Beyond Continuity: Challenges of Context Switching in Multi-Turn Dialogue with LLMs

Users interacting with Large Language Models (LLMs) in a multi-turn conversation routinely refine their requests or pivot to new topics. LLMs, however, often miss these topic shifts and carry over irrelevant context from previous turns, leading to inaccurate responses. In this paper, we stress-test the multi-turn understanding of LLMs and study the following two sub-tasks: (1) detecting whether the user pivots or refines in the current turn, and (2) shortlisting relevant context from previous turns. To this end, we construct synthetic benchmarks based on real-world datasets from varied domains, as to simulate context shifts of different levels of difficulty. We then evaluate the zero-shot performance of ten LLMs (open-weight, closed-source and reasoning), and demonstrate that only some reasoning and strongly instructed LLMs are accurate in detecting pivots; open-weight LLMs struggle with the task and frequently carry stale context even with explicit cues; and all models suffer from a position bias. Based on the results, we discuss key takeaways for improving long-term robustness in multi-turn capabilities for LLMs.

G$^2$TR: Generation-Guided Visual Token Reduction for Separate-Encoder Unified Multimodal Models

The development of separate-encoder Unified multimodal models (UMMs) comes with a rapidly growing inference cost due to dense visual token processing. In this paper, we focus on understanding-side visual token reduction for improving the efficiency of separate-encoder UMMs. While this topic has been widely studied for MLLMs, existing methods typically rely on attention scores, text-image similarity and so on, implicitly assuming that the final objective is discriminative reasoning. This assumption does not hold for UMMs, where understanding-side visual tokens must also preserve the model's capabilities for editing images. We propose G$^2$TR, a generation-guided visual token reduction framework for separate-encoder UMMs. Our key insight is that the generation branch provides a task-agnostic signal for identifying understanding-side visual tokens that are not only semantically relevant but also important for latent-space image reconstruction and generation. G$^2$TR estimates token importance from consistency with VAE latent, performs balanced token selection, and merges redundant tokens into retained representatives to reduce information loss. The method is training-free, plug-and-play, and applied only after the understanding encoding stage, making it compatible with existing UMM inference pipelines. Experiments on image understanding and editing benchmarks show that G$^2$TR substantially reduces visual tokens and prefill computation by 1.94x while maintaining both reasoning accuracy and editing quality, outperforming baselines on almost all benchmarks.

Grounded Satirical Generation with RAG

Humor generation remains challenging task for Large Language Models (LLMs), due to their subjective nature. We focus on satire, a form of humor strongly shaped by context. In this work, we present a novel pipeline for grounded satire generation that uses Retrieval-Augmented Generation (RAG) over current news to produce satirical dictionary definitions in the Finnish context. We also introduce a new task-specific evaluation framework and annotate 100 generated definitions with six human annotators, enabling analysis across multiple experimental conditions, including cultural background, source-word type, and the presence or absence of RAG. Our results show that the generated definitions are perceived as more political than humorous. Both topic-based word selection and RAG improve the political relevance of the outputs, but neither yields clear gains in humor generation. In addition, our LLM-as-a-judge evaluation of five state-of-the-art models indicates that LLMs correlate well with human judgments on political relevance, but perform poorly on humor. We release our code and annotated dataset to support further research on grounded satire generation and evaluation.

Primal-Dual Guided Decoding for Constrained Discrete Diffusion

Discrete diffusion models generate structured sequences by progressively unmasking tokens, but enforcing global property constraints during generation remains an open challenge. We propose primal-dual guided decoding, an inference-time method that formulates constrained generation as a KL-regularised optimisation problem and solves it online via adaptive Lagrangian multipliers. At each denoising step, the method modifies token logits through an additive, constraint-dependent bias, with multipliers updated by mirror descent based on constraint violation. The bias arises as the optimal KL-regularised projection of the constraint, so the constrained distribution remains as close as possible to the model's unconstrained distribution while still satisfying the constraint. The method requires no retraining and no additional model evaluations beyond standard sampling, supports multiple simultaneous constraints, and provides formal bounds on constraint violation. We evaluate our approach on topical text generation, molecular design, and music playlist generation, showing that a single algorithm instantiated via domain-specific scoring functions improves constraint satisfaction while preserving relevant domain-specific quality metrics.

Output Composability of QLoRA PEFT Modules for Plug-and-Play Attribute-Controlled Text Generation

Parameter-efficient fine-tuning (PEFT) techniques offer task-specific fine-tuning at a fraction of the cost of full fine-tuning, but require separate fine-tuning for every new task (combination). In this paper, we explore three ways of generalising beyond single-task training/inference: (i) training on combinations of multiple, related datasets; (ii) at inference, composing the weight matrices of separately trained PEFT modules; and (iii) at inference, composing the outputs of separately trained PEFT modules. We test these approaches on three different LLMs, QLoRA as the PEFT technique, and three sets of controlled text generation datasets for sentiment control, topic control, and multi-attribute control. We find that summing PEFT module outputs is a particularly strong composition method, which consistently either outperforms or matches the performance of alternative approaches. This is the case even when comparing against single-task specialised modules on the single-task test set, where three-module output composition achieves an average 2% point performance increase across all models for sentiment control.

Building Korean linguistic resource for NLU data generation of banking app CS dialog system

Natural language understanding (NLU) is integral to task-oriented dialog systems, but demands a considerable amount of annotated training data to increase the coverage of diverse utterances. In this study, we report the construction of a linguistic resource named FIAD (Financial Annotated Dataset) and its use to generate a Korean annotated training data for NLU in the banking customer service (CS) domain. By an empirical examination of a corpus of banking app reviews, we identified three linguistic patterns occurring in Korean request utterances: TOPIC (ENTITY, FEATURE), EVENT, and DISCOURSE MARKER. We represented them in LGGs (Local Grammar Graphs) to generate annotated data covering diverse intents and entities. To assess the practicality of the resource, we evaluate the performances of DIET-only (Intent: 0.91 /Topic [entity+feature]: 0.83), DIET+ HANBERT (I:0.94/T:0.85), DIET+ KoBERT (I:0.94/T:0.86), and DIET+ KorBERT (I:0.95/T:0.84) models trained on FIAD-generated data to extract various types of semantic items.

Cited but Not Verified: Parsing and Evaluating Source Attribution in LLM Deep Research Agents

Large language models (LLMs) power deep research agents that synthesize information from hundreds of web sources into cited reports, yet these citations cannot be reliably verified. Current approaches either trust models to self-cite accurately, risking bias, or employ retrieval-augmented generation (RAG) that does not validate source accessibility, relevance, or factual consistency. We introduce the first source attribution evaluation framework that uses a reproducible AST parser to extract and evaluate inline citations from LLM-generated Markdown reports at scale. Unlike methods that verify claims in isolation, our framework closes the loop by retrieving the actual cited content, enabling human or model evaluators to judge each citation against its source. Citations are evaluated along three dimensions. (1) Link Works verifies URL accessibility, (2) Relevant Content measures topical alignment, and (3) Fact Check validates factual accuracy against source content. We benchmark 14 closed-source and open-source LLMs across three evaluation dimensions using rubric-based LLM-as-a-judge evaluators calibrated through human review. Our results reveal that even the strongest frontier models maintain link validity above 94% and relevance above 80%, yet achieve only 39-77% factual accuracy, while fewer than half of open-source models successfully generate cited reports in a one-shot setting. Ablation studies on research depth show that Fact Check accuracy drops by approximately 42% on average across two frontier models as tool calls scale from 2 to 150, demonstrating that more retrieval does not produce more accurate citations. These findings reveal a critical disconnect between surface-level citation quality and factual reliability, and our framework provides the evaluation infrastructure to assess the disconnect.

Reward Hacking in Rubric-Based Reinforcement Learning

Reinforcement learning with verifiable rewards has enabled strong post-training gains in domains such as math and coding, though many open-ended settings rely on rubric-based rewards. We study reward hacking in rubric-based RL, where a policy is optimized against a training verifier but evaluated against a cross-family panel of three frontier judges, reducing dependence on any single evaluator. Our framework separates two sources of divergence: verifier failure, where the training verifier credits rubric criteria that reference verifiers reject, and rubric-design limitations, where even strong rubric-based verifiers favor responses that rubric-free judges rate worse overall. Across medical and science domains, weak verifiers produce large proxy-reward gains that do not transfer to the reference verifiers; exploitation grows over training and concentrates in recurring failures such as partial satisfaction of compound criteria, treating implicit content as explicit, and imprecise topical matching. Stronger verifiers substantially reduce, but do not eliminate, verifier exploitation. We also introduce a self-internalization gap, a verifier-free diagnostic based on policy log-probabilities, which tracks reference-verifier quality, detecting when the policy trained using the weak verifier stops improving. Finally, in our setting, stronger verification does not prevent reward hacking when the rubric leaves important failure modes unspecified: rubric-based verifiers prefer the RL checkpoint, while rubric-free judges prefer the base model. These disagreements coincide with gains concentrated in completeness and presence-based criteria, alongside declines in factual correctness, conciseness, relevance, and overall quality. Together, these results suggest that stronger verification reduces reward hacking, but does not by itself ensure that rubric gains correspond to broader quality gains.

DataDignity: Training Data Attribution for Large Language Models

Auditing language-model outputs often requires more than judging correctness: an auditor may need to identify which source document most likely supports the knowledge expressed in a response. We study this as pinpoint provenance: given a prompt, a target-model response, and a candidate corpus, rank the documents that best support the response. We introduce FakeWiki, a controlled benchmark of 3,537 fabricated Wikipedia-style articles designed to preserve ground-truth provenance while weakening lexical shortcuts. FakeWiki includes QA probes, source-preserving paraphrases, retro-generated variants, hard anti-documents that remain topically similar while removing answer-critical facts, and five query conditions: clean prompting plus four jailbreak-inspired transformations. We evaluate seven retrieval baselines, a training-free activation-steering retrieval-fusion method, SteerFuse, and a supervised contrastive provenance ranker, ScoringModel. ScoringModel maps response and document features into a shared space and is trained with InfoNCE using in-batch, retrieval-mined, and anti-document negatives. Across nine open-weight instruction-tuned LLMs and five query conditions, ScoringModel improves mean Recall@10 from 35.0 for the strongest retrieval baseline to 52.2, without inference-time fusion, and wins 41/45 model-by-condition cells. SteerFuse is usually second-best despite requiring no supervised training, showing that activation-space evidence can efficiently complement text retrieval. On jailbreak-inspired transformed queries, ScoringModel improves Recall@10 by 15.7 points on average over the best baseline. Overall, our work shows that robust training data attribution requires evaluation settings that separate true answer support from topical or lexical resemblance.

An Investigation of Linguistic Biases in LLM-Based Recommendations

We investigate linguistic biases in LLM-based restaurant and product recommendations given prompts varying across Southern American English (AE), Indian English (IE), and Code-Switched Hindi-English dialects, using the Yelp Open dataset (Yelp Inc., 2023) and Walmart product reviews dataset (PromptCloud,2020). We add lists of restaurant and product names balanced by cuisine type and product category to the prompts given to the LLM, and we zero-shot prompt the LLMs in a cold-start setting to select the top-20 restaurant and product recommendations from these lists for each of the dialect-varied prompts. We prompt LLMs using different list samples across 20 seeds for better generalization, and aggregate per cuisine-type and per category response counts for each seed, question/prompt, and LLM model. We run mixed-effects regression models for each model family and topic (restaurant/product) with the aggregate response counts as the dependent, and conduct likelihood ratio tests for the fixed effects with post-hoc pairwise testing of estimated marginal means differences, to investigate group-level differences in recommendation counts by model size and dialect type. Results show that dialect plays a role in the type of restaurant selected across the models tested with the mistral-small-3.1 model and both the llama-3.1 family models tested showing more sensitivity to Indian English and Code-Switched prompts. In terms of product recommendations, the llama-3.1-70B-model is particularly sensitive to Code-Switched prompts in four out of seven categories, and more beauty and home category recommendations are seen when using the Indian English and Code-Switched prompts for larger and smaller models, respectively. No broad trends are seen in the model-size based differences, with differing recommendations based on model sizes conditioned by the type of dialect.