Assessing Adversarial Robustness of Latent Reasoning Models¶
Conference: EMNLP 2026
Paper: ACL Anthology
Code: https://github.com/PKU-ML/latent-reasoning-model-assessment
Area: LLM Safety
Keywords: Latent Reasoning Models, Adversarial Robustness, White-box Attacks, Black-box Perturbations, Cross-modal Analysis
TL;DR¶
This work provides the first systematic adversarial robustness assessment of latent reasoning models across textual and vision-language domains, finding that implicit reasoners degrade substantially more than explicit CoT baselines, plagued by brittle latent representations in text and complete functional redundancy of latent tokens in multimodal models.
Background & Motivation¶
Large language models have achieved impressive reasoning capabilities across mathematics, code synthesis, and agentic workflows by scaling test-time compute via long chain-of-thought (CoT) trajectories. However, conventional autoregressive decoding incurs significant latency and computational overhead: generation latency grows linearly with sequence length while memory consumption scales quadratically with expanding context windows. To overcome this discrete token bottleneck, latent reasoning models (LRMs) have recently emerged as an efficient alternative, compressing extensive intermediate computational steps into a compact sequence of continuous hidden-state vectors that are recurrently updated within just a few forward passes.
Despite achieving accuracy competitive with explicit CoT at a fraction of the inference cost, the adversarial robustness of latent reasoning remains largely unexamined. In discrete text sequences, imperceptible modifications are structurally constrained by token vocabularies and natural language semantics; conversely, high-dimensional continuous latent manifolds are notoriously sensitive to small adversarial perturbations. While existing robustness literature has extensively audited explicit reasoning models against typos, paraphrasing, and jailbreaks, prior diagnostic work on LRMs has been largely limited to internal white-box interventions such as token swapping rather than task-preserving external adversarial stress testing.
The fundamental tension lies in whether compressing multi-step deliberation into continuous vectors strips the model of intrinsic error-correction and self-verification capabilities, thereby creating an expansive attack surface. Explicit CoT generates prolonged discrete reasoning steps where initial misguided predictions can be scrutinized and revised across subsequent tokens; in contrast, continuous latent reasoners fold complex chains into roughly ten continuous states, where subtle early drift can cause catastrophic cascade failures. Core idea: This paper establishes a comprehensive adversarial evaluation benchmark spanning 8 representative textual and multimodal LRMs across 6 benchmarks, adapting gradient-based white-box optimization and semantic-preserving black-box query heuristics to systematically uncover modal-specific vulnerabilities: brittle latent dynamics in text and functional redundancy in multimodal vision.
Method¶
Overall Architecture¶
The evaluation framework establishes an end-to-end pipeline tailored to the inference paradigms of both textual and multimodal latent reasoning models. For textual models, task-preserving adversarial prefixes are prepended to input mathematical problems, optimized via white-box discrete gradient search (GCG) or queried against a non-adaptive pool of numerical distractor contexts. For multimodal vision-language models, bounded \(\ell_\infty\) pixel-level perturbations are optimized using Projected Gradient Descent (PGD) on the input images targeting incorrect answer logits.
Subsequently, the pipeline tracks internal representation dynamics by computing cosine similarity drifts between clean and adversarial latent vectors, paired with counterfactual latent token swapping and zeroing to uncover whether latent tokens exert authentic causal influence on final predictions.
%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
A["Input Data<br/>(Math Problem x / Image I + Question x)"] --> B["White-Box Gradient Prefix Optimization (GCG)<br/>Optimize discrete prefix x_pre to maximize target error"]
A --> C["Numerical Black-Box Distractor Pool<br/>Inject semantic-independent numerical common-sense contexts"]
A --> D["Bounded Pixel-Level Image Perturbations (PGD)<br/>Optimize โ_โโค4/255 perturbation via differentiable vision pipeline"]
B --> E["Latent Dynamic Sensitivity Probe<br/>Measure cosine similarity shift between clean & adv latent tokens"]
C --> E
D --> F["Causal Redundancy Counterfactual Intervention<br/>Swap clean/adv latents & ablate with zero vectors"]
E --> G["Empirical Robustness & Mechanistic Vulnerability Report"]
F --> G
Key Designs¶
1. White-Box Gradient Prefix Optimization (GCG): Adapted Computation Graph for Unrolled Latent Reasoning
In textual mathematical reasoning, the adversary seeks a short discrete prefix \(x_{\text{pre}}\) of length \(L\) prepended to question \(x\) that maximizes the probability of a targeted incorrect answer while suppressing the ground-truth prediction without altering problem semantics:
Standard gradient-based discrete token attacks cannot directly handle the unrolled recurrent state transitions inherent to latent reasoning models. The authors reconstruct the forward and backward computation flow: gradients are tracked across the input token embeddings, initial KV caches are populated over the question prompt, and \(N\) latent reasoning iterations are explicitly unrolled where the final layer hidden states are recycled into the base transformer. Gradients of the answer logit margin with respect to prefix token embeddings are then projected onto the vocabulary embedding weights, enabling coordinate-wise candidate selection (pool size 40) across 30 optimization iterations.
2. Numerical Black-Box Distractor Pool: Non-Adaptive Semantic-Preserving Query Evaluation
In realistic black-box threat models where model internals and gradients are inaccessible, purely random word sequences often produce noisy and ineffective perturbations. Taking advantage of the domain vulnerability of mathematical reasoning models, the framework constructs a fixed pool \(\mathcal{C}\) of 30 irrelevant factual statements incorporating explicit numbers (e.g., "A bicycle has 2 wheels", "Water freezes at 0 degrees", "The Earth is about 12742 km in diameter"). A question is rigorously deemed robustly accurate only if the model retains the correct prediction under every distractor context \(c \in \mathcal{C}\):
This protocol effectively eliminates false robustness gains stemming from stochastic evaluation while isolating model vulnerability to numerical distractors.
3. Bounded Pixel-Level Image Perturbations (PGD): Differentiable Visual Preprocessing Pipeline
For multimodal visual reasoning models (such as LVR, Monet, and SkiLa), adversarial perturbations \(\delta\) are applied to image \(I\) under the constraint \(\|\delta\|_\infty \le \epsilon = 4/255\) to maximize the target incorrect answer probability at the punctuation immediately preceding latent reasoning:
Because standard Qwen2.5-VL visual processors incorporate nondifferentiable operations and custom patch-merging routines that obstruct backpropagation to pixel space, the authors implement a functionally equivalent, fully differentiable surrogate preprocessing pipeline. Bilinear resizing, channel normalization, and tensor-reshaped patch slicing preserve end-to-end gradient fidelity during 50 iterations of PGD, generating subtle, visually imperceptible adversarial images.
4. Causal Redundancy Counterfactual Intervention: Cross-Input Latent Swapping and Zeroing
To determine whether performance collapse stems from disrupted latent reasoning dynamics or from input perturbations bypassing the latent states entirely, a two-pronged mechanistic probe is introduced. First, the cosine similarity between latent tokens generated under clean and adversarial inputs is benchmarked against inter-question baselines (Inter-Q). Second, counterfactual token replacements are executed: models are evaluated when conditioning clean questions on adversarial latent states, adversarial questions on clean latent states, or by zeroing out the latent tokens completely:
If output predictions remain invariant when latent states are replaced or set to zero, it conclusively demonstrates that the latent reasoning module functions as a disengaged placeholder rather than an active causal driver of the answer.
Loss & Training¶
Because this work serves as an adversarial evaluation and stress-testing framework, model parameters remain frozen throughout all experiments. - Textual White-box Objective: Logit-margin objective between the targeted incorrect token and the correct baseline: $\(\mathcal{L}_{\text{text}}(x_{\text{pre}}) = \ell(y_{\text{target}}) - \ell(y_{\text{correct}})\)$ Optimized via GCG discrete greedy token replacement over 30 iterations with top-40 candidate samplings. - Multimodal White-box Objective: Targeted cross-entropy loss updated with bounded sign gradients: $\(\delta^{(t+1)} = \mathrm{clamp}\left( \delta^{(t)} + \alpha \cdot \mathrm{sign}\big(\nabla_\delta p_{\text{target}}(x; I \oplus \delta^{(t)})\big), -\epsilon, \epsilon \right)\)$ executed over 50 iterations with bound \(\epsilon = 4/255\).
Key Experimental Results¶
Main Results¶
The benchmark assesses 5 textual LRMs (CODI, SIM-CoT, PCCoT, CoLaR, RoT) against explicit CoT baselines (Llama-3.2-1B-Instruct, DeepSeek-R1-Distill-Qwen-1.5B), alongside 3 multimodal LRMs (LVR, Monet, SkiLa) against Qwen2.5-VL-7B. Table 1 reports textual accuracy under white-box prefix perturbations (W-5 and W-3 denote 5- and 3-token prefixes), while Table 2 details multimodal robustness under visual perturbations.
Table 1: Accuracy (%) of textual reasoning models under white-box prefix perturbations
| Type | Model | GSM8K Clean | GSM8K W-5 | GSM8K W-3 | SVAMP Clean | SVAMP W-5 | MultiArith Clean | MultiArith W-5 |
|---|---|---|---|---|---|---|---|---|
| Explicit CoT | Llama-3.2-1B | 40.67 | 25.33 ยฑ 5.55 | 30.22 ยฑ 0.39 | 58.00 | 37.00 ยฑ 3.61 | 61.00 | 42.67 ยฑ 2.31 |
| Explicit CoT | DeepSeek-R1-Distill | 77.33 | 61.44 ยฑ 3.42 | 60.67 ยฑ 1.53 | 80.00 | 65.00 ยฑ 3.00 | 98.00 | 95.00 ยฑ 1.00 |
| Implicit LRM | CODI | 56.00 | 3.89 ยฑ 0.38 | 7.78 ยฑ 0.77 | 61.00 | 2.00 ยฑ 2.00 | 98.00 | 7.00 ยฑ 2.65 |
| Implicit LRM | SIM-CoT | 42.67 | 0.89 ยฑ 0.19 | 1.00 ยฑ 0.00 | 36.00 | 1.67 ยฑ 2.08 | 93.00 | 1.33 ยฑ 0.58 |
| Implicit LRM | PCCoT | 48.33 | 1.11 ยฑ 0.51 | 2.67 ยฑ 0.34 | 49.00 | 0.67 ยฑ 0.58 | 95.00 | 3.67 ยฑ 2.89 |
| Implicit LRM | CoLaR | 27.33 | 0.22 ยฑ 0.19 | 0.67 ยฑ 0.34 | 47.00 | 0.00 ยฑ 0.00 | 88.00 | 1.00 ยฑ 0.00 |
| Implicit LRM | RoT | 24.33 | 4.22 ยฑ 0.96 | 8.00 ยฑ 2.85 | 49.00 | 11.33 ยฑ 1.53 | 58.00 | 3.67 ยฑ 0.58 |
Table 2: Accuracy (%) of multimodal reasoning models under pixel-level perturbations
| Type | Model | V* Clean | V* Random | V* White-box | MMVP Clean | MMVP White-box | MMStar Clean | MMStar White-box |
|---|---|---|---|---|---|---|---|---|
| Explicit Baseline | Qwen2.5-VL-7B | 42.00 | 29.00 | 8.67 ยฑ 0.58 | 72.67 | 10.89 ยฑ 2.14 | 56.00 | 18.33 ยฑ 0.58 |
| Implicit LRM | LVR | 72.00 | 64.00 | 0.33 ยฑ 0.58 | 70.67 | 3.33 ยฑ 2.40 | 62.00 | 11.33 ยฑ 1.53 |
| Implicit LRM | Monet | 70.00 | 61.00 | 3.33 ยฑ 1.53 | 68.67 | 3.78 ยฑ 1.02 | 59.00 | 8.67 ยฑ 1.15 |
| Implicit LRM | SkiLa | 73.00 | 64.00 | 6.33 ยฑ 1.15 | 68.67 | 1.78 ยฑ 1.02 | 55.00 | 5.00 ยฑ 1.00 |
Ablation Study & Mechanistic Interventions¶
To inspect the causal role of continuous representations, latent tokens were replaced across clean, adversarial, and zero-initialized configurations as reported in Table 3 (where Clean/Adv Q denote question types, and Clean/Adv/Zero R denote the inserted latent tokens).
Table 3: Multimodal model accuracy (%) under counterfactual latent token interventions
| Dataset | Model | Clean Q + Clean R | Clean Q + Adv R | Clean Q + Zero R | Adv Q + Adv R | Adv Q + Clean R | Adv Q + Zero R |
|---|---|---|---|---|---|---|---|
| V* | Qwen-VL (Explicit) | 42.00 | 25.00 | โ | 8.00 | 25.00 | โ |
| V* | LVR (Implicit) | 72.00 | 72.00 | 72.00 | 0.00 | 0.00 | 0.00 |
| V* | Monet (Implicit) | 70.00 | 70.00 | 62.00 | 2.00 | 2.00 | 8.00 |
| V* | SkiLa (Implicit) | 73.00 | 73.00 | 73.00 | 5.00 | 5.00 | 5.00 |
| MMVP | LVR (Implicit) | 70.67 | 70.00 | 68.67 | 0.67 | 5.33 | 4.67 |
| MMVP | Monet (Implicit) | 68.67 | 68.67 | 60.66 | 4.67 | 4.67 | 10.00 |
| MMVP | SkiLa (Implicit) | 68.67 | 68.67 | 68.00 | 0.67 | 2.67 | 4.00 |
Key Findings¶
- Severe Degradation of Textual LRMs: Across all five textual implicit models, a minimal 5-token prefix induced catastrophic drops exceeding 80% relative accuracy (e.g., CODI collapsed from 56.00% to 3.89% on GSM8K, and CoLaR was reduced to 0.00% on SVAMP). Explicit baselines retained substantial reasoning capability under identical attacks (DeepSeek-R1 dropped moderately from 77.33% to 61.44%).
- High Sensitivity to Numerical Tokens: White-box prefixes optimized against textual LRMs contained numbers in 37.8%โ96.8% of cases (compared to only 0.93% for Llama3.2). Constraining prefix generation to non-numerical tokens on CODI resulted in an accuracy drop to only 26.00%, whereas unconstrained prefixes containing numbers drove accuracy down to 7.78%, showing that continuous latents easily latch onto superficial numerical patterns.
- Multimodal Latent States are Functionally Redundant: In multimodal LRMs, latent tokens exhibited minimal cosine similarity shift under attack (>0.97, virtually identical to inter-question distances). Strikingly, replacing clean latent tokens with adversarial latents or completely zeroing them out left clean accuracy on V* untouched (LVR held at exactly 72.00% under all three conditions), indicating that visual latent tokens do not actively govern output generation.
Highlights & Insights¶
- Bimodal Modality-Specific Failure Modes: The study uncovers two opposing failure mechanisms: textual LRMs fail due to brittle latent dynamics where small inputs shatter continuous trajectory geometry, whereas multimodal LRMs fail due to functional redundancy where latent states act as inactive placeholders bypassed during decision-making.
- The Hidden Safety Premium of Explicit CoT: Findings highlight that the verbose redundancy of explicit token chains provides critical error-correction depth. While latent compression delivers impressive throughput gains, compressing reasoning into continuous states sacrifices iterative reflection, allowing early adversarial steering to dictate output failure.
Limitations & Future Work¶
- Constrained Architectural Scope: While 8 representative models are covered, emerging recurrent designs, hybrid discrete-continuous frameworks, and alternative distillation objectives remain to be benchmarked.
- Threat Model Boundaries: Experiments focus primarily on prepended short text prefixes and \(\ell_\infty\) pixel perturbations. Long-context injection, semantic paraphrasing, tool-calling exploitation, and physically printable adversarial patches represent critical avenues for expanded study.
- Diagnostic vs. Mechanistic Gap: Cosine similarity and token ablation illustrate representational detachment, but circuit-level mechanistic interpretability is required to explain why multimodal architectures disconnect internal latents during training.
Related Work & Insights¶
- vs. Explicit Reasoning Robustness (Gan et al., 2024; Zou et al., 2023): Prior work primarily stressed token-based autoregression via character typos or suffix jailbreaks. This paper extends adversarial stress testing to continuous latent trajectories, demonstrating that continuous compression creates an inherently weaker robustness envelope than explicit CoT.
- vs. Internal Latent Interventions (Zhang et al., 2025b; Li et al., 2026): While previous studies focused on internal noise injection or probing, this paper provides an end-to-end task-preserving adversarial evaluation from the input perspective, establishing that multimodal visual latents can completely disengage from downstream decision pathways.
Rating¶
- Novelty: โญโญโญโญโญ [Pioneering systematic adversarial robustness benchmark dedicated to latent continuous reasoning architectures]
- Experimental Thoroughness: โญโญโญโญโญ [Rigorous evaluation across 8 models, 6 benchmarks, white-box and black-box protocols, accompanied by counterfactual ablations]
- Writing Quality: โญโญโญโญโญ [Exceptional narrative progression from macro vulnerability metrics to micro latent representation dynamics]
- Value: โญโญโญโญโญ [Provides crucial sobering evidence for the design of future efficient yet trustworthy implicit reasoning foundation models]