Resource Consumption Threats in Large Language Models: A Survey¶
Conference: EMNLP 2026
Paper: ACL Anthology
Code: https://github.com/shuita2333/resource-consumption-threats-in-llms
Area: LLM Safety
Keywords: Resource Consumption Threats, Denial-of-Service, Overthinking, Unbounded Drift, Inference Security
TL;DR¶
This survey provides the first comprehensive synthesis of resource consumption threats in LLMs, introducing an outcome-based taxonomy that categorizes attacks into non-degenerate "Overthinking" and degenerate "Unbounded Drift", while dissecting underlying generation dynamics, active defenses, and standardized evaluation protocols.
Background & Motivation¶
The real-world serving of generative large language models (LLMs) operates under stringent computational and memory bottlenecks, where inference throughput and latency dictate deployment feasibility and economic viability. With the advent of long-context LLMs, reasoning models (RLLMs), vision-language architectures (MLLMs), and autonomous agent frameworks expanding token budgets to hundreds of thousands or even millions of tokens, the upper bound of operational resource investment per interaction has increased dramatically. Recent security investigations reveal a troubling paradigm shift: computational resource dependence is actively being weaponized by adversaries. Instead of attempting to elicit toxic text or extract private training data, attackers craft inputs designed to induce disproportionate, redundant, or non-terminating generation, amplifying infrastructure burdens and undermining service availability.
Despite its severe implications, research in this domain has remained fragmented and conceptually ambiguous across different subfields. Previous inquiries have separately studied length-increasing prompt attacks in text LLMs, reasoning over-expansion in chain-of-thought models, verbose image triggers in multimodal settings, and recursive execution loops in agentic tool-use environments. Crucially, the operational boundary between malicious resource amplification and legitimate, compute-heavy inference (such as thorough mathematical self-verification) remains poorly delineated. In commercial black-box deployments, providers lack reliable criteria to distinguish benign resource-intensive requests from covert denial-of-service attempts, rendering naive length thresholds and standard rate limits either prone to severe false positives or ineffective against non-halting trajectories.
To resolve this fragmentation, this survey reframes resource consumption from a traditional efficiency optimization problem into an urgent foundational safety concern. The authors argue that protecting serving infrastructure against adversarial compute exhaustion is just as vital as defending against jailbreaking or data exfiltration. Core idea: The survey establishes an outcome-based taxonomy that bifurcates resource consumption threats into bounded, non-degenerate "Overthinking" and self-sustaining, degenerate "Unbounded Drift", providing operational criteria grounded in internal decoding dynamics and bridging attack triggers, generation mechanisms, active defenses, and standardized benchmark metrics.
Method¶
Overall Architecture¶
The survey formulates a unified analytical pipeline that systematically contextualizes resource consumption risks across generative architectures. The overarching architecture progresses hierarchically: adversarial triggers across modalities induce anomalous generation trajectories, which are diagnostically partitioned into the Overthinking and Unbounded Drift outcome regimes; these regimes are then explained via microscopic decoding dynamics and representation biases, mitigated through targeted model-side and system-level active defenses, and audited via a three-tiered evaluation protocol.
%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
A["Adversarial Input Triggers<br/>(Text Prompts / Visual Perturbations / Tool Outputs)"] --> B["Outcome-Based Taxonomy<br/>(Overthinking vs. Unbounded Drift)"]
B --> C["Underlying Generation Dynamics<br/>(Induction Head Toxicity / Low-Entropy Lock-in / Weak EOS)"]
C --> D["Active Multi-Tier Defenses<br/>(Utility-Aware Gating / Recurrent Detection / Chunk Interruption)"]
D --> E["Three-Level Standardized Benchmark<br/>(Generation Behavior / Hardware Load / Service Throughput)"]
Key Designs¶
1. Outcome-Based Taxonomy: Resolving threat ambiguity across heterogeneous interfaces
To unify disparate attack vectors that span textual prompt injection, multimodal triggers, and multi-agent interaction chains, the authors organize threats strictly by observable generation trajectory outcomes rather than transient technical implementations. The taxonomy delineates two fundamental regimes: Overthinking describes trajectories where the model produces unnecessarily verbose, redundant, or prolonged reasoning and tool invocations while maintaining linguistic fluency, local semantic progression, and probability distributions comparable to benign tasks (e.g., Engorgio, OverThink, and ClawDrain); conversely, Unbounded Drift characterizes degenerate, self-sustaining trajectories where the model enters repetitive loops, cyclic reasoning, or recursive agent calls with negligible semantic progression, exhibiting collapsed token entropy, elevated token confidence, and severely suppressed End-of-Sequence (EOS) probabilities (e.g., LoopLLM, RECUR, and CORBA). This operational distinction directly dictates the required defensive posture.
2. Underlying Generation Dynamics: Unveiling the microscopic mechanisms sustaining excessive compute
To explain why models fail to terminate naturally once an attack trajectory is initiated, the survey synthesizes empirical findings across multiple analytical layers. At the representation level, self-attention induction heads and repetition neurons, which normally facilitate in-context copying and associative recall, become over-activated by adversarial patterns and drive attention away from informative continuations toward repetitive fragments. At the probabilistic level, autoregressive decoding exhibits a "winner-takes-all" positive feedback loop: once an adversarial prefix forces the output into a low-entropy state space, the conditional probability of repeating previous n-grams escalates non-linearly, trapping the trajectory in an attractor state. At the training level, standard likelihood maximization and reinforcement learning objectives that reward elaborate chain-of-thought verification inadvertently penalize early stopping, conditioning the model to mistake endless reflection for task reward.
3. Active Multi-Tier Defenses: Intercepting compute amplification across model and system stacks
Recognizing that standard inference acceleration (e.g., static token truncation or speculative skipping) cannot distinguish adversarial exploitation from legitimate complex reasoning, the survey categorizes active defenses by their alignment with the two outcome regimes. Against Overthinking, defenses implement utility-aware budget control and reflection management: frameworks such as CGRS and MoT monitor reasoning traces during decoding, identifying points where information gain saturates and proactively suppressing redundant self-affirmation or forcing early termination. Against Unbounded Drift, defenses focus on state recurrence detection and structural interruption: tools like RecurrentDetector track activation recurrence in Transformer hidden states to detect and halt decoding loops with sub-millisecond overhead, while WSC segments and excises repetitive reasoning chunks. Complementing these, system-level frameworks such as PD3F incorporate dynamic request polling and EOS probability boosting to safeguard multi-tenant queues against concurrent DoS pressure.
Benchmark & Evaluation Protocol¶
To overcome the pervasive limitation where prior studies report only visible token lengths, the authors establish a standardized three-tiered benchmarking framework: 1. Generation Level: Assesses behavioral trajectory anomalies, capturing total generated tokens (Output Length), end-to-end request latency (Time Cost), and runtime verification overhead introduced by defensive modules (Defense-time Cost). 2. Hardware Level: Measures physical infrastructure stress, focusing on GPU occupancy (GPU Utilization) and peak tensor allocation (GPU Memory Usage) to detect memory fragmentation, cache thrashing, and compute saturation. 3. Service Level: Evaluates real-world production impact, incorporating system concurrency (Throughput / Completed Requests per Unit Time), initial token response delay (Time-to-First-Token, TTFT), and downstream task accuracy preservation (Task Performance) to verify that defenses do not degrade benign task utility.
Key Experimental Results¶
Main Results: Generation Expansion Across Model Paradigms¶
The following table synthesizes quantitative trajectory expansion data compiled across representative studies (from Table 2 and thematic survey tables), showing average generation length and peak maximum length in braces across standard LLMs, reasoning LLMs (RLLMs), multimodal LLMs (MLLMs), and agentic pipelines:
| Setting & Max Budget | Outcome Regime | Representative Attack | Output Length (Avg {Max}) | Observed Amplification / Disruption |
|---|---|---|---|---|
| Text LLM (8K) | Overthinking | Engorgio (Dong et al., 2024) | 1.4K | Maintained natural syntax; elevated latency |
| Text LLM (8K) | Unbounded Drift | LoopLLM (Li et al., 2025a) | 2.1K | Low-entropy token loops; saturated token window |
| Reasoning LLM (64K) | Overthinking | OverThink (Kumar et al., 2025) | 4.9K | Redundant auxiliary reasoning; 46ร compute cost |
| Reasoning LLM (64K) | Unbounded Drift | RECUR (Wang et al., 2026b) | 5.3K | Self-reinforcing reflection; maxed 64K context |
| Multimodal LLM (32K) | Overthinking | Verbose Images (Gao et al., 2024a) | 247.4 | Visual perturbations; 8ร energy-latency increase |
| Multimodal LLM (32K) | Unbounded Drift | LingoLoop (Fu et al., 2025) | 11.5K | Visually trapped repetitive linguistic cycles |
| Agentic System (Multi-round) | Overthinking | Beyond Max Tokens (Zhou et al., 2026b) | 70.9K | Stealthy tool invocation expansion in trace |
| Agentic System (Multi-round) | Unbounded Drift | Overthinking Loops (Lee et al., 2026) | 9.0K | Poisoned tool returns; cyclic agent lock-in |
Ablation & Defense Comparison: Operational Trade-offs of Active Mitigations¶
The table below compares representative active defense mechanisms reviewed in the survey (compiled from Table 3, Table 14, and Table 15), detailing target regimes, underlying mechanisms, quantitative impact, and operational limitations:
| Defense Framework | Target Regime | Underlying Mechanism | Quantitative Efficacy | Key Operational Limitation |
|---|---|---|---|---|
| PD3F (Zhang et al., 2025d) | Overthinking & Drift | Resource consumption indexing + dynamic request polling and EOS amplification | Total service time reduced by 50% โ; halts queue DoS | Generates false positives on compute-heavy benign tasks |
| CGRS (Huang et al., 2025b) | Overthinking | Internal confidence and reflection cue tracking to suppress redundant checking | Suppresses uninformative reflection; preserves accuracy | Narrow pattern coverage; dependent on trigger phrases |
| MoT (Zhu et al., 2025b) | Overthinking | Periodic reasoning trajectory monitoring and automated early stopping | Attack success rate > 90%; unnecessary steps cut ~80% โ | Requires transparent or controllable reasoning traces |
| WSC (Xie et al., 2025) | Overthinking & Drift | Hidden state tracking at chunk boundaries to excise repetitive decoding blocks | Reasoning trace compressed by 13%โ57% | Model-specific classifier; regeneration may loop |
| RecurrentDetector (Yu et al., 2025) | Unbounded Drift | Online activation recurrence monitoring to detect state looping | Halts loops within 272 attempts vs 1679 baseline | Minimal 0.36 ms overhead; requires white-box state access |
Key Findings¶
- Unbounded Drift inflicts devastating resource monopolization: In terms of absolute sequence length, drift attacks push generation to the absolute limits of context windows (e.g., RECUR reaching 64.4K tokens on a 64K model), but their reliance on specific decoding dynamics makes them less transferable than syntax-preserving Overthinking attacks.
- Agent execution traces conceal massive resource overhead: In multi-agent and tool-use ecosystems, the visible final response often appears concise, yet intermediate tool calling and message exchange can covertly consume over 90K tokens (Beyond Max Tokens), bypassing naive output-length monitors.
- Hardware-level metric reporting is critically absent: The survey highlights an evaluation blind spot where less than half of existing security studies report GPU memory occupancy, hardware utilization, or cluster-level serving throughput, frequently conflating visible token truncation with genuine infrastructure protection.
Highlights & Insights¶
- Reframing resource management from efficiency to security: The survey delivers an essential mindset shift, demonstrating that long-form generation under adversarial manipulation cannot be treated merely as a latency-cost trade-off, but constitutes an infrastructure availability vulnerability.
- Quantitative decoding diagnostic signals: By analyzing top-1 logits, EOS ranking, token-level entropy, and output confidence across decoding steps, the authors provide concrete, model-agnostic criteria to distinguish natural contemplation from malicious lock-in.
- Holistic cross-modal threat taxonomy: The work successfully bridges isolated security findings from vision models, reasoning engines, and autonomous agents into a cohesive threat model.
Limitations & Future Work¶
- Impracticability of white-box defenses for closed commercial APIs: The authors acknowledge that the most effective loop-detection and chunk-interruption defenses (such as RecurrentDetector and MoT) mandate access to intermediate hidden states or logit distributions, which commercial proprietary API providers do not expose.
- Under-explored dynamics in multimodal and physical agent regimes: Current mechanistic explanations remain heavily grounded in textual autoregression; the dynamics of non-convergent generation in long-form video comprehension and embodied robotics remain largely uncharacterized.
- Absence of unified red-teaming benchmark suites: The literature currently relies on fragmented, small-scale adversarial prompts without standardized testbeds that integrate physical server load metrics.
Related Work & Insights¶
- vs. Precursor Sponge Attacks (Sponge Examples / NMTSloth): Early sponge attacks focused on breaking energy sparsity in static CNNs or forcing translation models to delay EOS tokens via gradient-based pixel or token perturbations. In contrast, modern LLM resource threats exploit deep semantic reasoning chains, agentic tool workflows, and multi-modal contexts, operating at a far higher cognitive abstraction level.
- vs. Standard LLM Jailbreaking and Refusal Attacks: Conventional jailbreaking techniques aim to induce policy-violating harmful text, whereas resource consumption attacks preserve apparent task utility or benign compliance while deliberately maximizing asymmetric compute expenditure to exhaust serving infrastructure.
Rating¶
- Novelty: โญโญโญโญโญ Establishes the first unified taxonomy and operational boundary for LLM resource consumption threats.
- Experimental Thoroughness: โญโญโญโญโญ Analyzes and synthesizes empirical results from over 120 foundational and contemporary publications across all major architectures.
- Writing Quality: โญโญโญโญโญ Exceptionally structured narrative with precise mechanistic formulations and rigorous technical clarity.
- Value: โญโญโญโญโญ Serves as an indispensable blueprint for building denial-of-service resilience and sustainable serving infrastructure in production LLM systems.