Unbounded Revision Dialogue

• The paper introduces a framework where dialogue states, goals, and memory can be revised indefinitely, enabling iterative corrections across turns.
• It applies to task-oriented dialogue, ontology revision, and prompting-based reasoning, using methods like dataflow-graph semantics and program stacks.
• Empirical evaluations demonstrate improved accuracy and adaptability, though they reveal trade-offs in stability and complexity that require robust revision protocols.

Unbounded Revision Dialogue denotes an interactional regime in which previously produced goals, computations, beliefs, answers, or memory states remain revisable across arbitrarily many turns rather than being overwritten in a fixed slot inventory or frozen after a single pass. In task-oriented dialogue, it is the user’s ability to “correct, refine, or modify previously specified goals or computations at any point, across arbitrarily many turns, and without being constrained to a fixed set of slots or a pre-scripted workflow” (Machines et al., 2020). In ontology revision, it is an ongoing interaction in which a receiver ontology $K$ is updated step by step by triggers $\varphi_1,\varphi_2,\ldots$, with Preservation and Reconstruction ensuring that earlier knowledge is not lost under indefinite sequences (Özçep, 2016). In prompting-based reasoning, it is the requirement that a reasoner “continuously expose, question, and refine its own work under external pressure until there are no unresolved concerns,” with iteration depth varying from a single pass to potentially unbounded cycles (Zhang et al., 2 Oct 2025).

1. Conceptual scope and the bounded–unbounded distinction

In task-oriented dialogue, the central motivation is that users often revise not only entities but “entire prior computations.” Examples include replacing an event name in an earlier computation (“Sorry, I meant all-hands”) or altering an upstream date constraint that changes a downstream computation such as dayOfWeek (“What about in 2021?”). Traditional slot-filling and finite-state approaches struggle because they represent state as a finite set of slots keyed to a single current intent, lack explicit compositional structure, and rely on scripted policies for clarification and recovery (Machines et al., 2020).

In ontology revision, the same theme appears as iterated reinterpretation under symbol ambiguity. The receiver ontology $K$ is updated by a sequence of sender triggers, and consistency is maintained by separating receiver and sender meanings via reinterpreted signatures and reconnecting them through bridging axioms. The adequacy properties called Preservation and Reconstruction are explicitly presented as crucial for indefinite sequences (Özçep, 2016).

In prompting-based reasoning, the design principle is externalized as “objection-led revision.” A Defender remains the sole author of the answer, an Objectioner asks only question-style objections, and a Host enforces consistency and closure. The protocol is “open-ended by design”: controlled experiments cap the number of rounds, but the intended pattern allows repetition until a closure condition is met or a budget is exhausted (Zhang et al., 2 Oct 2025).

Bounded variants are defined by explicit stopping or commitment mechanisms. In Read, Revise, Repeat (R3), iterative revision stops when either no new edits are proposed, $E_t=\varnothing$, or a predefined maximum revision depth $T=t_{\max}$ is reached; in the user studies, $t_{\max}=3$ (Du et al., 2022). In incremental sequence labelling, bounded revision policies restrict the editable region $S^{(t)}$, the number of edits per position, or the time after which a position is committed, whereas unbounded revision sets $S^{(t)}=\{1,\ldots,t-1\}$ and imposes no explicit limit on revision timing or frequency (Madureira et al., 2023).

2. State representations for revisable interaction

A prominent realization of unbounded revision is the dataflow-graph semantics for task-oriented dialogue. At turn $t$, dialogue state is a growing acyclic graph $G_t=(V_t,E_t,r_t,res_t)$, where $\varphi_1,\varphi_2,\ldots$0 contains functions, constructors, constants, and exceptions; $\varphi_1,\varphi_2,\ldots$1 are dependency edges; $\varphi_1,\varphi_2,\ldots$2 is the returned node; and $\varphi_1,\varphi_2,\ldots$3 is a partial result-edge function. Evaluation proceeds in topological order over $\varphi_1,\varphi_2,\ldots$4, and node values are defined by $\varphi_1,\varphi_2,\ldots$5. Each user turn is mapped to a program $\varphi_1,\varphi_2,\ldots$6 that nondestructively extends $\varphi_1,\varphi_2,\ldots$7 to $\varphi_1,\varphi_2,\ldots$8. The metacomputation operators refer and revise make prior computations directly reusable: refer(C) retrieves a salient node satisfying a constraint, while revise(rootLoc, oldLoc, new) constructs a modified copy of a salient prior computation, sharing nodes where possible. reviseConstraint performs structural, non-destructive transformation of constraints, including weakening when necessary to remain satisfiable (Machines et al., 2020).

Dialog2API adopts a different but related state model: a stack of programs with attached execution results. The dialogue state at turn $\varphi_1,\varphi_2,\ldots$9 is $K$0, with the most recently created or revised program on top. Each program has a goal ID, a status in $K$1, and executable code. Drafting programs may contain placeholders such as ?1, may be revised even when they are not on top of the stack, and move to the top when revised; final programs are executed immediately and become immutable; abandoned programs encode dropped goals. This stack discipline operationalizes “unrestricted program revision” while keeping execution outcomes explicit in state (Shu et al., 2022).

Long-horizon dialogue generation introduces a third representation: revisable per-thread compression states stored in compact dialogue memory. Context-Driven Incremental Compression (C-DIC) treats a conversation as interleaved contextual threads and maintains a dialogue memory $K$2 of latent slot states $K$3. At each turn, it retrieves relevant slots using semantic similarity with mild recency decay, generates the response from the retrieved latents and the current query, compresses the updated turn into a new latent $K$4, and then applies a revise-or-insert write-back rule: $K$5

$K$6

This “retrieve $K$7 revise $K$8 write-back” loop makes stale thread memories revisable across hundreds of turns without full-history re-encoding (Jung et al., 10 Jun 2026).

3. Logical and epistemic foundations

Iterated ontology revision by reinterpretation gives Unbounded Revision Dialogue a model-theoretic form. Ontologies are triples $K$9, where $E_t=\varnothing$0 is the public signature and $E_t=\varnothing$1 is a disjoint internal signature. Ambiguity-compliant substitutions internalize conflicting public symbols, and minimal conflicting symbol sets

$E_t=\varnothing$2

identify the smallest sets of public symbols whose reinterpretation restores consistency. Weak and strong reinterpretation operators then combine the internalized receiver ontology, the trusted trigger, and a selected set of maximal compatible bridging axioms. The weak operator is

$E_t=\varnothing$3

and the strong operator replaces $E_t=\varnothing$4 by the stronger family $E_t=\varnothing$5. Preservation states that some internalized copy $E_t=\varnothing$6 survives in the revised result for any finite trigger sequence; Reconstruction states that, after a suitable renaming, both the original receiver ontology and all triggers are embedded in the result. For atomic triggers, reinterpretation satisfies adapted Darwiche–Pearl postulates $E_t=\varnothing$7, while $E_t=\varnothing$8 need not hold; for complex triggers, fulfillment of the iteration postulates is not guaranteed (Özçep, 2016).

A second formalization appears in rational dialogue over large state spaces. Agents $E_t=\varnothing$9 have information partitions $T=t_{\max}$0 over an arbitrary non-empty state space $T=t_{\max}$1, communicate through a stationary directed graph $T=t_{\max}$2, and send messages determined by a common message function $T=t_{\max}$3 satisfying the Sure Thing Principle. The working partition $T=t_{\max}$4 induced by $T=t_{\max}$5 captures the informational content of agent $T=t_{\max}$6’s message. Dialogue unfolds over ordinal time: at successor stages,

$T=t_{\max}$7

and at a limit ordinal $T=t_{\max}$8,

$T=t_{\max}$9

The update for each agent joins the current partition with the working partitions of its senders, and limit-stage joins aggregate all prior refinements. Under like-mindedness, the Sure Thing Principle, strong connectivity, and symmetry, the transfinite dialogue reaches consensus: there exists an ordinal $t_{\max}=3$0 such that all message functions coincide, and if $t_{\max}=3$1 is the least ordinal with $t_{\max}=3$2, then $t_{\max}=3$3, hence $t_{\max}=3$4 when $t_{\max}=3$5 is infinite (Crescenzi, 2021).

These two lines of work differ in ontology language and epistemic substrate, but both treat revision as structure-preserving accumulation rather than destructive replacement. A plausible implication is that unbounded revision is most naturally expressed when the state formalism itself can retain prior structure—via graph fragments, internalized signatures, or ordinally accumulated partitions—rather than compressing history into a single flat record.

4. Revision protocols, control loops, and interaction roles

FOR-Prompting defines an asymmetric revision protocol with explicit roles and turn semantics. Given a user question $t_{\max}=3$6 and an iteration budget $t_{\max}=3$7,

$t_{\max}=3$8

$t_{\max}=3$9

$S^{(t)}$0

$S^{(t)}$1

The Objectioner may probe “logical gaps, missing constraints, counterfactuals, or ambiguities,” but must not propose fixes, hints, intermediate steps, or an alternative answer. The Host may terminate on fixed rounds or under a closure rule requiring that no substantive unresolved objections remain and that the answer be coherent and consistent. The protocol preserves “a single accountable reasoning chain” because the Defender remains the sole author at every round (Zhang et al., 2 Oct 2025).

R3 implements iterative text revision as a human–machine dialogue over edit suggestions. The document state at revision depth $S^{(t)}$2 is $S^{(t)}$3; model-generated edit suggestions are $S^{(t)}$4; user-accepted edits are $S^{(t)}$5; and the update operator is

$S^{(t)}$6

At each depth, the system performs sentence-level edit prediction, edit-intention classification over $S^{(t)}$7, intent-conditioned rewriting, and diff extraction. Humans inspect the resulting atomic edits and accept or reject them. The process is iterative, but the implemented system remains bounded because it stops at $S^{(t)}$8 or at $S^{(t)}$9 (Du et al., 2022).

Incremental sequence labelling generalizes revision control to streaming NLU. At time $S^{(t)}=\{1,\ldots,t-1\}$0, the model outputs a prefix $S^{(t)}=\{1,\ldots,t-1\}$1, and an edit at position $S^{(t)}=\{1,\ldots,t-1\}$2 occurs when

$S^{(t)}=\{1,\ldots,t-1\}$3

A revision occurs when at least one such edit exists: $S^{(t)}=\{1,\ldots,t-1\}$4 The revision policy is $S^{(t)}=\{1,\ldots,t-1\}$5, where $S^{(t)}=\{1,\ldots,t-1\}$6 is the incremental chart. Unbounded revision lets any past position be changed at any future time step; bounded policies impose positional windows, frequency caps, or time limits. The evaluation methodology characterizes revisions by convenience, effectiveness, novelty, local recurrence and oscillation, company and connectedness, distance from the right frontier, definiteness, and whether they are intermediate or final (Madureira et al., 2023).

5. Empirical performance and evaluation regimes

In task-oriented dialogue as dataflow synthesis, explicit metacomputation improves both representability and predictability. On SMCalFlow, full dataflow with explicit metacomputation achieves agent action exact-match accuracy after inlining of Dev $S^{(t)}=\{1,\ldots,t-1\}$7 and Test $S^{(t)}=\{1,\ldots,t-1\}$8, compared with Dev $S^{(t)}=\{1,\ldots,t-1\}$9 and Test $t$0 for a baseline trained on inlined metacomputation. The gains are larger on cross-turn phenomena: on reference turns, Test accuracy rises from $t$1 to $t$2; on revision turns, from $t$3 to $t$4. On MultiWOZ 2.1 test state tracking, the dataflow representation yields Joint Goal $t$5 versus $t$6 for TRADE, dialogue-level exact match $t$7 versus $t$8, and prefix length $t$9 versus $G_t=(V_t,E_t,r_t,res_t)$0 (Machines et al., 2020).

Dialog2API evaluates unrestricted program revision with execution-based metrics. Its AWS S3 dataset contains 100 dialogues with average 10.5 turns per dialogue, average 2.7 goals per dialogue, average 4.9 lines of code per program, and average 1.4 APIs per program. Because multiple programs can have the same effect, the paper introduces Execution Match Ratio (EMR),

$G_t=(V_t,E_t,r_t,res_t)$1

where $G_t=(V_t,E_t,r_t,res_t)$2 is the last turn up to which predicted and gold executions match in environment signature. Under in-context learning, the best reported baseline is Codex Davinci with API documentation plus in-domain examples, reaching EMR $G_t=(V_t,E_t,r_t,res_t)$3; API documentation alone yields $G_t=(V_t,E_t,r_t,res_t)$4, and examples alone $G_t=(V_t,E_t,r_t,res_t)$5 (Shu et al., 2022).

FOR-Prompting reports both factual accuracy and process-quality gains. On GSM8K, single-prompt accuracy is approximately $G_t=(V_t,E_t,r_t,res_t)$6, CoT accuracy approximately $G_t=(V_t,E_t,r_t,res_t)$7, and FOR-Prompting with 3 rounds approximately $G_t=(V_t,E_t,r_t,res_t)$8, yielding “about a 22 percentage point gain over single-prompt on accuracy.” Under a uniform GPT‑4.1 judge, FOR-Prompting also raises coherence from approximately $G_t=(V_t,E_t,r_t,res_t)$9 to $\varphi_1,\varphi_2,\ldots$00 and reasoning from approximately $\varphi_1,\varphi_2,\ldots$01 to $\varphi_1,\varphi_2,\ldots$02 relative to CoT. On Llama‑3.2:1B over the full GSM8K test set of 1,319 items, single-prompt achieves $\varphi_1,\varphi_2,\ldots$03, FOR with 1 round $\varphi_1,\varphi_2,\ldots$04, and FOR with 3 rounds $\varphi_1,\varphi_2,\ldots$05 (Zhang et al., 2 Oct 2025).

R3 evaluates iterative human–machine revision through acceptance rate and final-document quality. The training data comprise 44,270 / 5,152 / 6,226 unique documents for train/validation/test, with average human revision depths of approximately $\varphi_1,\varphi_2,\ldots$06. Empirically, R3’s acceptance rates are “comparable to human-written revisions at depths 1 and 2.” On an ArXiv subset, final documents produced with human interaction outperform system-only revision: SYSTEM-HUMAN reports Avg. depths $\varphi_1,\varphi_2,\ldots$07, #Edits $\varphi_1,\varphi_2,\ldots$08, Quality $\varphi_1,\varphi_2,\ldots$09, whereas SYSTEM-ONLY reports Avg. depths $\varphi_1,\varphi_2,\ldots$10, #Edits $\varphi_1,\varphi_2,\ldots$11, Quality $\varphi_1,\varphi_2,\ldots$12 (Du et al., 2022).

C-DIC evaluates unbounded revision at long horizons through perplexity, overlap metrics, QA accuracy, and latency. On the main table, C-DIC obtains MSC PPL $\varphi_1,\varphi_2,\ldots$13 and REALTALK per-session PPL $\varphi_1,\varphi_2,\ldots$14, while ICAE incremental collapses to MSC PPL $\varphi_1,\varphi_2,\ldots$15 and REALTALK PPL $\varphi_1,\varphi_2,\ldots$16. The paper states that static baselines “rise sharply after 3–4 consecutive compressions” and that, moving from single-turn to multi-turn evaluation, perplexity for static models “explodes by at least ~1900% while C-DIC decreases by 70%.” On REALTALK all-sessions, C-DIC maintains approximately $\varphi_1,\varphi_2,\ldots$17–$\varphi_1,\varphi_2,\ldots$18 seconds per dialogue up to 428 turns, while Full prompting grows with turn cap and becomes OOM beyond 30 turns under the reported setup (Jung et al., 10 Jun 2026).

Incremental sequence-labelling experiments contribute a complementary evaluation perspective: the restart-incremental Transformer recomputes at every step, giving $\varphi_1,\varphi_2,\ldots$19 on NER, POS, and slot filling, but only $\varphi_1,\varphi_2,\ldots$20. TAPIR-LTReviser and TAPIR-TrfReviser greatly reduce recomputation and revision frequency, while many observed revisions remain effective: for TAPIR-TrfReviser, effective revision rates are $\varphi_1,\varphi_2,\ldots$21 on slot filling, $\varphi_1,\varphi_2,\ldots$22 on POS, and $\varphi_1,\varphi_2,\ldots$23 on NER (Madureira et al., 2023).

6. Trade-offs, failure modes, and open directions

The literature repeatedly identifies a trade-off between recoverability and stability. In dataflow dialogue, reference resolution can be brittle with long-range or multiple candidates; revision scope selection through rootLoc and oldLoc can be ambiguous; long dialogues stress attention and copying; and reviseConstraint requires domain-specific logic to remain satisfiable and faithful (Machines et al., 2020). In incremental sequence labelling, unbounded revision “maximizes the system’s ability to correct incremental hypotheses” but increases “volatility, late corrections, and user-facing instability,” including oscillation and flicker; bounded policies improve stability and user experience by restricting position, frequency, or timing (Madureira et al., 2023).

Logical adequacy does not eliminate iteration-level pathology. Reinterpretation operators guarantee Preservation and Reconstruction, but the adapted Darwiche–Pearl postulates are only partially satisfied: for atomic triggers, $\varphi_1,\varphi_2,\ldots$24 hold, whereas $\varphi_1,\varphi_2,\ldots$25 may fail; for complex triggers, failures arise from multiple symbol redefinitions, non-local effects of bridging axioms, and interference between trivial and non-trivial steps (Özçep, 2016). In the rational-dialogue setting, consensus requires strong connectivity, symmetry, and the opportunity to engage in dialogues of transfinite length; if reciprocity or connectedness fails, fixed points with persistent disagreement may remain (Crescenzi, 2021).

Prompting-based and human-in-the-loop systems inherit additional costs. FOR-Prompting increases token and latency overhead relative to single prompts; missed errors remain possible when the Objectioner fails to ask the right questions; and Host bias can distort closure. The paper also notes that symmetric debate or ensemble voting may outperform objection-only revision when the goal is top‑1 accuracy regardless of provenance (Zhang et al., 2 Oct 2025). R3 is explicitly bounded in its current implementation, and deeper revision depths show distribution shift and error compounding; proposed extensions include persistent history $\varphi_1,\varphi_2,\ldots$26, dynamic stopping policies, learned utility ranking for edits, continual or online learning from feedback, conflict detection, and stronger context management for long documents (Du et al., 2022).

Memory-based long-horizon revision introduces a different cluster of risks. C-DIC notes memory growth under frequent topic shifts, reliance on the pretrained compressor, latent privacy and stale content, and mis-threading when semantically close topics collide. The reported mitigations include bounded retrieval, modest recency decay, and simple replacement-based write-back, while future work points toward learned slot eviction or merging, adaptive retrieval thresholds, hierarchical threading, and integration with external retrieval systems (Jung et al., 10 Jun 2026).

Several forward directions recur across these otherwise distinct settings. One is better learned control over what to retrieve and what to revise: salience and constraint-selection models for refer and revise in dataflow dialogue, learned recomputation or revision policies in incremental labelling, and retrieval-aware threading in dialogue memory (Machines et al., 2020). Another is richer closure and recovery: declarative user goals and preferences that persist across turns, Host criteria for “no unresolved objections,” and exception-driven clarification or recovery that treats follow-up answers as ordinary revisions rather than bespoke policies (Zhang et al., 2 Oct 2025). A final common direction is stronger structure-sensitive modeling—graph transformers, program induction with execution-guided decoding, memory-conditioned compressors, and other neuro-symbolic mechanisms—so that revision remains both expressible and predictable as interaction histories lengthen (Jung et al., 10 Jun 2026).