A Context Alignment Pre-processor for Enhancing the Coherence of Human-LLM Dialog

A Contex t Alignment Pre-processor for E nhancing the Coherence of Human –LLM Dialog DING WEI College of Architecture, Nanjing Tech University kxrdk@163.com Abstract: Large language models (LLMs) have made remarkable progress in generating fluent text, b ut they still face a critical challenge of contextual misalignment in long-term and dynamic d ialog ue. When human users omit premises, si mplify references, or shift contex t abruptly during interactions with LLMs, the models may fail to capture their a ctual intentions, producing mechanical or off-topic re sponses that weaken the c ollabor ative potential of dialogue. To address this problem, this paper proposes a computational framewo rk ca lle d the Context Alignment Pre-processor (C.A.P.). Rather than operating during generation, C.A.P. functions as a pre- processing modul e between u ser input and response generation. The framework includes thre e core processes: (1) semantic ex pansion, which e xtends a user ins truction to a broader semantic span including its premises, literal meaning, and implications; (2) t ime- weighted context retrieval, which p rioritizes recent dialog ue history through a tempor al deca y fun ction appr oximating h uman c onvers ational focus; and (3) alignment verification and decision branching, which evaluates whether the dialo gue remains on t rack b y measuring the semantic similarity between the c urre nt prompt and the weighted historical context. When a significant deviation is d etected, C.A.P. initiates a st ructured clarification protocol to help users and t he system recalibrate the conversatio n. This study p resents the architecture and theoretical b asis of C.A.P., d rawing on cog nitive science and Common G round theory in human-comp uter interaction. We argue that C.A .P. is not o nly a te chnical re finement but also a step tow ard shifting human -computer d ialog ue from one- way command- execution patterns to two-way, self-correcting, partnership-based collaboration. Finally , we discuss implementatio n paths, evaluation methods, and imp lications fo r the future design of interactive intelligent systems . Keyw ords: Large Language Models, Human–Computer In teraction, Dialog S ystems, Contextual Understanding, Common Ground, Int ent Alignment, Co mputational Framework 1 INTRODUCTION Since th e advent of the transformer architecture, such large language models (LLMs) as the GPT series and LlaMA have emerged as the most transformative force in natural language p rocessing [1, 2]. They have d emonstrated remarkable capabilities in tasks such as the generation , summarizatio n, an d translation of text as well as answering questions, and are being used increasingly co mmonly as tools for hum an k nowledge worke rs and creators. Ho wever , as human–machine interactions evolve from simple, single-turn questions and answers (Q&A) to complex, long-term, and multi-turn collaborative dialogs, a d eep limitation o f LLMs h as gradually emerged: contextual misalignment. Human conversation is in herently an efficient but “uncertainty-rich” co llaborative process. P articipants rely on shared knowledge, common focal points, and a continually updated “common ground” [3 ] to comprehend each other's implied premises, simplified references, and non-linear leaps in reason ing. However, cu rrently available LLMs remain largely “faithful but n aïve” execu tors. They primarily rely on li mited co ntextual windows and attent ion mechanisms t o interpret user input. When u sers issue seemingly simple comman ds th at i mply co mplex historical con texts or subtle shifts in int ent, the models often deviate d ue to their inability to d ynamically track deeper contextual nuances. This deviation manifests in t he following forms: (1) Mechan ical respo nses: M odels strictly execute instructions b ased o n their literal meaning, whi le igno ring their true p urpose with in th e given conversational flow. 2 (2) Focus drift: The model is co mpletely derailed from the established conversational thread by n ew instructions, which results in logical di sconnects. (3) Missed opportunities: The model fails to recognize the creativ e potential of or invitati ons for deeper exploration that are embedded in new user instructions, an d maintain s the con versation at a superficial level. The above issues not only reduce the efficiency of in teraction, b ut, more crit ically , also hinder the formation of genuine “intellectual partn erships” between humans and LLMs. T o ad dress th is co re challen ge, t he relevan t research has primarily focused on expanding context windows, o ptimizing attention mechani sms, and enhancing model compliance through in struction tuning [4]. Ho wever , these approaches are largely “passive” adaptations that fail to fund amentally endow mod els with th e ability to actively calibrate th e context. This paper proposes a solution to t he above p roblem called th e Context Al ignment Pre-processor (C. A.P .). The C.A.P . is a lightweight an d mod ular computational framework that operates prior t o the LLM's primary task (response generation), with the sole objective o f ensuring that the model and user are “on t he same page” at the given t ime step. By simulating the human conversational mechanisms of reflection and confirmation, it dynamically evaluates the consistency between t he given instructions and th e history of the d ialog. Once it d etects a potenti al “misalignment,” it pauses generation and initiates a clarification protoco l, th us tran sforming th e b urden o f “guessin g” into an explicit and collaborative contextual calibration wi th th e u ser . The main co ntributions of t his paper are as follows: (1) It introduces a compu tational framework (C.A.P.) that is specifically designed to proactively manage and calibrate the understanding of t he conversation al cont ext by th e LLM before response generation. (2) It formally defines the three core components of t he C.A.P. —semantic expansio n, time-weighted context retracing, and verification o f alignment with decisio n bran ching—and elucid ates their collaborative mechanism. (3) It grounds the C.A.P. framework in robust theoretical foundations from cognitive science and Human- Computer Interaction (HCI), p articularly its connection to the Common Ground theory and mechani sms o f conversational repair. This establi shes i ts theoretical legitimacy. (4) This p aper explores path way s for the implementation of the C.A.P, metrics for its evaluation, and directions for future research in the area. This o ffers a feasible technical ro admap to elevate LLMs from “tools” to “partners.” The remainder of this p aper i s structu red as follows: Section 2 reviews related work in the field, Section 3 details the architecture and workflow of the C.A.P . framework, while Section 4 explo res its theoretical foundations and deeper implications. Section 5 d escribes methods for implementing and evaluating the C.A.P , Section 6 d iscusses its p otential impacts and limitations, an d Sectio n 7 summarizes the conclusio ns o f this paper. 2 RELA T ED WORK The motivation for research on the C.A.P . framework an d i ts d esign philo sophy are closely related to curren t research in three domains: context processing for LLMs, dialog management system s, and foun dational t heories of human–computer interaction. 2.1 C ontext Handling an d Limitations of LLMs Modern LLMs are built upon the transformer architecture, th e self-attenti on mechanism of which enables th em to weigh the imp ortance of parts within the i nput sequ ences [1 ]. Theoretically, th is allows the models to captu re l ong-range 3 dependencies. In practice, however, the contextual understanding o f LLMs remains constrained by several factors. The first is the finite con text window . Although window sizes are in creasing (from thousands to millions of to kens), they ultimately face physical limits. Critical early in formation for extremely long dialogs may thus be discarded. The second constraint on the windo w size is the “lost in the midd le” phenomenon. Research suggests that when processing long inputs, LLMs focus pr imarily on information at t he b eginning and end , such th at th e midd le section s can be easily overlooked [5]. F inally , th e recency bias causes mod els to o veremphasize recent rounds of dialog while potentially overlooking early groundwork th at sets th e t one for the overall conversation. The C.A.P .'s time-weighted backtracking mechanism specifically counters this bias by algorithmically forcing the model to revisit and evaluate the importance o f the h istorical co ntext. 2.2 D ialog Management The Dialo g M anager handles Dialog State Tracking (DST) in traditional task-oriented dialog systems [6]. DST aims to accurately represent the u ser's in tent and the slot values in each time step based on the dialog history . However, traditional DST primarily applies to well-defined an d d omain-restricted tasks (e.g., b ooking tickets and weather-related queries). For open-domain and creative collaborative dialogs, the user's “intent” is fluid and emergent, and is difficult to characterize b y using p redefined slot s. Recent research has attempted to leverage the LLMs themselves for d ialog state management [7], but this approach often couples it with t he generation task, such th at a d edicated mechanism for “reflective” context alignment is lacking. The C.A.P . can be viewed as a novel an d light weight dialog manager t hat tracks higher-leve l “semantic coherence,” rather than specific “slots,” an d can proactively in terrupt and rep air when incoherence is detected. 2.3 C ommon Ground in Human–Computer Interaction (HCI) The “common ground” theo ry , proposed by Clark and Brennan, p ertains to th e knowledge, beli efs, and assump tions shared by the participants o f a d ialog [3]. Establishing and maintain ing common ground is crucial for successful communication. When one party p erceives potential d eviations in the common ground, they initiate “repair mechanisms,” such as requesting clarificatio n (“Did you mean ...?”). In human–computer interaction, enabl ing machin es to effectively participate in the construction of a common ground remains a core challenge [8 ]. Current research has primarily focused on en abling systems to generate more “context-aware” responses, such as by referencing prior dialogic content. However, these approaches are passive. The C.A.P .'s u niqueness lies in its explicit algori thmic formalization of repair mechanisms. Its “alignment check” process simulates the computatio nal assessment of t he stability of a shared ground, while its “clarification protocol” d irectly b orrows from human co nversational repair mechanisms. This end ows the AI with the unprecedented capabi lity of acknowledging that it may have “lost track” and req uesting assistance from its human p artner . This marks a significant shift from p ursuing “omniscient” AI toward pursuing “ honest, collaborative” AI. In summ ary, the C.A.P . framework fills a critical gap in prevalent research by introducing a p re-processing stage that is independent of the generatio n task, and by explicit ly simulating human conversational reflection and rep air mechanisms. This provides LLMs with a structured metho d for actively managing and calibrating the conversation al context. 3 DET AILED EXPLANA TION OF C. A.P . FRAMEWORK The core d esign philosophy of the C.A.P . is to “think b efore acting.” Upon receivin g a new user inst ruction A at ti me 4 point T A , it does not immediately p ass it to the LLM for generation. Inst ead, it initiates a preprocessing task comprising three sequ ential p rocesses. 3.1 O verall Framework Architecture The C.A.P . fun ctions as middleware, and is positioned between the user an d the LLM's core generation module. Its workflow is as follows: (1) In put: Real-time requ est A submitted by th e user at time point T A . (2) C. A.P. Pro cessing: Process 1: S emantic expansion . Process 2: Time-weighted context retrieval. Process 3: Alignmen t check and decisio n bran ching. (3) Ou tput: If aligned: Pass the original in struction A (possibly with a context summary appended) to the LLM generation module. If misaligned: S uspend th e main task and present the user with the “ Clarification Protocol” interface. 3.2 Pr ocess One: Semantic Expansion This p rocess is designed to overcome the li mitation of literal in terpretation of u ser inst ructions by the LLM. It expands a single instructio n A in to a set Set ( A ) t hat encompasses its potential semantics, thereby tran sforming a “point-like” instruction i nto an “interval-like ” seman tic space. 󰇛󰇜        A (Literal): This is th e user's ori ginal instruction, serving as the center point of t he semantic space. A− (Prerequisite/Foun dation) : This constitutes implicit prerequisites, foundatio nal definiti ons, or a more specific version required to execu te command A . For example, if A is “Provid e the formula for th e functio nal synergy index of each distri ct in a city,” A− may be “ First define what constitutes functional complementarity an d activity correlation amon g d istricts in a city.” A+ (Implication/Application) : This is the l ogical extensio n, scenario of applicatio n, o r a b roader an d more exploratory version of in struction A . For example, if A is “Provide a formula for the functional synergy index of urban districts,” A+ cou ld be “Explore how this index can be used to construct urban functional networks and perform a commun ity analysis o f urban dist ricts.” A− and A+ can be generated through a single, small LLM invocation by using such meta-promp ts as “What prerequisites are needed to execute thi s instruction?” and “What is the next step for this instructio n?” This step aims to capture b roader seman tic associatio ns for the subseq uent verification o f alignment. 3.3 Pr ocess T wo: Time-weighted Context Retrieval This process simulates the focal n ature of h uman memory , in which recent dialogic content is typically most relevant. However, the crucial early context should not b e forgotten either . It retrieves a weighted context subset from t he complete d ialog history .   󰇝    󰇞 5 (1) Retrieval : Extract the k most recent ro unds o f d ialogic h istory, e.g., Hcontext  󰇝      󰇞 where Ai represent s u ser in structions and Ri denotes model responses. (2) Weighting : Assign a weight Wi to each h istorical dialogic round Hi (or each instructio n). This weight is a decreasing function of its temporal dist ance from the current time TA . A simple yet effective function to this en d is the in verse proportio nal functio n:   󰇛  󰇜  󰇛  󰇜    where T i is the t imestamp of the h istorical i nstruction Hi , and τ is a temporal scale-related parameter that contro ls the rate of decay o f the weights. This formula ensures that th e largest weight is assigned to the most recent dialogic tu rn, while earlier turns exhibit a smoo th, non-zero weight d ecay . 3.4 Pr ocess Three: Alignment Check and Decision Branching The C.A.P . is responsible for making t he final decision. It does so by calcul ating the alignment score between t he semantic space S et ( A ) of th e current prompt and the weighted historical context. (1) Vectorization : By u sing a pre-trained mod el of sentence embedding (e.g., S entence-BERT [9]), convert each element in S et(A) (A−, A, A+), and each instruction Hi in the weighted history in to a high-dimensio nal semantic vector v ( x ). (2) Calculate Alignment Score : The alignment score Salign  is defined as th e maximum weighted similarity between S et ( A ) and the hi storical con text:  align 󰇛 󰇜   󰇛󰇜          󰇛󰇜    where sim( v 1 , v 2) is th e cosine similarity . It measures the extent to which the current instruction (and it s latent semantics) can b e “explained” by the most recent and relevant di alog history . (3) Decision B ranch : Compare the computed alignment score Salig n with a preset threshold θ . If Sal ign ≥ θ (Alig nm ent Confirmed): Conclusion : The cu rrent in struction A is a natural continuation of the flow o f the con versational lo gic. Action : Execute normally. Pass instruction A to th e LLM core generation module. To further enhance coherence, inject t he most similar h istorical ent ry H j into the prompt as additional context. If Sal ign  < θ (Misalignment Alert): Conclusion : A p otential jump in con text o r ambiguous instruction has been d etected. This means that the user intent may h ave shifted significantl y, o r a simplified expression exceeds the boundaries of safe i nference. Action : Initiate the Clarificati on P rotocol. Pause the p rimary task and present the user with a structured interface: Repeat : “You r current real-ti me requ est is: ‘[Repeat in struction A].’” Alert : “I n ote that t his request appears substantially different in subject matter from o ur previous discussion of ‘[Repeat most similar h istorical in struction Hj ].’” 6 Empower : “To b etter understand your in tent, I need your assistance. Would you like to:” Offer Choices : a) P roceed with this new req uest; b) Co rrect my understand ing—your req uest is actu ally a deepening or variation of t he previou s topic; c) Alternatively, provide a clearer new request. This p rotocol is designed to p olitely and non-confrontationally return con trol to the user and collaboratively resto re a “shared foundatio n.” 4 THEORETICAL FOUNDA TIONS AND SIGNIFICANC E The C.A.P . framework is not merely an engineering solution; i t is deeply rooted in the theoretical foun dations of cognitive science and human–computer interaction, which endows it with significance beyond pure technical optimization. 4.1 Fr om Cognitive Science: Simulating Human Reflection and Repair Human dialog is far from a perfect linear process. It is filled with interruptions, corrections, and clarifications. These “disruptions” are p recisely the key mechani sms ensuring su ccessful communication . When one party in a conversation is uncertain about th eir understanding of the meaning of the other, they instinctively pau se and seek con firmation through questioning, p araphrasing, or other means. This is a form o f metacognit ive ability as th e awareness of o ne's o wn cognitive state. The C.A.P .'s “alignment check” is a computational simul ation of th is metacognitive reflection. It prevents AI from being overconfident, an d teaches it t o practice “self-do ubt.” The “clarification p rotocol” directly implements dialog repair mechanisms. Through this process, AI t ransforms from a passive i nformation processor in to an active participant in commun ication. It can identify potential barriers to communication and i nvite it s human partner t o collaboratively overcome t hem. 4.2 Fr om an HCI Perspective: Building and Maintaining a “Shared Ground” As previously noted, shared ground is the cornerstone o f collaborative activities. Clark and Brennan [3] have noted that diff erent communication media carry varying “grounding costs” in supporting the construction o f shared ground. Face- to-face human interaction in curs the lowest co st, as th e particip ants can rapidly confirm a common understanding through multiple channels, li ke eye contact and gestures. By cont rast, text-based human–computer in teraction entail s significantly h igher costs for establ ishing a shared ground. The C.A.P . framework can b e viewed as a mechanism designed to reduce th e costs of grounding in human–computer dialog. When it detects potenti al instability in grou nding (i.e., low ali gnment scores), it rap idly rebui lds con sensus through a low-cost clarificatory in teraction, thereby avoiding the substantial sun k co sts o f subsequent rounds of dialog caused b y misunderstanding. From thi s perspective, the C. A.P . carves out an efficient path for maintaining a shared foundation between h umans and machines with in th e limited text-based channel of interaction. 4.3 Pa radigm Shift: F rom “T ool” to “Partner” The u ltimate significance of t he C.A.P . lies in the fact that it represents a paradi gm shift in human–machine relations. Tool Paradigm : AI acts as a passive execu tor, and humans b ear the respo nsibility of issui ng clear and unambiguous in structions. The burden of communication th us rests entirely on the human side. 7 Partner Paradigm : AI acts as an active col laborator. It recognizes ambi guities in communication an d shares the responsibility for clarifi cation wit h h umans. Communicatio n th en becomes bidirectio nal, and is jointly constructed. By endowing AI with th e capabi lities of “reflection” and “seeking assistance,” the C.A.P . enables patterns of AI behavior th at closely resemble th ose of a true conversational partner. This partnersh ip is built on trust, which stems from AI's ability to acknowledge it s li mitations and co mmit t o achieving a deep understanding in i ts co llaboration with humans. 5 P A THW A YS OF IMPLEMENT A TION AND EV ALUA TION As a con ceptual framework, the value of th e C.A.P . ultimately r equires demon stration through its implementatio n and rigorous evalu ation. 5.1 Pa th of Implem entation The C.A.P . can be implemented as a standalone Python li brary or an API service, so that it can encapsulate calls to the underlying LLMs (e.g., the GPT -4 API). (1) Storage of Conversation History : A simp le in -memory queue can be u sed to this end. V ector datab ases (e.g., Pinecone, Chrom a) can store embeddings of historical conversations for efficient retrieval f or applications t hat require persistence. (2) Semantic Ex pansion : This is achieved by sending carefully crafted metaprompts to th e same LLM or ano ther, smaller LLM. (3) Vectorization : Efficient models of sent ence embedd ing, like all -MiniLM-L6-v2, are recommen ded for this task o wing to their balan ced performance and speed. (4) P arameter Tuning : Key parameters of the framework, li ke the co efficient of time decay τ and threshold of alignment  , require t uning through experiments o n benchmark datasets. Setting t he value o f  is particularly critical because too high a value can cause excessive clarification such that this impairs th e fluency of th e LLM, while t oo l ow a value reduces the effectiveness of its “alert” function. 5.2 Meth ods of Evaluation Evaluating the C.A.P .'s eff ectiveness requ ires a multi-dimensional framework that combines quantitative and qualitative metrics. A/B Testin g : This serves as t he core method of evaluati on. Recrui t a grou p of users to complete a series of complex, multi-round collab orative tasks (e.g., jointly developi ng a business plan, writing a sh ort story) by using two versions of the system. Control group: Users interact directly wi th the base LLM. Experimental group: Users interact wit h th e LLM in tegrated with th e C.A.P. Quantitative metrics: Task S uccess Rate : It measures the extent to which the user completes predefined tasks. Dialogic Efficiency : It is the total n umber of rounds or total time requi red to complete t he task. We anticipate that the C.A.P. may requ ire more rounds (due to clarifications) but can reduce the to tal time wasted due to misunderstandings. Frequency of Clarification : It i s the number of times that the C.A.P. triggers the clarification protocol. 8 User Satisfaction : It is assessed by using standardized q uestionnaires, such as the System Usability Scale (SUS) [10] or the PARADISE framework [11], to evaluate the users' subjective perceptions of the quality of interaction, and the in telligence and co operativeness of th e system. Qualitative Metrics: Conversation Analysis : It involves qualitati vely coding tran scribed dialogic texts to an alyze the occurrence of “catastrophic misunderstand ings”—instances where users express frustration—and moments reflectin g “deep collaboration.” Post-task Interview : It i nvolves conducting semi-structu red interviews with t he users to gain insights in to t heir perception of d ifferences b etween t he systems in terms of “understanding,” “sense of cooperation,” and “trust.” W e ant icipate that a system that i ncorporates t he C.A.P . will significantly outperform the baselin e system on key metrics, including t he rate of task success, user satisfactio n, and “sense of collaboration.” 6 DISCUSSION AND LIMIT A TIONS The C.A.P . framework offers a promising path for enhancing the quality of human–computer dialog, b ut its implementation and ap plication remain l imited, an d face several chall enges. First, the compu tational overhead of th e framework requires consideration. Each p reprocessing step in the C.A.P ., particularly the in vocation of the LLM for semantic expansion and vector computations, i ncreases its respo nse latency. Optimizing t he C.A.P .'s efficiency of execu tion with out significantl y impactin g its fluency of interaction remains a critical engineering challenge. Second, p arametric sensitivity—especially the t hreshold of alignment  —is critical to u ser experience. A fixed threshold may fail to accommodate all users and types o f dialogs. Future research should explo re dynamic mechanisms of th reshold adjustment, such as automatically ad apting  based on the domain of the dialog, user expertise, or historical patterns of in teraction. Third, t he design of th e clarificati on protocol requires refinement. Excessively frequ ent or poorly d esigned clarifications may ann oy users with “ over-interruption.” Designing clarifi cation-related interactions that are b oth eff ective and natural is an HCI problem th at requires it erative optimizatio n th rough extensive user research. Finally , t he C.A.P . primarily add resses semantic coherence, and h as a limited cap ability for deeper “alignment” involving emotio ns, values, or complex social dynamics. While it represents an imp ortant starting point for such investigation, it is far from the endpoint of complete human–AI alignment. Despit e t hese l imitations, we think that the design philosophy embod ied by t he C. A.P ., which empowers AI with self-reflectio n and th e in itiative to seek assistance, holds profound value. It enco urages us to reth ink the essence of intelligence, and move beyond the pursuit o f raw performance to prioritize the authenticit y of human–AI collaboration. 7 CONCLUSION In an era of deepening human–LLM int egration, th e quality o f human–machine dialog d irectly d etermines t he upper limit of co llaborative creatio n. This paper has con sidered t he pervasive issu e o f “context misalignment” in long-term conversations involving LLMs, and h as proposed a computational framework to solve the problem called the “Context Alignment P reprocessor” (C.A.P .). By introducing t hree co re processes—semantic expansion, time-weighted context recall, and alignment verification — 9 prior to text generation, the C.A.P . endows AI with the unprecedented capability to activel y assess its own understanding of u ser intent and, upon detectin g p otential misalignment, request its h uman p artner to jointly calibrate the in teraction. This framework represents not merely a techn ical o ptimization, but a profound paradigm shift that en ables h uman– computer interaction to evolve from unidirectional co mmand execution toward a b idirectional, self-repairing collaborative partnership. By grounding C.A.P . in robust theories o f cognitive science and human–computer interactio n, an d clearly o utlining pathways for its implementation and evaluation, C.A.P . provides a sound foundation for building smarter, more reliable, and more t rustworthy next-generation in teractive AI systems. Fu ture work in the area should focus on th e engineering implementation o f this framework, large-scale user stud ies, and contin ual optimizatio n of i ts cor e algorith ms. The ultimate goal i s to ensu re that every conversation between humans and AI becomes a truly meaningful resonance of ideas. 8 REFERENCES [1] Ashish V aswani, Noam Shazee r, Niki Parmar, Jakob Us zkoreit, Llion Jones, Aidan N. G omez, Łuk asz Kais er, and Illia Polosuk hin. 2017. Attention is all you nee d. In Proceed ings of the 31st Internat ional Conference on Neural Informat ion Pro cessing Systems (NIPS 2017) . C urran Associa tes, Inc., Red Hook, NY , USA, 5998– 6008. [2] T om B. Brown, Benj amin Mann, Nic k Ryder , Melanie Subbia h, Jared Ka plan, Prafulla Dhariwal, A rvind Neelaka ntan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-V oss, Gretchen Kruege r, To m Henigha n, Re won Child, Aditya R amesh, Da niel M. Ziegler, Jeffrey Wu, Clemens Winter, Christophe r H esse, Ma rk Chen, Eric Sigler, Mateus z Litwin, Scott Gray, Benjamin Chess , Jac k Clark, Christopher Berner, Sam McCandlish , Alec Radford, Ilya Sutskever, and Dario Amodei. 2020 . Lan guage models are few-shot lea rners. In Advances in Neural Information Process ing Systems, V ol. 33. Curran Associates, Inc., Red Hook, NY , USA, 1877– 1901. [3] Herbert H. C lark and Susan E. Brennan. 1991. Grounding in communic ation. In Perspec tives on Socially Shared Cognition. Lauren B. Resnick, John M. Levine, and Stephanie D. Te asley (Eds.). American Ps ychological Ass ociation, Was hington, DC, USA, 127 –149. https:// doi.org/10.1037/100 96- 006 [4] Long Ouyang, Je ff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wai nwright , Pa mela M ishkin, C hong Zhang, Sandhini Agarwal, Katarina Slama , Alex Ray, J ohn Schulma n, Jacob Hilton, Fraser Kelton, Luke M iller, Maddie Simens, Amanda As kell, Peter Wel ind er, Paul Christia no, Jan Leike, and Ryan Lowe. 2 022. Training langua ge models to follow ins tructions with huma n fe edback. arXiv:2203.0 2155. Retrieved from https://arxiv.org /abs/2203. 02155 [5] Nelson F . Liu, Kevin Lin, John Hew itt, Ashwin Paranjape , Michele Bevilacqua , Fabio Petroni, and Percy Liang. 2023. Lost in the middle: How language mode ls use long contexts. arXiv:2 307.03172 . Retrieved from https://ar xiv. org/abs/2307.03172 [6] Matthew Henderson, Blais e Thomson, and Jason D. Williams. 2014. The second dialog state tracking challe nge. In Procee dings of the 15th Annual Meeting of the Special Inter est Group on Discours e and Dialogue (SIGDIAL 2014). Association for Computa tional Linguistics , Philad elphia, P A, USA, 263–272. https://doi. org/10.3115/v1/W14-4337 [7] Michael Heck, Carel van Niekerk, Nurul Lubis, Christia n Geishauser, Hsien-chin Lin, M arco Moresi, a nd Milica Ga šić. 2020. Tri ppy: A Triple-Copy Strategy for V alue-Independen t Neural Dialog State Tracking. arXiv:2 005.02877. R etrieved from [8] David R. Traum . 1999. C omputational models of grounding in c ollaborative systems. In Wo rking Notes of the AAAI Fall Sympo sium on Psychologica l Models of Communicatio n in Collabor ative Systems . AAAI Pres s, North Falmout h, MA, USA, 124–131. [9] Nils Re imers and Iryna Gurevych. 2019. Sentenc e-BERT : Sentenc e embeddings using siamese BE RT -networks. In Proceedings of the 2019 Conference on Empirical M ethods in N atural Language Proce ssing and the 9th Internat ional Joint Conference on Natural Langua ge Processing (EMNLP-IJCNLP ). Assoc iation for Computat ional Linguis tics, Hong Kong, China , 3982–3992. https://doi.org/10. 18653/v1/D19 -1410 [10] John Brooke. 1996. SUS-A quick and dirty usability s cale. Usability Evaluation In Industry. Patrick W . J ordan, B. Thoma s, Ian Lyall McClelland, Bernard W eerdmeester (Eds. ). Tay lor & Francis , London, UK, 189– 194. [11] Marilyn A. W alker, Diane J . Litman, Candace A. Kamm, and Alicia Abella. 1997. P ARADISE: A framework for evalua ting spoken dialogue a gents. In Pro ceedings of the 35th Annua l Meeting of the Association for Computational Linguistics. Association for Computatio nal Linguistics, M adrid, Spain, 271–280. https://doi. org/10.3115/976909.979652