Tag
This paper identifies a novel failure mode in reasoning models called unfaithful capitulation, where the chain-of-thought remains factually correct across adversarial multi-turn dialogues but the final answer flips wrong, highlighting limitations of current evaluation methods.
This paper theoretically identifies and mitigates context distribution shift in multi-turn dialogue RL, proposing Calibrated Interactive RL that couples interactive RL with simulator alignment to reduce the sim-to-real gap and achieve state-of-the-art performance.
Proposes SKG-Eval, a quasi-deterministic evaluation framework for multi-turn dialogue that uses incremental semantic knowledge graphs to detect cross-turn inconsistencies, contradiction, and topic drift, achieving higher correlation with human judgments.
SCICONVBENCH is a benchmark that evaluates LLMs on multi-turn clarification for ill-posed scientific queries across computational science domains, finding that even frontier models struggle with disambiguation and frequently make silent assumptions.
This paper describes a system for SemEval-2026 Task 8 that uses a three-stage pipeline involving query rewriting with a fine-tuned Qwen model, hybrid retrieval, and cross-encoder reranking to improve multi-turn retrieval performance.
This paper introduces SOMA, a framework for efficient multi-turn LLM serving that uses small language models adapted via soft prompts and LoRA fine-tuning to reduce latency and cost.
This paper introduces TRACE, a framework for turn-aware credit assignment in multi-turn LLM jailbreaking attacks using reinforcement learning, claiming significant improvements in attack success rates and defense alignment.
Presents TurnGate, a turn-level monitor that detects hidden malicious intent in multi-turn dialogues by identifying the earliest turn where a response would enable harmful action, along with the Multi-Turn Intent Dataset (MTID) to support training and evaluation.
Study shows GPT and Claude exhibit distinct, unreliable repair behaviors in multi-turn math dialogues, with some models resisting correction and others over-correcting.
This paper introduces Bipredictability (P) and the Information Digital Twin (IDT), a lightweight method to monitor conversational consistency in multi-turn LLM interactions using token frequency statistics without embeddings or model internals. The approach achieves 100% sensitivity in detecting contradictions and topic shifts while establishing a practical monitoring framework for extended LLM deployments.
Context-Agent proposes a novel framework that models multi-turn dialogue history as dynamic tree structures rather than flat sequences, better capturing the hierarchical and branching nature of natural conversation. The paper introduces the NTM benchmark for evaluating non-linear dialogue scenarios and demonstrates improved task completion rates and token efficiency across various LLMs.