dialogue system developers

Abstract: A full-duplex spoken dialogue model must yield when the user cuts in and hold when the user merely backchannels. Sequence-level RL with a timing reward (GRPO) is the usual recipe; we test its whole-rollout form on four open models and find two failures. A binary window reward gives zero advantage on exactly the windows that most need learning (a zero-gradient trap), and the one score per rollout is spread over listening-state decisions the reward is not written in terms of (credit leakage), so the model talks into silence or stops answering. Both come from sampling a reward that is a known function of a few binary control decisions. We propose \emph{window likelihood} (\WL): sample a continuation per window, but integrate in closed form over the stop decisions it visits and maximise the log-probability of acting correctly---a noisy-OR for yielding, a survival product for holding. On HumDial-FDBench, \WL\ gives the best yield/hold F1 on all six structurally different open models we post-train, exceeds Kyutai's released GRPO checkpoints (trained on different data), and adds no re-opening of the floor on any model.

Abstract: In long-term interactions with large language model (LLM) agents, users reveal habits through their choices and actions. These habits capture implicit user information that agents must infer across conversations. However, individual observations may provide only partial evidence about a habit or when it applies. Retaining past interactions alone does not ensure that agents can infer habits and apply them appropriately. We introduce HABIT-Bench to evaluate how well agent memory systems infer habits from weak evidence and apply them to later requests. We construct synthetic user histories from task-oriented dialogue data and controlled habit graphs. The histories span hundreds of sessions across food, finance, software, and travel. Four-choice tasks assess agents' ability to infer user habits and apply them appropriately. They include temporary exceptions, changes in habits, and suggestions that users have not accepted. Sampled tasks undergo independent human review. We evaluate eight memory methods with Qwen3-235B-A22B and GLM-5.3-Flash. The best-performing memory method achieves 33.61\% macro accuracy across the four domains. Performance remains limited even when annotated evidence is supplied directly. These results expose a persistent gap between access to past interactions and the ability to infer and appropriately use user habits.

Abstract: Brain alignment of large language models (LLMs) offers a way to investigate representational similarities between LLMs and the human brain during language processing. While previous studies have mainly focused on passive language comprehension, little is known about such alignment during dialogue, in which speakers engage in both language production and comprehension. In this study, using fMRI data collected during dialogue, we examine brain alignment during language production and comprehension across LLM layers and brain regions. We assess its association with the intrinsic dimension (ID) of LLM representations and the contributions of several interpretable linguistic and communicative features. Our results reveal distinct patterns of brain alignment between production and comprehension, as well as strong associations with ID in brain regions associated with each processing mode. Furthermore, we show that the contributions of linguistic and communicative features also vary across processing modes, LLM layers, and brain regions.

Abstract: We study partial-revelation panel forecasting: given $k$ of $n$ annotator votes on the previous dialogue turn, predict the full count distribution of the next turn's panel. This cross-turn predictive information differs from same-item annotation value. Under any nested revelation mechanism, the population decrement satisfies $V(k)-V(k+1)=I(Y;R_{k+1}\mid X_k)$; a fitted model also contributes adjacent excess-risk differences. Coarsened-completely-at-random (CCAR) masking supplies the exchangeable random-rater interpretation used in our operators. We define coarsening-consistent panel operators with a closed-form Dirichlet-multinomial realization, learned completion, and rater-count transfer. Against per-level lookup, cold-start count negative log-likelihood differences are -0.422/-0.166/-0.014 nats on IEMOCAP/M3ED/EmoWOZ. A held-out rater experiment measures same-item value: first and second additional votes reduce NLL by 0.583/0.221, 0.578/0.184, and 0.180/0.069 nats, respectively. Cross-turn and same-item gains are not interchangeable, and CCAR masks are controlled interventions rather than observed arrival processes. Consistency is auditable but does not guarantee predictive improvement; per-level marginalization can outperform a shared table. Target-free allocation beats uniform only where row-level signal is predictable: at budget one it recovers 83--93% of retrospective feature class headroom on IEMOCAP/M3ED but does not improve EmoWOZ. Gains are modest and corpus-dependent, with several small or null partial-revelation differences. Code and aggregate result artifacts accompany the preprint.

Abstract: When a speaker says ``some students passed'', listeners infer that not all of them did. Reasoning models now let users decide how long a model thinks before it answers, but we do not know what extra thinking does to such everyday inferences. We study this directly: we keep a model's weights fixed, change only how much it thinks, and record which reading it chooses. A second tool cuts the model's own reasoning at successive points and asks for an answer, showing where the reading changes. Thinking makes models read logical words as logic: it cancels the implicatures of \emph{some} (\emph{not all}) and \emph{or} (\emph{not both}) in most tested open reasoning models and in GPT-5, while context-driven inferences such as indirect answers improve. The trace shows why: models first entertain the enriched reading and cancel it later. Answering from that early reading, an exploratory inference-time remedy, recovers these implicatures without losing literal accuracy. Evaluations of reasoning models should measure how they interpret, not only how accurate they are.

Abstract: Proactive dialogue agents must guide multi-turn interactions toward task goals while adapting to users whose intentions and attitudes evolve. Existing planners improve system side action selection through search, learned policies, or user aware planning, but are often trained and evaluated against fixed or overly compliant simulators; as a result, high simulated success can reflect soft user bias rather than robust dialogue ability. We introduce MASP, a Mentalized Adversarial Self Play framework that treats proactive dialogue as a partially observable game between a system and a rationally resisting user. MASP represents the user's evolving mental state with a BDI based latent state, separates reward side teacher mentalization from policy side student mentalization, and alternates self play updates for the system and user policies. This turns robustness to rational resistance into a training objective rather than only an evaluation condition. Across diverse proactive dialogue benchmarks, compared with the widely used SFT and Planner-only approaches, MASP substantially improves success rate (SR) and rationality constrained success (RCSR), reduces average turns (AT), and performs better against hard users who make the task goal more difficult to achieve.

Abstract: Task-oriented dialogue understanding remains underexplored for Bangla, particularly in multilingual settings involving English and romanized Bangla (Banglish). We introduce BanglaTOD, a parallel tri-lingual benchmark comprising 209 multi-turn dialogues across nine Bangladesh-specific service domains and 64 intents. Each conversational turn is aligned across Bangla, English, and Banglish, yielding 4,689 language-specific turn records annotated with intents, token-level slots, state updates, and accumulated dialogue states. We establish context-aware baselines for joint intent classification and slot filling using mBERT, BERT-base-uncased, and BanglaBERT. Cross-language analysis reveals substantial performance differences across models and linguistic representations, while removing dialogue history reduces intent accuracy by 19.0--25.7 percentage points. BanglaTOD provides a foundation for multilingual dialogue understanding and future dialogue state tracking research in Bangladesh-specific conversational applications. The dataset is publicly available at \href{https://huggingface.co/datasets/afri7here/BanglaTOD}{Hugging Face}.

Abstract: As agentic systems gain commercial popularity, user simulators increasingly serve as measurement instrument for their evaluation. However, the fidelity of simulated users in comparison to real human users is generally low, and typically assessed by costly, subjective LLM judges. In this pilot study, we ask whether fidelity can instead be measured deterministically by treating a user persona sociolinguistically: as a social type that emerges from observable linguistic style, rather than one predicted by labels or descriptions a model must extrapolate into behaviour. We author personas as concrete stylistic rates, which lets us transfer two established, model-free instruments -- authorship-verification stylometry and lexicon-based content analysis -- as fidelity diagnostics. We A/B-test the sociolinguistic schema against a flat descriptive baseline across five task-oriented customer-service agents. Results show that the sociolinguistic schema improves both stylistic adherence and stylometric distinguishability for most of the tested models, with a caveat that persona style fidelity does not necessarily equal persona "naturalness". We argue that a sociolinguistic approach to persona design is a promising path towards more diverse and representative user personas, and that these metrics are most valuable in an error-attribution analysis, localizing where fidelity breaks down. This is a first step towards interventions that move user simulations closer to faithful renderings of diverse and variable linguistic outputs.

Abstract: Reinforcement learning with verifiable rewards (RLVR) has driven substantial progress in domains where task outcomes can be reliably evaluated, but long-horizon interaction remains challenging due to sparse terminal feedback and difficult credit assignment. Process rewards provide denser supervision, yet the capabilities most relevant for training can change as the policy evolves: a behavior that is easy to evaluate or frequently deficient need not be the bottleneck currently limiting task success. We introduce RewardWeaver, a self-evolving reward adaptation framework for language agents in long-horizon interaction. RewardWeaver maintains a validated capability space in which the semantics of admitted Rubrics remain fixed, and closes the loop between policy optimization, task evaluation, failure attribution, and reward adaptation. After each training stage, it performs outcome-grounded backward attribution on low-outcome trajectories, aggregates recurrent and policy-controlled capability bottlenecks, and dynamically selects the corresponding process rewards for the next stage. Recurrent failures not covered by the existing capability space trigger a separate, controlled expansion procedure. We evaluate REWARDWEAVER on SOTOPIA, Amazon?HistoryPrice, and a newly constructed Sales Benchmark. Across social interaction, bilateral bargaining, and domain-specific sales, REWARDWEAVER establishes new state-of-the-art (SOTA) results. Ablations further demonstrate the importance of dynamic reward allocation, failure-grounded attribution, and stable semantics for admitted capabilities.

Abstract: An agent that lacks the information it needs should ask rather than act, and the task definitions of agent leaderboards say so. BFCL multi-turn builds two of its four categories around a turn on which the model is supposed to ask, and its scorer never looks at that turn: the gold trajectory there is empty, the checker skips it, and the scripted user cannot answer, so asking earns nothing, guessing costs nothing on that turn, and asking twice loses the item. The benchmark also contains the control experiment for that decision. A should-ask item is a base item with one piece of information removed from one turn, so the same request appears twice at the same turn index, once complete and once not: on the first the model should make the call that changes the world, on the second it should ask. We score one decision per pair, whether the model attempted a world-changing call on that turn, read off the stored trajectories with no LLM judge; acting always and asking always both score 50. On the 223 pairs that pose this decision, gpt-5.4 attempts the call on 83.4% of the complete turns and holds back on 78.0% of the incomplete ones, the best decision accuracy of seven models at 80.7%; on the same items the official score ranks it sixth and puts first a model that lands in the middle here. One added line telling gpt-5.4 not to ask pushes it toward acting on both sides of the pair, so its decision accuracy shows no detectable change, while its official score rises by 13.5 to 23.5 points on the two should-ask categories and on the base twins; the opposite line, telling gemma-4-31B-it to ask first, improves its decision by 4.5 points and gains no official score. The score moves with the push toward action, not with the decision. We release the pairs, a turn-level scorer that runs on any BFCL output directory without an API key, and 31 manually verified bad items.

Abstract: Quantifying conversational dynamics requires reliable identification of interactional units and their temporal boundaries, but speech activity alone does not distinguish conversational turns from listener feedback or within-turn pauses. We present an automated pipeline for extracting turns and backchannels from separate-channel recordings of spontaneous dyadic conversation, designed to provide a consistent first-pass annotation for subsequent human review. The pipeline combines voice activity detection, channel-energy filtering, temporal merging, automatic speech recognition, and context-based post-processing. We evaluated the pipeline on 99 ten-minute Danish conversations from 33 dyads using segment-level detection reliability and temporal boundary error. Conversations were recorded under both normal and asymmetric listening conditions. In the latter, speech-shaped noise was delivered to one participant through bone-conduction headphones. Overall detection reliability was F1=0.621, with similar performance for turns F1=0.624 and backchannels F1=0.618. For successfully matched segments, median absolute onset and offset errors were 0.150 and 0.160s for turns and 0.130 and 0.180s for backchannels, respectively. Mean errors were substantially larger for turn boundaries, indicating a smaller number of large boundary mismatches. Performance did not differ significantly across the two experimental listening conditions. In a four-conversation case study, pipeline-human agreement was lower and more variable than human inter-annotator agreement and varied across parameter settings. These results support the pipeline as an automated first pass within a semi-automated annotation workflow, providing a consistent basis for more standardised and reproducible annotation of conversational dynamics.

Abstract: We empirically investigate Question Under Discussion based modelling in naturalistic dialogue by studying whether the salience of generated potential questions predicts their subsequent resolution. Building on Wu et al. (2024), we construct a dataset of 7,124 questions automatically generated from utterances and preceding context from the British National Corpus, and annotated for salience and answerability. We find a robust but low positive correlation between salience and answerability in dialogue, indicating that more salient questions are more likely to be addressed. However, this effect is markedly weaker than in monologic text, suggesting that conversational structure is less predictable. We further observe that structured interactions exhibit stronger alignment between annotators than less organised dialogues.

Abstract: Backchannels, brief acknowledgements like "uh-huh" produced while the other party may still be talking, are central to natural conversation, but full-duplex spoken dialogue models rarely model them explicitly. We introduce a lightweight backchannel head that predicts, from a full-duplex model's own hidden states, when a backchannel should begin. Once this probability crosses a tunable threshold, a backchannel is force-decoded. Attached to both a 7B (PersonaPlex) and a 1B (F-Actor) model, it generalizes across scale. Probing confirms the hidden states anticipate real human timing, and generation evaluation shows more frequent, better-timed backchannels. Human raters judge the resulting backchannels on par with real ones.

Abstract: Large Language Models (LLMs) are increasingly deployed as argumentative agents in persuasive dialogues, necessitating rigorous evaluation of their debating competence relative to human interlocutors. In this study, we focus on character attacks (ad hominem arguments), traditionally dismissed as fallacies, which play a pivotal role in political persuasive dialogues where ethos often rivals propositional content. Specifically, we investigate whether modern LLMs can replicate human competence to strategically use and respond to such attacks. We analyse a corpus of natural language political dialogues to identify defensive strategies human interlocutors naturally employ in ethos-centred debates and structure them into a dialogue game. Empirically, we benchmark LLM-generated dialogues against the ElecDeb60to16-fallacy corpus of U.S. presidential debates, contrasting human debaters' repertoire of defensive strategies with those of artificial agents. Results reveal a substantial difference: most LLMs rigidly prioritise logical defences, failing to exploit ethotic counterattacks as valid moves in political discourse. We argue that current safety fine-tuning constraints the strategic action space of these LLMs, making them unable to fully engage in naturalistic interactions within domains where character contestation is a normative expectation rather than a mere fallacy.

Abstract: Full-duplex spoken dialogue systems must distinguish interruptions that require yielding the floor from backchannels that permit continued speaking. Existing benchmarks typically evaluate events independently and may therefore reward fixed action preferences rather than context-sensitive decisions. We introduce ECHO, a paired diagnostic benchmark for Chinese full-duplex turn-taking. ECHO pairs examples with the same overlap transcript but contrasting preceding multi-turn dialogue contexts, with one requiring Yield and the other Keep. It additionally includes off-talk examples for diagnosing unnecessary yielding. We introduce pair accuracy, which requires correct decisions on both members of a pair and assigns no credit to constant-action policies. Experiments on multiple full-duplex systems show that most exhibit a pronounced bias toward \textsc{Yield}, performing substantially better on interruptions than on backchannels, while another system remains comparatively balanced. These findings demonstrate that interruption-only evaluation can overestimate practical turn-taking reliability. ECHO and its metadata will be publicly released.

Abstract: Turn-taking is a fundamental component of spoken interaction, and while humans naturally rely on both verbal and non-verbal signals, dialogue systems usually depend on audio cues alone. This paper investigates whether visual features from face-to-face conversations can enhance turn-taking prediction beyond what is achievable from audio-only. We extend the Voice Activity Projection (VAP) model, a self-supervised transformer-based model for predicting future voice activity, by incorporating visual features extracted from the large-scale Meta Seamless Interaction dataset of dyadic face-to-face conversations. The visual features include gaze direction, head movement, body and hand pose, and facial action units (FAU). For incorporating the visual features, we explore concatenation, cross-attention fusion, delta features, and trainable gating mechanisms. Results show that visual information improves performance over the audio-only baseline, with FAU being significantly more informative than other feature groups. Body and gaze features nevertheless contribute complementary information, as the model combining all features performs best. Furthermore, results indicate that performance on specific tasks varies depending on whether training and test data come from improvised (acted) or naturalistic (non-acted) conversations.

Abstract: Automatic sentence function identification is important for many downstream natural language processing (NLP) applications such as dialogue systems, text-to-speech synthesis, and machine translation. However, benchmark resources for Bangla sentence function classification remain limited. To mitigate this gap, this paper introduces a corpus of 10,000 Bangla sentences, manually annotated into four functional categories, namely declarative, interrogative, imperative, and exclamatory. The corpus is nearly balanced across the four classes, with high annotation reliability reflected by a Fleiss\' Kappa of 0.82. Furthermore, we evaluate multiple feature representations, including Bag-of-Words (BoW), TF-IDF, and Word2Vec, with several classical machine learning classifiers. In addition, two heterogeneous ensemble models, namely Single-Level Ensemble (SLE) and Double-Level Ensemble (DLE), are utilized to improve classification performance. Experimental results show that TF-IDF consistently outperforms Word2Vec, likely due to its ability to emphasize discriminative lexical cues associated with sentence functions, particularly given the relatively small corpus used to train Word2Vec. The DLE model with TF-IDF features achieves the best performance with accuracy and macro-F1 of 0.95, demonstrating the effectiveness of sparse lexical representations and heterogeneous ensemble learning for this task. Further cross-validation confirms the robustness of the approach, while LIME-based interpretability provides insights into model predictions. The developed corpus and model benchmarking establish strong baselines for Bangla sentence function classification.

Abstract: Large language model serving costs scale directly with output sequence length, yet standard preference alignment often inflates response verbosity without improving utility. We study whether the parameterization of post-training updates affects generation length: low-rank subspaces alter sequence length without modifying the alignment loss. We present LOCUS, a method that selects a task-aware low-rank adaptation subspace to minimize output-token cost subject to a utility constraint. Within this subspace, post-training retains the native preference objective with a frozen backbone. Across Anthropic HH-RLHF dialogue preferences, we evaluate two $\sim$3B decoder backbones, Pythia-2.8B and Qwen2.5-3B, against protocol-matched full-parameter DPO and DrDPO branches and the released SamPO checkpoint. LOCUS reduces continuation length by up to 39.84\% on Pythia-2.8B and by 14.87--17.58\% on Qwen2.5-3B while updating only 0.24--0.28\% of model parameters, with no material change in the internal preference diagnostic.

Abstract: To realize natural behavior in dialogue agents in multi-party dialogue scenarios, it is important to understand group emotion such as valence and arousal as a whole. Most prior work addressed this task at the utterance level or using a coarse-grained time window, which is not sufficient to capture emotional dynamics. In this study, we formulate continuous recognition of the Group Emotion at a one-second resolution. Moreover, we also introduce the Mixed state, which captures the emotional divergence among participants in the group. We constructed a dataset with frame-level soft labels based on the TEIDAN corpus and propose a multimodal temporal framework that integrates audio and video information using a sliding-window context. Experimental results demonstrate that the temporal Transformer outperforms simple baselines and shows stronger temporal agreement with the ground-truth labels than the LLM-based model. The effect of context length is limited, whereas audio-visual input outperforms either unimodal input on the continuous-label metrics. Additionally, our analysis shows larger Group Emotion recognition errors in intervals with high Mixed values, exposing emotional divergence as a key challenge for group emotion recognition.