Revisiting Early Detection of Sexual Predators via Turn-level Optimization (2025.naacl-long)
Copied to clipboard
| Challenge: | Existing methods to detect online grooming rely on chat-level risk labels and fail to identify optimal intervention points. |
| Approach: | They propose a speed control reinforcement learning strategy based on luring communication theory to capture the predator’s turn-level entrapment and a new reward function that balances the trade-off between speed and accuracy based upon the LCT. |
| Outcome: | The proposed method preempts online grooming while identifying optimal early intervention points. |
Similar Papers
TROJail: Trajectory-Level Optimization for Multi-Turn Large Language Model Jailbreaks with Process Rewards (2026.acl-long)
Copied to clipboard
| Challenge: | Existing approaches to training multi-turn attackers to probe model safety vulnerabilities rely on turn-level optimization, which is insufficient for learning long-term attack strategies. |
| Approach: | They propose a multi-turn reinforcement learning problem that optimizes the harmfulness of the final-turn response as the outcome reward. |
| Outcome: | The proposed approach improves attack success rates across multiple models and benchmarks, highlighting the effectiveness of the proposed approach. |
Learning to Conceal Risk: Controllable Multi-turn Red Teaming for LLMs in the Financial Domain (2026.acl-long)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) are increasingly deployed in high-stakes domains such as finance where unsafe behavior can lead to serious regulatory risks. |
| Approach: | They propose a black-box multi-turn risk-concealed redteaming framework that progressively conceals surface-level risk while exploiting regulatory-violating behaviors. |
| Outcome: | Experiments on nine widely used LLMs show that the proposed framework achieves 93.19% average attack success rate (ASR) and improves the average ASR to 95.00%. |
Prior Prompt Engineering for Reinforcement Fine-Tuning (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing studies have focused on algorithms, reward shaping, and data curation, but prior prompt engineering is understudied. |
| Approach: | They investigate prior prompt engineering (pPE) in reinforcement fine-tuning . they translate five representative iPE strategies into corresponding pPE approaches . |
| Outcome: | The proposed approaches outperform iPE-prompted models on in-domain and out-of-domain benchmarks. |
Turn-PPO: Turn-Level Advantage Estimation with PPO for Improved Multi-Turn RL in Agentic LLMs (2026.findings-eacl)
Copied to clipboard
| Challenge: | Reinforcement learning (RL) has re-emerged as a natural approach for training interactive LLM agents in real-world environments. |
| Approach: | They propose a variant that operates on a turn-level MDP formulation, instead of the commonly used token-level one. |
| Outcome: | The proposed method is more robust than the widely used GRPO algorithm and more efficient than token-level MDPs. |
Don’t Click That: Teaching Web Agents to Resist Deceptive Interfaces (2026.acl-long)
Copied to clipboard
| Challenge: | Existing approaches to deception detection and defenses are inadequate . Existing methods do not integrate with agent decision-making . |
| Approach: | They propose a framework that integrates hybrid-reward learning with asymmetric penalties and experience summarization to distill failure patterns into transferable guidance. |
| Outcome: | The proposed framework reduces deception susceptibility by 53.8% while maintaining task performance, establishing an effective foundation for robust web agent deployment. |
Token-level Inference-Time Alignment for Vision-Language Models (2026.findings-acl)
Copied to clipboard
Kejia Chen, Junjun Zheng, Jiawen Zhang, Manxi Lin, Xiao Pan, Jiacong Hu, Jian Lou, Zunlei Feng, Mingli Song
| Challenge: | Vision-Language Models (VLMs) often prioritize linguistic fluency over visual fidelity . despite widespread adoption, VLMs often exhibit a critical failure mode: hallucination . |
| Approach: | They propose a framework for Token-level Inference-Time Alignment that steers the decoding process without updating the base model parameters. |
| Outcome: | The proposed framework improves performance on 13 benchmarks across architectures . it boosts LLaVA-1.5-7B by 8.6% on MMVet and achieves a 74.0 MMStar score . |
Proactive Guidance of Multi-Turn Conversation in Industrial Search (2025.acl-industry)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) have advanced multi-turn conversation systems, emphasizing the need for proactive guidance to enhance users’ interactions. |
| Approach: | They propose a goal-adaptive supervised fine-tuning framework that generates proactive guidance for users to click for the next turn of the conversation. |
| Outcome: | The proposed framework achieves 86.10% accuracy in offline evaluation (+23.95% over baseline) and 25.28% CTR in online deployment (149.06% relative improvement). |
Harmonizing Dense and Sparse Signals in Multi-turn RL: Dual-Horizon Credit Assignment for Industrial Sales Agents (2026.acl-industry)
Copied to clipboard
| Challenge: | Large language models for industrial sales require balancing long-term commercial objectives with immediate linguistic constraints such as fluency and compliance. |
| Approach: | They propose a framework that disentangles optimization across time scales by normalizing advantages from turn-level and session-level rewards before fusion. |
| Outcome: | The proposed framework outperforms the state-of-the-art GRPO model in conversion rate and identity detection rate. |
TRAP: Targeted Random Adversarial Prompt Honeypot for Black-Box Identification (2024.findings-acl)
Copied to clipboard
| Challenge: | Large Language Model (LLM) services and models often come with legal rules on who can use them and how they must use them. |
| Approach: | They propose a method that uses adversarial suffixes to get an answer from a target LLM. |
| Outcome: | The proposed method detects the LLMs with over 95% true positive rate at under 0.2% false positive rate even after a single interaction. |
What If Consensus Lies? Selective-Complementary Reinforcement Learning at Test Time (2026.acl-long)
Copied to clipboard
| Challenge: | Existing TTRL methods rely on positive pseudo-labeling strategies to enhance reasoning capabilities. |
| Approach: | They propose a test-time reinforcement learning framework that mitigates label noise amplification by deriving pseudo-rewards from majority voting consensus. |
| Outcome: | The proposed framework mitigates label noise amplification by implementing selective positive pseudo-labeling and entropy-gated negative p-labeled pruning. |