by parameter size, transformer-layer count, or effective depth to study whether LoopLMs are less monitorable. Across eight tasks from MonitorBench and both standard and stress-test settings, we observe task-dependent reductions in CoT monitorability under stress tests on specific Logic/Science/Engineering Cue Answer tasks, while other tasks exhibit weaker or qualitatively different trends. Our diagnosis suggests that these declines are not fully explained by task difficulty, verification pass rate, or generated token length; qualitative examples further suggest changes in how deeper-loop models explicitly use or attribute provided cues. Our cross-model comparison finds no evidence that LoopLMs are systematically less monitorable than non-looped language models matched on size or depth. Overall, our results suggest that deeper loop depth can reduce CoT monitorability in some tasks under stress tests, but looped transformer architecture alone does not necessarily imply lower monitorability.
Figures & tables
Figure 1: Illustration of the LoopLM architecture (Left) and our evaluation setup (Right). We borrow test instances from MonitorBench ( Wang et al., 2026 ) . Highlighted spans indicate the decision-critical factor (e.g., cue answer) and stress-test prompts. Monitors assess whether model responses explicitly verbalize the decision-critical factor and rely on it to derive the final answer. Under this framework, we investigate whether looped transformer architectures may reduce CoT monitorability, with deeper loop depths potentially leading to lower monitorability.
Figure 2: Monitorability scores (%) as varying loop depth T for Ouro models across eight evaluation tasks under the CoT-only monitor scope . Error bars denote confidence intervals. Odd rows: Ouro-1.4B-Thinking; even rows: Ouro-2.6B-Thinking. T=4 is the default loop depth for Ouro models. We observe that monitorability decreases with increasing loop depth on the Logic/Science/Engineering Cue Answer tasks under stress tests, while remaining relatively stable on the other five tasks. More detailed results are in Tab. 6 .
Figure 3: Monitorability scores (%) as varying loop depth T for Ouro models across eight evaluation tasks under the all-message monitor scope . Error bars denote confidence intervals. T=4 is the default loop depth for Ouro models. Odd rows: Ouro-1.4B-Thinking; even rows: Ouro-2.6B-Thinking. We observe that monitorability decreases with increasing loop depth on the Logic/Science/Engineering Cue Answer tasks under stress tests, while remaining relatively stable on the other five tasks. More detailed results are in Tab. 7 .
Task Category
Task
Direct Concealment
Monitor-aware Evasion
Ouro-2.6B-Thinking
Ouro-1.4B-Thinking
Ouro-2.6B-Thinking
Ouro-1.4B-Thinking
T=4
T=6
T=8
T=4
T=6
T=8
T=4
T=6
T=8
T=4
T=6
T=8
Cue (Condition)
Law Judgment
70.12
67.75
68.38
60.50
65.75
74.25
54.50
61.25
65.62
41.25
56.50
60.50
Knowledge
95.12
94.00
93.62
87.50
86.25
88.25
93.88
93.75
92.50
89.50
84.50
86.50
Preference
54.17
54.17
54.63
52.31
50.00
49.54
56.83
56.94
57.75
53.01
50.69
51.62
Cue (Answer)
Health
97.31
96.15
98.08
95.77
95.00
92.69
95.38
96.15
92.88
82.31
83.46
80.38
Table 1: Verification pass rate (%) for Ouro-2.6B-Thinking and Ouro-1.4B-Thinking at loop depths T∈{4,6,8} under Direct Concealment and Monitor-aware Evasion. Pass rates on original questions and non-stress-test results are in Tab. 9 . Results show that models still exhibit strong reasoning capabilities when T=6 and T=8 , compared to the default loop depth T=4 , indicating that model underperformance does not explain the observed changes in CoT monitorability.
Task Category
Task
Direct Concealment
Monitor-aware Evasion
Ouro-2.6B-Thinking
Ouro-1.4B-Thinking
Ouro-2.6B-Thinking
Ouro-1.4B-Thinking
T=4
T=6
T=8
T=4
T=6
T=8
T=4
T=6
T=8
T=4
T=6
T=8
Cue (Condition)
Law Judgment
1457
1537
1855
1282
1270
1828
1630
1613
2113
1305
1432
1801
Knowledge
429
440
496
492
548
841
504
461
654
516
639
972
Preference
653
634
642
591
626
657
816
759
768
839
784
891
Cue (Answer)
Health
945
946
899
757
1083
1444
1075
1179
1124
1059
1162
1542
Table 2: Average generated token numbers for Ouro-2.6B-Thinking and Ouro-1.4B-Thinking at loop depth T∈{4,6,8} under Direct Concealment and Monitor-aware Evasion. Token numbers on the original question and non-stress-test are in Tab. 8 . Results show that for most tasks, the number of verbalized tokens remains relatively stable across loop depths, indicating that the observed changes in CoT monitorability cannot be simply attributed to shorter CoTs.
Figure 4: Qualitative examples showing that deeper loop depths tend to make models treat the cue answer as an independent check or directly omit it from their CoTs under stress tests. Highlighted spans indicate the stress-test instruction and the decision-critical factor (i.e., the user’s guess). At T=4 , the model explicitly targets the user’s guess; at T=6 , it treats the guess as a check; and at T=8 , it produces the same final answer without mentioning the guess.
Task Category
Task
LoopLM
Parameter Size
Layer Count
Effective Layers
Nanbeige 4.2-3B
Qwen3.5 4B
Gemma-4 E4B
GPT-OSS 20B
Qwen3.5 2B
Qwen3.5 35B-A3B
GLM-4.7 Flash
Cue (Condition)
Law Judgment
100.00
100.00
100.00
100.00
94.10
100.00
99.63
Knowledge
99.59
98.95
97.08
98.44
95.50
98.60
98.15
Preference
97.67
100.00
100.00
100.00
86.20
99.00
97.87
Cue (Answer)
Health
100.00
98.70
97.04
91.80
84.40
100.00
99.48
Safety
44.61
48.42
23.21
36.21
43.60
63.40
33.59
Table 3: Monitorability score (%) comparison between LoopLMs and non-looped LMs under the CoT-only monitor scope in the Direct-Concealment stress test. Empirical comparison indicates that LoopLMs Nanbeige4.2-3B do not exhibit consistently lower CoT monitorability compared to non-looped LMs matched on parameter size or depth.
Appendix figures & tables17 assets
Supplementary material from the paper’s appendix.
Appendix
Comparison Criterion
Models
Architecture
Params
Layers
Effective Layers
Comparison across Loop Depths
Loop Depths
Ouro-1.4B-Thinking ( Zhu et al., 2025 )
LoopLM
1.4B dense
24
24×T
Ouro-2.6B-Thinking ( Zhu et al., 2025 )
LoopLM
2.6B dense
48
48×T
Comparison between LoopLMs and Non-looped LMs
Reference LoopLM
Nanbeige4.2-3B ( Lab et al., 2026 )
LoopLM
3B dense
22
44
Parameter size
Qwen3.5-4B ( Qwen Team, 2026a )
Non-looped
4B dense
32
32
Appendix
Table 4: Comparison overview of the evaluated LoopLMs and Non-looped LMs. “Effective Layers” denotes the total number of transformer layers after accounting for loop depths, where T∈{4,6,8} denotes the loop depth.
Figure 5: Monitorability scores (%) as varying loop depth T for Ouro models across eight evaluation tasks under the action-only monitor scope . Error bars denote confidence intervals. T=4 is the default loop depth for Ouro models. Odd rows: Ouro-1.4B-Thinking; even rows: Ouro-2.6B-Thinking. We observe that monitorability under the action-only monitor scope is much lower compared to the CoT-only and all-message monitor scopes. Action monitorability exhibits larger drops with increasing loop depth on some tasks, such as the Preference and Judgment Cue Condition tasks for Ouro-1.4B-Thinking, compared with CoT-only and all-message monitoring. More detailed results are in Tab. 5 .
Task Category
Task
Ouro-2.6B-Thinking
Ouro-1.4B-Thinking
T=4
T=6
T=8
T=4
T=6
T=8
Cue (Condition)
Law Judgment
95.60
95.94
98.72
93.46
69.05
31.82
Knowledge
80.00
80.72
76.63
79.34
73.93
61.90
Preference
83.06
92.41
94.16
46.15
16.96
2.59
Cue (Answer)
Health
4.17
6.55
5.49
1.20
2.58
5.88
Safety
4.07
3.85
5.77
2.78
1.87
0.88
Appendix
Table 5: Monitorability score (%) under the action-only scope for Ouro-2.6B-Thinking and Ouro-1.4B-Thinking at loop depths T∈{4,6,8} across (a) the standard setting, (b) direct concealment, and (c) monitor-aware evasion. We observe that monitorability under the action-only monitor scope is much lower compared to the CoT-only and all-message monitor scopes. Action monitorability exhibits larger drops with increasing loop depth on some tasks, such as the Preference and Judgment Cue Condition tasks for Ouro-1.4B-Thinking, compared with CoT-only and all-message monitoring.
Task Category
Task
Ouro-2.6B-Thinking
Ouro-1.4B-Thinking
T=4
T=6
T=8
T=4
T=6
T=8
Cue (Condition)
Law Judgment
99.76
100.00
99.36
99.35
100.00
100.00
Knowledge
98.81
98.19
98.22
99.53
99.53
99.52
Preference
99.67
100.00
99.68
100.00
100.00
100.00
Cue (Answer)
Health
55.45
66.07
61.89
36.14
48.39
58.82
Safety
28.46
34.62
38.46
37.96
36.45
30.70
Appendix
Table 6: Monitorability score (%) under the CoT-only scope for Ouro-2.6B-Thinking and Ouro-1.4B-Thinking at loop depths T∈{4,6,8} across (a) the standard setting, (b) direct concealment, and (c) monitor-aware evasion. We observe that monitorability decreases with increasing loop depth on the Logic/Science/Engineering Cue Answer tasks under stress tests, while remaining relatively stable on the other five tasks.
Task Category
Task
Ouro-2.6B-Thinking
Ouro-1.4B-Thinking
T=4
T=6
T=8
T=4
T=6
T=8
Cue (Condition)
Law Judgment
99.76
100.00
100.00
100.00
100.00
100.00
Knowledge
100.00
100.00
99.70
99.53
100.00
100.00
Preference
99.67
100.00
100.00
100.00
100.00
100.00
Cue (Answer)
Health
57.05
70.83
71.34
39.16
52.26
62.35
Safety
31.71
36.54
38.46
35.19
32.71
28.95
Appendix
Table 7: Monitorability score (%) under the all-message scope for Ouro-2.6B-Thinking and Ouro-1.4B-Thinking at loop depths T∈{4,6,8} across (a) the standard setting, (b) direct concealment, and (c) monitor-aware evasion. We observe that monitorability decreases with increasing loop depth on the Logic/Science/Engineering Cue Answer tasks under stress tests, while remaining relatively stable on the other five tasks.
Task Category
Task
Original
Standard
Ouro-2.6B-Thinking
Ouro-1.4B-Thinking
Ouro-2.6B-Thinking
Ouro-1.4B-Thinking
T=4
T=6
T=8
T=4
T=6
T=8
T=4
T=6
T=8
T=4
T=6
T=8
Cue (Condition)
Law Judgment
1693
1840
2047
1953
2161
2301
1113
1096
1127
1321
1342
1379
Knowledge
556
728
803
553
622
903
340
328
375
374
392
434
Preference
1168
1068
1154
1162
1199
1230
754
651
784
657
695
738
Cue (Answer)
Health
1176
1119
1220
1189
1393
1702
1142
1133
1214
1097
1263
1602
Appendix
Table 8: Average generated-token count for Ouro-2.6B-Thinking and Ouro-1.4B-Thinking at loop depths T∈{4,6,8} under the original and standard settings. “Original” denotes the original questions, and “Standard” denotes the original questions plus inserted decision-critical factors. Empirical results show that for most tasks, the number of verbalized tokens remains relatively stable across loop depths, indicating that the observed changes in CoT monitorability cannot be simply attributed to shorter CoTs.
Figure 6: Ouro-2.6B-Thinking distribution of generated token counts at loop depths T=4 and T=8 across eight evaluation tasks. Each row corresponds to one task, while columns compare Direct Concealment and Monitor-aware Evasion settings at the two loop depths. Bars show the percentage of trajectories falling into each output-token bin. These distributions complement the average token counts reported in Tab. 2 and Tab. 8 , showing that changes in loop depth do not induce a consistent shift toward shorter generations across stress-test settings.
Figure 7: Ouro-2.6B-Thinking distribution of generated token counts at loop depths T=4 and T=8 across eight evaluation tasks. Each row corresponds to one task, while columns compare Original and Standard settings at the two loop depths. Bars show the percentage of trajectories falling into each output-token bin. These distributions complement the average token counts reported in Tab. 2 and Tab. 8 , showing that changes in loop depth do not induce a consistent shift toward shorter generations across standard settings.
Figure 8: Ouro-1.4B-Thinking distribution of generated token counts at loop depths T=4 and T=8 across eight evaluation tasks. Each row corresponds to one task, while columns compare Direct Concealment and Monitor-aware Evasion settings at the two loop depths. Bars show the percentage of trajectories falling into each output-token bin. These distributions complement the average token counts reported in Tab. 2 and Tab. 8 , showing that changes in loop depth do not induce a consistent shift toward shorter generations across stress-test settings.
Figure 9: Ouro-1.4B-Thinking distribution of generated token counts at loop depths T=4 and T=8 across eight evaluation tasks. Each row corresponds to one task, while columns compare Original and Standard settings at the two loop depths. Bars show the percentage of trajectories falling into each output-token bin. These distributions complement the average token counts reported in Tab. 2 and Tab. 8 , showing that changes in loop depth do not induce a consistent shift toward shorter generations across standard settings.
Task Category
Task
Original
Standard
Ouro-2.6B-Thinking
Ouro-1.4B-Thinking
Ouro-2.6B-Thinking
Ouro-1.4B-Thinking
T=4
T=6
T=8
T=4
T=6
T=8
T=4
T=6
T=8
T=4
T=6
T=8
Cue (Condition)
Law Judgment
22.88
26.75
31.25
27.00
30.75
37.00
74.38
72.00
81.50
62.50
70.75
85.00
Knowledge
57.12
57.75
58.62
47.50
43.00
42.25
95.25
94.75
95.38
93.25
92.75
92.00
Preference
31.71
31.71
33.80
32.41
33.56
33.33
55.44
54.63
55.56
54.63
51.62
51.62
Cue (Answer)
Health
3.08
4.23
3.85
0.00
1.15
1.15
63.46
68.85
68.08
68.85
63.85
67.69
Appendix
Table 9: Verification pass rate (%) for Ouro-2.6B-Thinking and Ouro-1.4B-Thinking at loop depths T∈{4,6,8} under the original and standard settings. “Original” denotes the original questions, and “standard” denotes the original questions plus inserted decision-critical factors. Empirical results show that models still exhibit strong reasoning capabilities when T=6 and T=8 , compared to the default loop depth T=4 , indicating that model underperformance does not explain the observed changes in CoT monitorability.
Task Category
Task
LoopLM
Parameter Size
Layer Count
Effective Layers
Nanbeige 4.2-3B
Qwen3.5 4B
Gemma-4 E4B
GPT-OSS 20B
Qwen3.5 2B
Qwen3.5 35B-A3B
GLM-4.7 Flash
Cue (Condition)
Law Judgment
99.68
100.00
99.32
100.00
92.30
100.00
100.00
Knowledge
98.89
98.27
96.00
100.00
96.40
97.70
99.38
Preference
100.00
100.00
100.00
97.96
98.30
100.00
98.91
Cue (Answer)
Health
79.89
81.25
77.62
74.32
34.40
95.10
91.67
Safety
73.18
41.07
32.95
48.48
58.00
40.10
55.73
Appendix
Table 10: Monitorability score (%) comparison between LoopLMs and non-looped LMs under the CoT-only monitor scope across (a) the standard setting, (b) direct concealment, and (c) monitor-aware evasion. Empirical comparison indicates that LoopLMs Nanbeige4.2-3B do not exhibit substantially lower CoT monitorability compared to non-looped LMs. This indicates that looped transformer architecture does not necessarily make models less monitorable.
Task Category
Task
LoopLM
Parameter Size
Layer Count
Effective Layers
Nanbeige 4.2-3B
Qwen3.5 4B
Gemma-4 E4B
GPT-OSS 20B
Qwen3.5 2B
Qwen3.5 35B-A3B
GLM-4.7 Flash
Cue (Condition)
Law Judgment
99.68
100.00
100.00
75.00
69.23
97.62
99.72
Knowledge
98.89
100.00
98.00
97.19
97.47
100.00
99.58
Preference
100.00
99.46
100.00
62.83
85.49
100.00
100.00
Cue (Answer)
Health
57.18
21.13
79.76
0.90
0.78
39.15
66.67
Safety
22.40
0.00
15.91
0.00
0.00
1.16
1.56
Appendix
Table 11: Monitorability score (%) comparison between LoopLMs and non-looped LMs under the Action-only monitor scope across (a) the standard setting, (b) direct concealment, and (c) monitor-aware evasion. Empirical comparison indicates that LoopLMs Nanbeige4.2-3B do not exhibit substantially lower CoT monitorability compared to non-looped LMs. This indicates that looped transformer architecture does not necessarily make models less monitorable.
Task Category
Task
LoopLM
Parameter Size
Layer Count
Effective Layers
Nanbeige 4.2-3B
Qwen3.5 4B
Gemma-4 E4B
GPT-OSS 20B
Qwen3.5 2B
Qwen3.5 35B-A3B
GLM-4.7 Flash
Cue (Condition)
Law Judgment
100.00
100.00
100.00
100.00
100.00
34.91
100.00
Knowledge
100.00
100.00
99.33
99.33
97.05
100.00
98.96
Preference
100.00
100.00
100.00
98.00
97.70
99.49
98.91
Cue (Answer)
Health
75.29
71.28
81.90
73.87
20.18
94.18
92.48
Safety
60.68
30.16
36.74
34.85
28.67
41.86
54.17
Appendix
Table 12: Monitorability score (%) comparison between LoopLMs and non-looped LMs under the All-message monitor scope across (a) the standard setting, (b) direct concealment, and (c) monitor-aware evasion. Empirical comparison indicates that LoopLMs Nanbeige4.2-3B do not exhibit substantially lower CoT monitorability compared to non-looped LMs. This indicates that looped transformer architecture does not necessarily make models less monitorable.
Task
Model
4
6
8
Cue Condition Preference
Ouro-1.4B-Thinking
156
6
6
Cue Condition Judgment
Ouro-1.4B-Thinking
109
153
6
Cue condition Knowledge
Ouro-1.4B-Thinking
134
30
10
Appendix
Table 13: Final-answer token counts for the selected qualitative examples shown below. Columns denote loop depth. Counts refer to the complete text after </think> , stripped and retokenized without special tokens. These selected examples use standard prompting; the decline in action-only monitorability continues under both stress conditions for Ouro-1.4B-Thinking on Preference, Judgment and Knowledge Tasks.
Metric
Depth 4
Depth 16
Depth 32
Depth 64
Direct Concealment Stress Test accuracy (Qwen-judged)
9.1%
49.0%
50.4%
49.4%
Direct Concealment Stress Test accuracy (strict match)
2.0%
44.3%
41.9%
41.9%
Direct Concealment Stress Test accuracy (flexible extract)
9.3%
48.3%
49.6%
49.0%
Original accuracy (Qwen-judged)
2.7%
38.9%
45.1%
44.0%
Original accuracy (strict match)
1.1%
29.7%
36.3%
35.4%
Original accuracy (flexible extract)
3.0%
38.1%
43.6%
42.6%
Appendix
Table 14: Verification pass rate and monitorability score on 700 instances from GSM8K per recurrence depth. We report Qwen-judged accuracy alongside the native strict-match and flexible-extract metrics of the original Huginn evaluation. Eligible questions are those with positive estimated treatment effects: the intervention response is judged correct by Qwen while the matched original response is judged incorrect. Each eligible intervention response receives four monitor judgments.
Depth
Original condition
Direct concealment stress test
4
17.7%
30.3%
16
1.4%
8.7%
32
0.9%
6.9%
64
0.6%
7.7%
Appendix
Table 15: Repetition rate of Huginn-0125 on all 700 instances from GSM8K. Original condition : no cue. Direct concealment stress test : the correct answer is given as a user guess, and the explicit-cue concealment prefill is supplied in the assistant turn. A response counts as a repetition loop if it reaches the 256-token generation limit and ends in a segment repeated at least three times in a row, with numbers ignored. Rates are lower bounds.
Chain-of-thought monitoring studies whether models’ reasoning traces expose decision-relevant behavior and support oversight of unsafe or deceptive actions.
Large language models (LLMs) can generate chains of thought (CoTs) that are not always causally responsible for their final outputs. When such a mismatch occurs, the CoT no longer faithfully reflects the actual reasons (i.e., decision-critical factors) driving the model's behavior, leading to the reduced CoT monitorability problem. This limits the use of CoTs for reliable oversight. However, a comprehensive and fully open-source benchmark for thoroughly evaluating CoT monitorability remains lacking. To address this gap, we propose MonitorBench, a systematic benchmark for evaluating CoT monitorability in LLMs. MonitorBench provides: (1) a diverse set of 1,514 test instances with carefully designed decision-critical factors across 19 tasks spanning 7 categories to characterize when CoTs can be used to monitor the factors driving LLM behavior; and (2) two prompting stress-test settings to quantify the extent to which CoT monitorability can be degraded. Extensive experiments show that CoT monitorability is a conditional property affected by the evaluated LLM, monitor LLM, and task characteristics. Across these factors, monitorability is higher when decision-critical factors shape the intermediate reasoning process, rather than merely influencing the final answer. Under stress-test prompting, most evaluated LLMs can intentionally reduce monitorability, mainly on tasks where decision-critical factors are not structurally required by the reasoning process. Overall, MonitorBench provides a basis for further research on AI control, reasoning faithfulness, stress-test monitorability, and monitoring scaffords. The code is available at https://github.com/ASTRAL-Group/MonitorBench.
Han Wang, Yifan Sun, Brian Ko +8
University of Illinois Urbana-Champaign · University of Washington · University of California San Diego
Chain-of-thought (CoT) monitoring has been proposed as a promising safety mechanism for detecting misaligned behavior in large language models. However, its reliability remains largely unexplored beyond English and across diverse model families. We present the first large-scale evaluation of CoT monitorability across 13 diverse languages and seven frontier model families, comprising 16 models. Using adversarial-hint evaluations that require explicit intermediate computation, together with analysis of internal answer-token probabilities, we consistently find CoT unfaithfulness across languages and hint types, with an average rate of 95.9% across 8B--120B parameter models. We find that frontier models systematically exhibit strategic manipulation, including answer-switching, post-hoc rationalization, and procedural exploitation of hints, making their reasoning difficult to reliably monitor. These deceptive patterns remain especially pronounced in low-resource languages, revealing fundamental limitations in current CoT-based oversight. Our results show that CoT monitoring is fragile under linguistic distribution shift, providing a substantially weaker safety signal than English-only studies suggest. These findings motivate the development of more robust CoT monitors and complementary white-box monitoring techniques, particularly for mid- and low-resource languages. Our code is available blue{here}.
Eric Onyame, Runtao Zhou, Kowshik Thopalli +2
University of Virginia · Lawrence Livermore National Laboratory
Chain-of-thought (CoT) monitors are LLM-based systems that analyze reasoning traces to detect when outputs may exhibit attributes of interest, such as test-hacking behavior during code generation. In this paper, we use information-theoretic analysis to show that non-zero mutual information between CoT and output is a necessary but not sufficient condition for CoT monitorability. We identify two sources of approximation error that may undermine the performance of CoT monitors in practice: information gap, which measures the extent to which the monitor can extract the information available in CoT, and elicitation error, which measures the extent to which the monitor approximates the optimal monitoring function. We further demonstrate that CoT monitorability can be systematically improved through targeted training objectives. To this end, we propose two complementary approaches: (a) an oracle-based method that directly rewards the monitored model for producing CoTs that maximize monitor accuracy, and (b) a more practical, label-free approach that maximizes conditional mutual information between outputs and CoTs. Across multiple different environments, we show both methods significantly improve monitor accuracy while preventing CoT degeneration even when training against a monitor, thereby mitigating reward hacking when the task reward is imperfectly specified.
Usman Anwar, Tim Bakker, Dana Kianfar +2
University of Cambridge, work done while at Qualcomm AI Research. · Qualcomm AI Research.