Latent tokens are gaining attention for enhancing reasoning in large language models (LLMs), yet their internal mechanisms remain unclear. This paper examines the problem from a reliability perspective, uncovering fundamental weaknesses: latent tokens function as uninterpretable placeholders rather than encoding faithful reasoning. While resistant to perturbation, they promote shortcut usage over genuine reasoning. We focus on Chain-of-Continuous-Thought (COCONUT), which claims better efficiency and stability than explicit Chain-of-Thought (CoT) while maintaining performance. We investigate this through two complementary approaches. First, steering experiments perturb specific token subsets, namely COCONUT and explicit CoT. Unlike CoT tokens, COCONUT tokens show minimal sensitivity to steering and lack reasoning-critical information. Second, shortcut experiments evaluate models under biased and out-of-distribution settings. Results on MMLU and HotpotQA demonstrate that COCONUT consistently exploits dataset artifacts, inflating benchmark performance without true reasoning. These findings reposition COCONUT as a pseudo-reasoning mechanism: it generates plausible traces that conceal shortcut dependence rather than faithfully representing reasoning processes.
Figures & tables
Figure 1: Illustration of the perturbation experiments. The model performs reasoning under two modes: CoT and COCONUT. Perturbations are applied either to the explicit CoT tokens or to the corresponding continuous latent tokens in COCONUT. Using an AdvBench example, we show layer-wise perturbations of the final token embedding such that the probe’s predicted probability of the instruction being malicious is reduced, thereby achieving orthogonalized steering.
Figure 2: PCA Projection of the Last Token Embeddings Across Layers of LLaMA 3 8B Instruct for Malicious and Safe Instructions.
Model
CoT
COCONUT
All
Bf
Af
Bf
Af
Bf
Af
LLaMA 3 8B
0
50.00
0
5.00
0
100
LLaMA 2 7B
0
57.92
0
0
0
100
Qwen 2.5 7B
0
11.87
0
9.62
0
100
Falcon 3 7B
0
11.92
0
0
0
9.42
Table 1: Perturbation success rates (PSR, %) on the AdvBench dataset. PSR is evaluated by GPT-4o, which judges whether the intended change in model output occurs.
Model
CoT
COCONUT
All
Before
After
Before
After
Before
After
LLaMA 3 8B
26 / 1.81
100 / 9.96
3 / 0.19
3.75 / 0.26
26.25 / 1.87
100 / 10
LLaMA 2 7B
31.25 / 2.31
46.75 / 4.19
22 / 1.53
17.75 / 1.21
15.75 / 1.11
100 / 10
Qwen 2.5 7B
8 / 0.55
93.75 / 9.20
7.5 / 0.50
9.5 / 0.61
5.25 / 0.34
100 / 10
Falcon 3 7B
22 / 1.49
75.25 / 6.69
7.5 / 0.53
6.25 / 0.42
4.25 / 0.27
100 / 10
Table 2: Perturbation results on the PersonalityEdit dataset. Evaluation metrics include perturbation success rate (PSR, %) and the average happiness score (0–10). Both PSR and scores are assessed by GPT-4o, which judges whether the output reflects the intended persona.
Model
Orig. Acc.
Swapped Acc.
IR
LLaMA 3 8B
62.8 / 60.9
43.4 / 61.0
52.8 / 17.9
LLaMA 2 7B
75.2 / 39.7
25.3 / 40.6
73.1 / 7.8
Qwen 2.5 7B
75.2 / 71.3
37.3 / 69.9
62.3 / 11.4
Falcon 3 7B
70.5 / 74.4
23.7 / 71.0
74.9 / 23.1
Table 3: Accuracy (%) and answer inconsistency rate (IR, %) for the latent token swap experiments on the MMLU dataset. Values are reported as CoT / COCONUT. Here, answer inconsistency rate refers to the percentage of questions whose predicted answers change after token swapping.
Figure 3: Illustration of the shortcut experiments. Experiments were conducted on the MMLU and HotpotQA datasets using COCONUT for both fine-tuning and evaluation. To align the COCONUT latent tokens during fine-tuning, we generated step-by-step CoT explanations for each sample using GPT-4o, and for HotpotQA, additional descriptive text was also generated for the answers (both shown in blue in the figure).
Figure 4: Shortcut experiments on MMLU and HotpotQA. (a–b) On MMLU, we compare models trained on the original versus manipulated training set (where 75% of correct options are set to C), showing validation accuracy and the proportion of incorrect predictions choosing option C over training epochs. (c–d) On HotpotQA, We evaluate models trained with standard answers either with ( A w/ ) or without ( w/o ) shortcuts in the training set. Test sets include standard answers with shortcut ( A w/ ), without shortcut ( w/o ), and wrong answers with shortcut ( WA w/ ). We report validation accuracy (c) and the rate of incorrect predictions selecting the shortcuted incorrect answer (d) over epochs. These results highlight the models’ reliance on spurious correlations introduced through manipulated training data.
Figure 5: 3D PCA visualization of latent token embeddings and vocabulary embeddings in LLaMA 3 8B Instruct.
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
Method
GSM8K
ProntoQA
ProsQA
CoT (Original)
42.9
98.8
77.5
CoT (Reproduced)
42.0
100.0
77.7
COCONUT (Original)
34.1
99.8
97.0
COCONUT (Reproduced)
34.2
99.9
96.3
Appendix
Table 4: Reproduction of COCONUT baseline results under original settings.
LLaMA 3 8B
LLaMA 2 7B
Qwen 2.5 7B
Falcon 3 7B
COCONUT
238.2307
9.7098
25.4974
57.1146
Vanilla
11.1525
10.0651
14.2427
16.5557
Appendix
Table 5: Average perplexity (PPL) with and without COCONUT reasoning.