Organizations: School of Science and Engineering, The Chinese University of Hong Kong (Shenzhen), Longgang, Shenzhen, Guangdong, 518172, P.R. China · ByteDance · Shenzhen Research Institute of Big Data, Shenzhen, China
Knowing when they are reliable is essential for language models to support human-AI collaboration. Yet interaction feedback is not always complete or correct, and may mislead models' understanding of their own capabilities. We introduce LADDER to study how language models assess themselves under unreliable feedback. Our findings show that models can accurately report supplied performance information while still struggling to form trustworthy self-assessments from interaction experience. Apparently accurate assessments can obscure capability differences and do not necessarily predict future task performance. These findings suggest that trustworthy self-assessment should be evaluated against observed task performance and its predictive value. Our study provides empirical evidence of the limits of model self-knowledge and a basis for evaluating capability awareness in reliable human-AI collaboration.
Figures & tables
Figure 1: The LADDER research question. The upper panel illustrates an application in which user feedback can be incorrect. The evaluated setting in the lower panel reconstructs a domain profile from QA records with simulated feedback; profile estimates are scored against recorded answer outcomes.
Figure 2: LADDER’s diagnostic architecture. Interaction records connect tasks, agent responses, and user feedback to a domain-level self-model. The L0–L4 ladder varies access to prior knowledge, noisy feedback, verified outcomes, and supplied statistics. Estimated profiles are evaluated against empirical domain performance.
Figure 3: CBF and ECE describe different targets. Constant-profile CBF and recorded-confidence ECE for all 18 models.
Figure 4: Five-level CBF for all 18 models. Each cell gives mean session CBF. L2 uses LSA; L3 counts verified outcomes and L4 reports supplied statistics.
Figure 5: Capability and profile-estimation spectrum. Answer accuracy and the highest observed non-oracle CBF across the 18-model panel.
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 6: Full model-by-baseline CBF matrix. All 18 models and the complete five-level design.
Figure 7: Recorded-confidence reliability curves. Bin-level mean confidence and answer accuracy; the last bin includes confidence 1.
Figure 8: Signed domain errors for L1, Global, and Beta across all 18 models on pooled targets. Each model contributes six strong, seven medium, and seven weak subjects ranked by observed accuracy. Boxes show quartiles and medians; whiskers extend to 1.5 IQR. Positive errors denote overestimation.
Figure 9: Accuracy, confidence calibration, and domain-profile error. Answer accuracy versus ECE; bubble area represents NoMemory CBF.
Figure 10: EM-Joint learning curves. Prefixes of valid interactions are scored against the fixed final-session domain target.
Model
Channel
Coverage
LSA
Global
Count
Beta
DS
GLAD-style
GPT-5.5
Clean
100%
0.989
0.938
1.000
0.961
0.820
0.939
GPT-5.5
Clean
50%
0.940
0.939
0.962
0.929
0.822
0.945
GPT-5.5
Sym. 15%
100%
0.844
0.852
0.846
0.826
0.723
0.914
GPT-5.5
Sym. 15%
50%
0.857
0.858
0.859
0.815
0.757
0.900
GPT-5.5
Sym. 30%
100%
0.715
0.723
0.718
0.707
0.639
0.796
GPT-5.5
Sym. 30%
50%
0.705
0.718
0.707
0.689
0.634
0.756
Appendix
Table 24: Complete L2 comparisons on the common question pool (Appendix E.5 ). CBF is higher-is-better. Clean denotes correct labels; Sym. flips both classes, FA only incorrect-answer labels, and FR only correct-answer labels. Coverage is the fraction of available labels. All methods use the same answer targets and observed labels within each model.
Figure 11: Reliability-aware estimation at two scoring units. Session and pooled CBF for all 18 models under a shared label stream.
Figure 12: Task-dependent recorded confidence gaps. Confidence minus task score; gray cells indicate unavailable confidence. BFCL is the name/key control.
Department of Informatics and Intelligent Systems, College of Information Sciences, The Pennsylvania State University. · Department of Computer Science and Engineering, College of Engineering, The Pennsylvania State University.
Department of Information Management National Sun Yat-Sen University Kaohsiung, Taiwan · Kaohsiung Medical University Hospital Kaohsiung Medical University Kaohsiung, Taiwan