acceptodds
Withdrawn from ICLR 2027

Asking Forever: Turn Amplification via Clarification-Seeking Directions in LLMs

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Multi-turn interaction length is a dominant factor in the operational costs of conversational LLMs. In this work, we present a new failure mode: *turn amplification*, in which a model consistently prolongs multi-turn interactions without completing the underlying task. We show that an adversary can exploit clarification-seeking behavior—commonly encouraged in multi-turn conversation settings—to consistently prolong interactions. Moving beyond prompt-level behaviors, we take a mechanistic perspective and identify query-independent (i.e., *universal* across prompts) activation directions associated with clarification-seeking responses. Unlike prior attacks that rely on per-turn prompt optimization, exploiting these directions requires no such optimization, and their effect persists across tasks. To assess whether it constitutes an exploitable vulnerability, we instantiate two well-established threat models—supply-chain attacks via fine-tuning and runtime attacks through hardware-level fault injection—and show that both reliably induce turn amplification across prompts and tasks. Across multiple LLMs and benchmarks, our attack substantially increases turn count while keeping individual responses benign. We evaluate potential countermeasures and show they offer limited protection.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
large language models, cost amplification, activation steering

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_kYxsOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies turn amplification, an attack that increases interaction length and inference costs through persistent clarification-seeking and delayed task completion. The authors induce this behavior through activation steering, LoRA fine-tuning, and targeted bit flips, and develop an interaction-level auditing framework to evaluate attack effectiveness and potential countermeasures.

Critical strengths
  1. Novel perspective with practical relevance. Under usage-based billing, LLM service providers may unnecessarily prolong user conversations to increase revenue, which is a plausible real-world risk. The paper extends the study of cost-amplification attacks from single-turn generation to entire interactions and systematically investigates this issue.

  2. Systematic empirical evaluation. The experiments cover multiple models, datasets, and two user engagement modes. Together with multiple attack implementations and ablation studies, they provide solid evidence for the feasibility of turn amplification.

Critical weaknesses
  1. Limited evidence for capability preservation. Accuracy is computed only over valid responses, while simulated users sometimes provide intermediate reasoning or even final answers. These factors limit the support for the claim that the target model’s capabilities are preserved.

  2. Some counterintuitive results remain insufficiently explained. For example, the Defensive System Message strengthens the turn amplification induced by Prefilling, but this effect is not adequately discussed.

Questions for authors
  1. Could the authors report the overall correct-completion rate across all questions and evaluate a user simulator that does not provide solution steps or final answers, to further validate capability preservation?

  2. For Llama3-8B on Alpaca under Mode 1, why does the Defensive System Message increase the average number of turns under Prefilling from 3.3 to 6.4 (Tables 1 and 3)? Is this effect consistent across repeated runs, and what might explain it?

Reviewer_A6MgOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces turn amplification, where a compromised LLM keeps asking unnecessary clarification questions instead of finishing the user’s task, increasing token costs. It shows that this behavior can be induced through activation steering, LoRA FT, or targeted bit flips to model weights. Across several models and benchmarks, these attacks can greatly increase conversation length and input-token usage while keeping individual responses seemingly normal. The paper also tests simple defenses, but finds that they do not reliably prevent the attack.

Critical strengths
  • This paper identifies an interesting interaction-level failure mode where models can amplify cost by unnecessarily prolonging conversations.
  • Relatively comprehensive evaluates the phenomenon across multiple model families and attack mechanisms
Critical weaknesses
  • The overall contribution feels limited beyond reframing unnecessary clarification-seeking as a cost-amplification attack.
  • All three attacks need white-box access. With that access an attacker can cause almost any behavior, so getting the model to ask clarifying questions is unsurprising. Given this level of access, it is unclear why an attacker would specifically choose to induce prolonged clarification-seeking rather than use simpler and potentially more effective ways to waste computation or degrade service. More broadly, the attacker’s practical motivation is not sufficiently convincing, making the real-world significance of the proposed attack difficult to assess.
  • Learning steering vectors with gradients, LoRA fine-tuning on synthetic data, and gradient-guided bit search all exist already. Hence, there's limited technical novelty.
  • There's limited validation of the evaluation, where all results come from simulated conversations and there is no study with real users. So it is unclear whether the reported turn and token increases would carry over to real deployments, where users would likely quit sooner or behaving in other ways.
Questions for authors
  • How realistic is the assumed attacker, given that all attack scenarios require white-box access to model parameters or internal activations?
  • How robust are the results to different user simulators, judge models, and more realistic conversational tasks?
  • Are there any evidence that the attack remains difficult to detect once monitoring operates at the conversation level rather than at individual turns?
Reviewer_zeFuOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper introduces turn amplification: an attack that raises the inference cost of conversational LLMs by making the model defer task completion across many turns, mainly through clarification-seeking, while each individual response looks benign. The authors learn per-layer steering vectors by gradient descent on synthetic stalling conversations. They also implement the attack through LoRA fine-tuning and through 25 targeted bit flips. To measure it, they build a framework that turns single-turn benchmarks (Alpaca, GSM8K) into multi-turn ones, with an LLM judge and two simulated user modes (cooperative, and impatient/abandoning). They report up to 9.8× more turns and 121.9× more input tokens across four models, and find that prompt-based defenses help little.

Critical strengths
  1. Provides a reusable multi-turn auditing framework that converts single-turn benchmarks, with two user modes (cooperative and abandoning) and a human-validated judge (95% agreement, κ=0.9).
  2. The work is transparent: an adaptive attack against the authors’ own heuristic detector, full prompts, verbatim example conversations, and released code.
Critical weaknesses
  1. Limited novelty relative to prior cost attacks. Prior cost attacks already use the same threat models and vectors: data poisoning in P-DoS and bit flips in BitHydra (both cited at l. 106–107). The methods are off-the-shelf; only the target metric (turn count) is new.
  2. No comparison with prior cost attacks, and an inflated cost metric. The paper compares only against prefilling and system-prompt baselines, not against existing cost attacks. Most of the reported cost comes from re-sent conversation history, which prompt caching and context truncation already reduce, so the real monetary amplification is likely far below the headline ratios.
  3. Stealth claims are unsupported, and the evaluation favors the attack. The LoRA model rarely answers, simulated users often abandon, and no monitor is evaluated. Results rely on a cooperative simulated user and fully specified tasks where clarification is never legitimate.
Questions for authors
  1. How does total monetary cost compare with P-DoS, BitHydra, or Engorgio at a matched attacker budget, under realistic pricing with cached-input rates and context truncation?
  2. How well does a simple monitor, such as your own completion judge or a per-conversation turn threshold, detect the attacks? What is its false-positive rate on benign multi-turn traffic where clarification is legitimate?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.