arXiv subjects

Kyomin Jung

Publications and source records attributed to Kyomin Jung.

At least 37 records · Page 2Linked to original sources

arXiv2026

CausalArmor: Efficient Indirect Prompt Injection Guardrails via Causal Attribution

AI agents equipped with tool-calling capabilities are susceptible to Indirect Prompt Injection (IPI) attacks. In this attack scenario, malicious commands hidden within untrusted content trick the agent into performing unauthorized actions. Existing defenses can reduce attack success but often suffer from the over-defense dilemma: they deploy expensive, always-on sanitization regardless of actual threat, thereby degrading utility and latency even in benign scenarios. We revisit IPI through a causal ablation perspective: a successful injection manifests as a dominance shift where the user request no longer provides decisive support for the agent's privileged action, while a particular untrusted segment, such as a retrieved document or tool output, provides disproportionate attributable influence. Based on this signature, we propose CausalArmor, a selective defense framework that (i) computes lightweight, leave-one-out ablation-based attributions at privileged decision points, and (ii) triggers targeted sanitization only when an untrusted segment dominates the user intent. Additionally, CausalArmor employs retroactive Chain-of-Thought masking to prevent the agent from acting on ``poisoned'' reasoning traces. We present a theoretical analysis showing that sanitization based on attribution margins conditionally yields an exponentially small upper bound on the probability of selecting malicious actions. Experiments on AgentDojo and DoomArena demonstrate that CausalArmor matches the security of aggressive defenses while improving explainability and preserving utility and latency of AI agents.

cs.CR↗

arXiv2026

Persona Switch: Mixing Distinct Perspectives in Decoding Time

Role-play prompting is known to steer the behavior of language models by injecting a persona into the prompt, improving their zero-shot reasoning capabilities. However, such improvements are inconsistent across different tasks or instances. This inconsistency suggests that zero-shot and role-play prompting may offer complementary strengths rather than one being universally superior. Building on this insight, we propose Persona Switch, a novel decoding method that dynamically combines the benefits of both prompting strategies. Our method proceeds step-by-step, selecting the better output between zero-shot and role-play prompting at each step by comparing their output confidence, as measured by the logit gap. Experiments with widely-used LLMs demonstrate that Persona Switch consistently outperforms competitive baselines, achieving up to 5.13% accuracy improvement. Furthermore, we show that output confidence serves as an informative measure for selecting the more reliable output.

cs.CL↗

arXiv2026

When Wording Steers the Evaluation: Framing Bias in LLM judges

Large language models (LLMs) are known to produce varying responses depending on prompt phrasing, indicating that subtle guidance in phrasing can steer their answers. However, the impact of this framing bias on LLM-based evaluation, where models are expected to make stable and impartial judgments, remains largely underexplored. Drawing inspiration from the framing effect in psychology, we systematically investigate how deliberate prompt framing skews model judgments across four high-stakes evaluation tasks. We design symmetric prompts using predicate-positive and predicate-negative constructions and demonstrate that such framing induces significant discrepancies in model outputs. Across 14 LLM judges, we observe clear susceptibility to framing, with model families showing distinct tendencies toward agreement or rejection. These findings suggest that framing bias is a structural property of current LLM-based evaluation systems, underscoring the need for framing-aware protocols.

cs.CL↗

arXiv2025

Confidence-Guided Stepwise Model Routing for Cost-Efficient Reasoning

Recent advances in Large Language Models (LLMs) - particularly model scaling and test-time techniques - have greatly enhanced the reasoning capabilities of language models at the expense of higher inference costs. To lower inference costs, prior works train router models or deferral mechanisms that allocate easy queries to a small, efficient model, while forwarding harder queries to larger, more expensive models. However, these trained router models often lack robustness under domain shifts and require expensive data synthesis techniques such as Monte Carlo rollouts to obtain sufficient ground-truth routing labels for training. In this work, we propose Confidence-Guided Stepwise Model Routing for Cost-Efficient Reasoning (STEER), a domain-agnostic framework that performs fine-grained, step-level routing between smaller and larger LLMs without utilizing external models. STEER leverages confidence scores from the smaller model's logits prior to generating a reasoning step, so that the large model is invoked only when necessary. Extensive evaluations using different LLMs on a diverse set of challenging benchmarks across multiple domains such as Mathematical Reasoning, Multi-Hop QA, and Planning tasks indicate that STEER achieves competitive or enhanced accuracy while reducing inference costs (up to +20% accuracy with 48% less FLOPs compared to solely using the larger model on AIME), outperforming baselines that rely on trained external modules. Our results establish model-internal confidence as a robust, domain-agnostic signal for model routing, offering a scalable pathway for efficient LLM deployment.

cs.CL↗

arXiv2025

Fooling the LVLM Judges: Visual Biases in LVLM-Based Evaluation

Recently, large vision-language models (LVLMs) have emerged as the preferred tools for judging text-image alignment, yet their robustness along the visual modality remains underexplored. This work is the first study to address a key research question: Can adversarial visual manipulations systematically fool LVLM judges into assigning unfairly inflated scores? We define potential image induced biases within the context of T2I evaluation and examine how these biases affect the evaluations of LVLM judges. Moreover, we introduce a novel, fine-grained, multi-domain meta-evaluation benchmark named FRAME, which is deliberately constructed to exhibit diverse score distributions. By introducing the defined biases into the benchmark, we reveal that all tested LVLM judges exhibit vulnerability across all domains, consistently inflating scores for manipulated images. Further analysis reveals that combining multiple biases amplifies their effects, and pairwise evaluations are similarly susceptible. Moreover, we observe that visual biases persist under prompt-based mitigation strategies, highlighting the vulnerability of current LVLM evaluation systems and underscoring the urgent need for more robust LVLM judges.

cs.CL↗

arXiv2025

FaithUn: Toward Faithful Forgetting in Language Models by Investigating the Interconnectedness of Knowledge

Various studies have attempted to remove sensitive or private knowledge from a language model to prevent its unauthorized exposure. However, prior studies have overlooked the complex and interconnected nature of knowledge, where related knowledge must be carefully examined. Specifically, they have failed to evaluate whether an unlearning method faithfully erases interconnected knowledge that should be removed, retaining knowledge that appears relevant but exists in a completely different context. To resolve this problem, we first define a new concept called superficial unlearning, which refers to the phenomenon where an unlearning method either fails to erase the interconnected knowledge it should remove or unintentionally erases irrelevant knowledge. Based on the definition, we introduce a new benchmark, FaithUn, to analyze and evaluate the faithfulness of unlearning in real-world knowledge QA settings. Furthermore, we propose a novel unlearning method, KLUE, which updates only knowledge-related neurons to achieve faithful unlearning. KLUE identifies knowledge neurons using an explainability method and updates only those neurons using selected unforgotten samples. Experimental results demonstrate that widely-used unlearning methods fail to ensure faithful unlearning, while our method shows significant effectiveness in real-world QA unlearning.

cs.CL↗

arXiv2025

Program Synthesis via Test-Time Transduction

We introduce transductive program synthesis, a new formulation of the program synthesis task that explicitly leverages test inputs during synthesis. While prior approaches to program synthesis--whether based on natural language descriptions or input-output examples--typically aim to generalize from training examples, they often struggle with robustness, especially in real-world settings where training examples are limited and test inputs involve various edge cases. To address this, we propose a novel framework that improves robustness by treating synthesis as an active learning over a finite hypothesis class defined by programs' outputs. We use an LLM to predict outputs for selected test inputs and eliminate inconsistent hypotheses, where the inputs are chosen via a greedy maximin algorithm to minimize the number of LLM queries required. We evaluate our approach on four benchmarks: Playgol, MBPP+, 1D-ARC, and programmatic world modeling on MiniGrid. We demonstrate that our method significantly improves program synthesis in both accuracy and efficiency. We release our code at https://github.com/klee972/SYNTRA.

cs.AI↗

arXiv2025

ReflAct: World-Grounded Decision Making in LLM Agents via Goal-State Reflection

Recent advances in LLM agents have largely built on reasoning backbones like ReAct, which interleave thought and action in complex environments. However, ReAct often produces ungrounded or incoherent reasoning steps, leading to misalignment between the agent's actual state and goal. Our analysis finds that this stems from ReAct's inability to maintain consistent internal beliefs and goal alignment, causing compounding errors and hallucinations. To address this, we introduce ReflAct, a novel backbone that shifts reasoning from merely planning next actions to continuously reflecting on the agent's state relative to its goal. By explicitly grounding decisions in states and enforcing ongoing goal alignment, ReflAct dramatically improves strategic reliability. This design delivers substantial empirical gains: ReflAct surpasses ReAct by 27.7% on average, achieving a 93.3% success rate in ALFWorld. Notably, ReflAct even outperforms ReAct with added enhancement modules (e.g., Reflexion, WKM), showing that strengthening the core reasoning backbone is key to reliable agent performance.

cs.CL↗

arXiv2025

Black-Box Hallucination Detection via Consistency Under the Uncertain Expression

Despite the great advancement of Language modeling in recent days, Large Language Models (LLMs) such as GPT3 are notorious for generating non-factual responses, so-called "hallucination" problems. Existing methods for detecting and alleviating this hallucination problem require external resources or the internal state of LLMs, such as the output probability of each token. Given the LLM's restricted external API availability and the limited scope of external resources, there is an urgent demand to establish the Black-Box approach as the cornerstone for effective hallucination detection. In this work, we propose a simple black-box hallucination detection metric after the investigation of the behavior of LLMs under expression of uncertainty. Our comprehensive analysis reveals that LLMs generate consistent responses when they present factual responses while non-consistent responses vice versa. Based on the analysis, we propose an efficient black-box hallucination detection metric with the expression of uncertainty. The experiment demonstrates that our metric is more predictive of the factuality in model responses than baselines that use internal knowledge of LLMs.

cs.CL↗

arXiv2025

Harmful Prompt Laundering: Jailbreaking LLMs with Abductive Styles and Symbolic Encoding

Large Language Models (LLMs) have demonstrated remarkable capabilities across diverse tasks, but their potential misuse for harmful purposes remains a significant concern. To strengthen defenses against such vulnerabilities, it is essential to investigate universal jailbreak attacks that exploit intrinsic weaknesses in the architecture and learning paradigms of LLMs. In response, we propose \textbf{H}armful \textbf{P}rompt \textbf{La}undering (HaPLa), a novel and broadly applicable jailbreaking technique that requires only black-box access to target models. HaPLa incorporates two primary strategies: 1) \textit{abductive framing}, which instructs LLMs to infer plausible intermediate steps toward harmful activities, rather than directly responding to explicit harmful queries; and 2) \textit{symbolic encoding}, a lightweight and flexible approach designed to obfuscate harmful content, given that current LLMs remain sensitive primarily to explicit harmful keywords. Experimental results show that HaPLa achieves over 95% attack success rate on GPT-series models and 70% across all targets. Further analysis with diverse symbolic encoding rules also reveals a fundamental challenge: it remains difficult to safely tune LLMs without significantly diminishing their helpfulness in responding to benign queries.

cs.AI↗

arXiv2025

Public Data Assisted Differentially Private In-Context Learning

In-context learning (ICL) in Large Language Models (LLMs) has shown remarkable performance across various tasks without requiring fine-tuning. However, recent studies have highlighted the risk of private data leakage through the prompt in ICL, especially when LLMs are exposed to malicious attacks. While differential privacy (DP) provides strong privacy guarantees, it often significantly reduces the utility of in-context learning (ICL). To address this challenge, we incorporate task-related public data into the ICL framework while maintaining the DP guarantee. Based on this approach, we propose a private in-context learning algorithm that effectively balances privacy protection and model utility. Through experiments, we demonstrate that our approach significantly improves the utility of private ICL with the assistance of public data. Additionally, we show that our method is robust against membership inference attacks, demonstrating empirical privacy protection.

cs.AI↗

arXiv2025

Avoidance Decoding for Diverse Multi-Branch Story Generation

Large Language Models (LLMs) often generate repetitive and monotonous outputs, especially in tasks like story generation, due to limited creative diversity when given the same input prompt. To address this challenge, we propose a novel decoding strategy, Avoidance Decoding, that modifies token logits by penalizing similarity to previously generated outputs, thereby encouraging more diverse multi-branch stories. This penalty adaptively balances two similarity measures: (1) Concept-level Similarity Penalty, which is prioritized in early stages to diversify initial story concepts, and (2) Narrative-level Similarity Penalty, which is increasingly emphasized later to ensure natural yet diverse plot development. Notably, our method achieves up to 2.6 times higher output diversity and reduces repetition by an average of 30% compared to strong baselines, while effectively mitigating text degeneration. Furthermore, we reveal that our method activates a broader range of neurons, demonstrating that it leverages the model's intrinsic creativity.

cs.CL↗

arXiv2025

Unplug and Play Language Models: Decomposing Experts in Language Models at Inference Time

Enabled by large-scale text corpora with huge parameters, pre-trained language models operate as multi-task experts using a single model architecture. However, recent studies have revealed that certain neurons play disproportionately important roles in solving specific tasks, suggesting that task-relevant substructures can be isolated and selectively activated for each task. Therefore, we introduce Decomposition of Experts (DoE), a novel framework that dynamically identifies and activates task-specific experts within a language model to reduce inference cost without sacrificing accuracy. We first define a task expert as a set of parameters that significantly influence the performance of a specific task and propose a four-step unplug-and-play process: (1) receiving a user request, (2) identifying the corresponding task expert, (3) performing inference using the expert-localized model, and (4) restoring the original model and waiting for the next task. Using attribution methods and prompt tuning, DoE isolates task-relevant neurons, minimizing computational overhead while maintaining task performance. We assume a setting where a language model receives user requests from five widely used natural language understanding benchmarks, processing one task at a time. In this setup, we demonstrate that DoE achieves up to a x1.73 inference speed-up with a 65% pruning rate, without compromising accuracy. Comparisons with various task expert localization methods reveal that DoE effectively identifies task experts, while ablation studies validate the importance of its components. Additionally, we analyze the effects of batch size, token count, and layer types on inference speed-up, providing practical insights for adopting DoE. The proposed framework is both practical and scalable, applicable to any transformer-based architecture, offering a robust solution for efficient task-specific inference.

cs.CL↗

arXiv2025

Can You Trick the Grader? Adversarial Persuasion of LLM Judges

As large language models take on growing roles as automated evaluators in practical settings, a critical question arises: Can individuals persuade an LLM judge to assign unfairly high scores? This study is the first to reveal that strategically embedded persuasive language can bias LLM judges when scoring mathematical reasoning tasks, where correctness should be independent of stylistic variation. Grounded in Aristotle's rhetorical principles, we formalize seven persuasion techniques (Majority, Consistency, Flattery, Reciprocity, Pity, Authority, Identity) and embed them into otherwise identical responses. Across six math benchmarks, we find that persuasive language leads LLM judges to assign inflated scores to incorrect solutions, by up to 8% on average, with Consistency causing the most severe distortion. Notably, increasing model size does not substantially mitigate this vulnerability. Further analysis demonstrates that combining multiple persuasion techniques amplifies the bias, and pairwise evaluation is likewise susceptible. Moreover, the persuasive effect persists under counter prompting strategies, highlighting a critical vulnerability in LLM-as-a-Judge pipelines and underscoring the need for robust defenses against persuasion-based attacks.

cs.CL↗

arXiv2025

Guaranteed Generation from Large Language Models

As large language models (LLMs) are increasingly used across various applications, there is a growing need to control text generation to satisfy specific constraints or requirements. This raises a crucial question: Is it possible to guarantee strict constraint satisfaction in generated outputs while preserving the distribution of the original model as much as possible? We first define the ideal distribution - the one closest to the original model, which also always satisfies the expressed constraint - as the ultimate goal of guaranteed generation. We then state a fundamental limitation, namely that it is impossible to reach that goal through autoregressive training alone. This motivates the necessity of combining training-time and inference-time methods to enforce such guarantees. Based on this insight, we propose GUARD, a simple yet effective approach that combines an autoregressive proposal distribution with rejection sampling. Through GUARD's theoretical properties, we show how controlling the KL divergence between a specific proposal and the target ideal distribution simultaneously optimizes inference speed and distributional closeness. To validate these theoretical concepts, we conduct extensive experiments on two text generation settings with hard-to-satisfy constraints: a lexical constraint scenario and a sentiment reversal scenario. These experiments show that GUARD achieves perfect constraint satisfaction while almost preserving the ideal distribution with highly improved inference efficiency. GUARD provides a principled approach to enforcing strict guarantees for LLMs without compromising their generative capabilities.

cs.CL↗

arXiv2025

MultiActor-Audiobook: Zero-Shot Audiobook Generation with Faces and Voices of Multiple Speakers

We introduce MultiActor-Audiobook, a zero-shot approach for generating audiobooks that automatically produces consistent, expressive, and speaker-appropriate prosody, including intonation and emotion. Previous audiobook systems have several limitations: they require users to manually configure the speaker's prosody, read each sentence with a monotonic tone compared to voice actors, or rely on costly training. However, our MultiActor-Audiobook addresses these issues by introducing two novel processes: (1) MSP (**Multimodal Speaker Persona Generation**) and (2) LSI (**LLM-based Script Instruction Generation**). With these two processes, MultiActor-Audiobook can generate more emotionally expressive audiobooks with a consistent speaker prosody without additional training. We compare our system with commercial products, through human and MLLM evaluations, achieving competitive results. Furthermore, we demonstrate the effectiveness of MSP and LSI through ablation studies.

cs.SD↗

arXiv2025

Drift: Decoding-time Personalized Alignments with Implicit User Preferences

Personalized alignments for individual users have been a long-standing goal in large language models (LLMs). We introduce Drift, a novel framework that personalizes LLMs at decoding time with implicit user preferences. Traditional Reinforcement Learning from Human Feedback (RLHF) requires thousands of annotated examples and expensive gradient updates. In contrast, Drift personalizes LLMs in a training-free manner, using only a few dozen examples to steer a frozen model through efficient preference modeling. Our approach models user preferences as a composition of predefined, interpretable attributes and aligns them at decoding time to enable personalized generation. Experiments on both a synthetic persona dataset (Perspective) and a real human-annotated dataset (PRISM) demonstrate that Drift significantly outperforms RLHF baselines while using only 50-100 examples. Our results and analysis show that Drift is both computationally efficient and interpretable.

cs.CL↗

arXiv2025

Are LLM-Judges Robust to Expressions of Uncertainty? Investigating the effect of Epistemic Markers on LLM-based Evaluation

In line with the principle of honesty, there has been a growing effort to train large language models (LLMs) to generate outputs containing epistemic markers. However, evaluation in the presence of epistemic markers has been largely overlooked, raising a critical question: Could the use of epistemic markers in LLM-generated outputs lead to unintended negative consequences? To address this, we present EMBER, a benchmark designed to assess the robustness of LLM-judges to epistemic markers in both single and pairwise evaluation settings. Our findings, based on evaluations using EMBER, reveal that all tested LLM-judges, including GPT-4o, show a notable lack of robustness in the presence of epistemic markers. Specifically, we observe a negative bias toward epistemic markers, with a stronger bias against markers expressing uncertainty. This suggests that LLM-judges are influenced by the presence of these markers and do not focus solely on the correctness of the content.

cs.CL↗