People increasingly turn to large language models (LLMs) for everyday advice, making ethically charged interpersonal problems a practical moral-advisory context. Most prior work has studied this context through single-turn judgments or pressure-laden rebuttals, assumptions that poorly match how guidance is sought in real-world contexts. These assumptions leave unclear whether narration alone, without an explicit opposing position, can shift model judgments during multi-turn moral consultation. Yet real-world moral-conflict conversation often elicits one party's self-justifying account, which can unfold over multiple turns and create information asymmetry. We introduce \textbf{narrative captivity}, a failure mode in which a model treats an unopposed one-sided account as complete and aligns with the narrator's interpretation without seeking missing perspectives. To measure this phenomenon, we build a benchmark of $5{,}078$ interpersonal-conflict scenarios spanning six moral dimensions. Across 17 LLMs, narrative captivity is widespread: end-state judgments under multi-turn narration shift by 25 percentage points on average beyond the matched single-turn baseline. Stage-level analysis identifies preference optimization as a major contributor, while four inference-time strategies provide only partial mitigation. We hope our project fosters LLM advisors that preserve independent judgment in real-world consultation.
Jihoon Tack, Philippe Laban, Jennifer Nevillecs.LG
As LLMs become more capable, they are increasingly deployed as collaborative agents, taking on user-delegated tasks through iterative interaction. Yet genuine interaction is inherently dynamic: users rarely specify their intent upfront, instead disclosing, revising, and reshaping it as the conversation unfolds. Despite this, LLMs are still predominantly evaluated or trained in single-turn, fully-specified settings, leaving open a fundamental question: how well do LLMs track and act on user intent as it evolves over the course of a conversation? To study this, we introduce a framework that transforms static, single-turn tasks into dynamic multi-turn conversations in which the user's intent evolves across turns--incrementally revealed, revised, and at times redirected mid-conversation--while preserving each task's original evaluation protocol, enabling existing benchmarks to be reused as controlled testbeds without new annotation. Across multiple tasks, we surface a consistent phenomenon: strong static-setting performance does not transfer to the evolving-intent setting, with substantial drops across model families. Our findings point to a fundamental gap: today's LLMs do not yet faithfully track and act on the user's evolving intent, a capability invisible to static evaluation yet critical for future collaborative agents.
When a language model produces a response in a multi-turn conversation, which tokens from prior turns shaped that answer, and how did those dependencies propagate across prior turns? Existing context attribution methods process the full context in a single pass, recovering surface-level dependencies but missing the layered, non-linear structure of real-world dialogues and multi-step reasoning tasks. We introduce multi-turn context attribution (MTCA): given a target span in a model response, the task of tracing attribution backward across turns to identify not only which prior turns were directly relevant, but also how those turns themselves depended on earlier context. We propose Tokengeist, an attribution-method-agnostic and scalable framework that recovers full dependency paths by casting attribution as a recursive traversal of a directed acyclic graph (DAG) over conversation turns. We will release MTCABench, a benchmark of 3,845 target spans across 665 multi-turn conversations, annotated with gold provenance graphs reaching depths of up to 14, across four dependency types. Across four open-weight models, flat attribution methods fail to recover multi-hop dependencies, achieving under 20% source recall, while Tokengeist reaches 90%. Our results reveal systematic failure modes of single-pass attribution -- which we term provenance collapse -- and motivate attribution methods that reason recursively across turns.
Production teams deploying LLM chat agents face a specific quality assurance gap: existing evaluation tools test individual responses or simulate social interactions, but none systematically verify whether real users can achieve their goals through multi-turn conversation. We introduce a three-layer dogfooding framework that bridges this gap by combining canonical question-bank testing (Layer 1), random-walk multi-turn evaluation (Layer 2), and a goal-directed NPC (Non-Player Character) simulator with five structured goal types and a ten-category failure taxonomy (Layer 3). In a longitudinal case study on a production multi-agent system over roughly three months (257 evaluation runs; a 108-scenario NPC suite), we find that the three layers produce complementary regression signals: cross-layer correlation for response quality is weak within a synchronized run (Spearman rho between -0.15 and 0.14) and negative across the longitudinal series (rho down to -0.46), confirming that canonical correctness does not predict goal-directed conversation success. The NPC simulator achieves 77 percent goal achievement at 0.17 dollars per run (6,272x cheaper than human evaluation), enabling daily CI/CD integration with automated PROMOTE/HOLD/ROLLBACK release decisions. We release full prompt templates, the failure taxonomy, and a Python-first replicability guide so that other teams can adopt the framework for their own LLM chat agents.