PaperlayerView exact revision on arXiv

Source-linked AI summary

Lost with a Map: Conversational State and Behavioral Reliability in Language Models

Atahan Dokme, Larry Heck

arXiv:2609.33883v1cs.CL

TL;DR

Language models can act from dialogue history without exposing an explicit belief-state object, raising questions about how conversational state is represented, updated, and used. The paper studies those questions and tests a controller that selectively edits model actions; it finds that reliable interaction depends on resolving which available constraints currently apply and ensuring they govern action.

  • Problem

    Language models can act directly from dialogue history without exposing an explicit belief-state object, leaving open how conversational state is represented, updated, and used.

    Source: §1 · p. 1

  • Method

    Across eight models on MultiWOZ and SGD, the paper measures conversational-state readouts and interventions, then tests SAC, which selectively edits Base LM actions using structural readouts and dialogue evidence.

    Source: §1 · p. 1, §3.1 · p. 3, §7 · p. 9

  • Results

    SAC nearly doubles exact-query accuracy and raises task success on every checkpoint without complete state reconstruction or an additional generation call.

    Source: §7 · p. 9

  • Takeaways & Limitations

    Reliable interaction requires resolving which available constraints currently apply and ensuring that the resulting state governs action.

    Source: §7 · p. 9

    The conclusions are limited to the readout interfaces, interventions, operational diagnostics, and schema-aware interaction setting; SAC is evaluated only in the schema-aware task-oriented dialogue settings described.

    Source: §7 · p. 9, ETHICS STATEMENT · p. 10

Abstract

from arXiv · show

Task-oriented dialogue requires maintaining and updating information across turns, yet language models expose no explicit belief-state object. We study how conversational state is represented, updated, and used inside eight instruction-tuned language models from four families on MultiWOZ and SGD. Structure and values separate: which domains, slots, and requests are active is linearly readable just before the model acts, whereas exact values are far more readable where the user stated them. After a user changes a value, both values remain accessible at their mentions, and causal interventions show that both continue to influence the model's action. In natural closed-loop interaction, query failures separate into cases of weak structural support, incorrect value resolution, and failure to deploy otherwise-supported constraints, with targeted interventions producing systematically different repair behavior across these cases. These findings motivate a state-action controller that starts from the base model action and selectively edits it using structural readouts, without requiring a complete predicted belief state as an intermediate representation. On held-out MultiWOZ interaction across five models, it raises the base model exact-query accuracy from .318 to .621 and task success from .272 to .371 at negligible added cost. Overall, reliable interaction requires not only retaining conversational information, but resolving which available constraints currently apply and ensuring that they govern action.

1 INTRODUCTION

The introduction asks how language models represent, update, and use conversational state without an explicit belief-state object. It argues that reliable action depends on selecting currently applicable information and deploying it, not merely retaining it.

1 INTRODUCTION

  • Across MultiWOZ and SGD, eight models show accessible structure near the decision site but exact values remain more accessible at their historical mentions.The paper frames this as a state-action problem rather than a clean turn-by-turn overwrite.Source: §1 · pp. 1–2
  • The paper names this pattern append–select–deploy: information accumulates, the model resolves which constraints apply, and the resolved constraints must reach the action.The introduction contrasts this process with assuming that earlier information is overwritten.Source: §1 · p. 1
  • Natural query failures involve weak structural support, incorrect value resolution, or failure to deploy supported constraints, and respond differently to targeted interventions.These profiles connect the representation findings to closed-loop behavior.Source: §1 · p. 2
  • SAC starts from the Base LM query and selectively edits it using structural readouts and visible dialogue evidence, without reconstructing a complete belief state.The controller preserves supported Base LM constraints while proposing targeted edits.Source: §1 · p. 2

2 RELATED WORK

The related work situates the study between explicit dialogue-state tracking and language models that act directly from dialogue history. It also distinguishes decoding information from demonstrating its causal or behavioral use.

2 RELATED WORK

  • Earlier task-oriented dialogue systems use explicit belief states, while later language-model approaches integrate tracking, policy, and response generation or elicit slot–value state.This paper instead uses latent state information to selectively correct executable actions rather than reconstructing a complete belief state on every inference path.Source: §2 · p. 2
  • Prior probing and retrieval work studies decodability, but decodability alone does not establish causal or behavioral use; matched interventions test whether source validity changes influence.The controller also differs from activation steering and emitted-program constraints by editing an executable action using a latent structural readout.Source: §2 · p. 2

3 CONVERSATIONAL STATE: SETTING AND ACCESSIBILITY

The section compares state information at decision sites and value-mention sites across eight instruction-tuned models. Its findings distinguish accessible structure and values from the harder problem of identifying which competing value is current.

3.1 SETTING AND MEASUREMENT FRAMEWORK

  • The study evaluates eight instruction-tuned models on MultiWOZ and SGD, comparing frozen readouts at the final pre-generation decision site with readouts at value-mention sites.The reference state includes active domains, constrained slots, requests, and exact values.Source: §3.1 · p. 3

3.2 STRUCTURAL INFORMATION AND EXACT VALUE ACCESSIBILITY

  • Decision-site structural readouts achieve domain F1 .926–.976, constrained-slot F1 .752–.808, and request F1 .662–.802 across the eight models.Rank-32 supervised projections outperform rank-matched PCA and shuffled-supervision controls in every model–task combination.Source: §3.2 · p. 3
  • On the same 3,255 instances, exact-value accuracy is .961–.993 at historical mentions versus .182–.313 at decision sites three turns later.When the slot key is teacher-forced, value accuracy at the key position is .618–.942, indicating that the values become readable once the key is chosen.Source: §3.2 · p. 3

3.3 EARLIER AND UPDATED VALUES REMAIN ACCESSIBLE

  • After SGD updates, old values remain readable at their original mentions with accuracy .868–.898 and new values at theirs with .885–.941.Both values can remain accessible simultaneously; the forced-choice readout’s chance accuracy is .5.Source: §3.3 · p. 3
  • A frozen decision-site readout identifies the currently valid value near chance, while native model behavior improves with scale; the readout interface does not expose all model competence.The figure distinguishes item-level readout accuracy from pair-level behavioral accuracy; nonlinear readouts do no better than linear readouts on four checkpoints. The findings frame current-value resolution as selecting among accessible constraints, not assuming an update erased earlier information.Source: §3.3 · pp. 3–4, Figure 1 · p. 4

4 EARLIER DIALOGUE SOURCES REMAIN CAUSALLY RELEVANT

The section tests whether historical dialogue sources affect behavior and whether their influence changes with conversational validity. It distinguishes source dependence from whether the model selects and uses the currently valid value.

4.1 EARLIER SOURCES CARRY BEHAVIORALLY CONSEQUENTIAL CONTENT

  • Blocking value-source spans lowers correct-value accuracy from .473–.605 to .202–.239, while matched controls remain within .007 of the unmodified model.The disruption centers on value-dependent behavior: tested structural readouts remain largely unchanged.Source: §4.1 · p. 4
  • Replacing an earlier source with its matched counterfactual shifts value preference by +6.55 to +14.22 log-probability units and reverses its sign in every model.Production of the substituted value rises from .085–.123 to .575–.670; random, unrelated-value, and wrong-span substitutions do not reverse the preference.Source: §4.1 · pp. 4–5

4.2 THE SAME SOURCE UNDER CURRENT AND SUPERSEDED STATUS

  • In matched dialogues, the same source has greater influence while current than after supersession: mean Δtarget is −14.66 versus −6.31 across models.The control-adjusted binary effect is positive in all eight models, but the comparison does not identify a distinct internal validity-tracking mechanism.Source: §4.2 · p. 5

4.3 DEPENDENCE ON EARLIER SOURCES CHANGES AFTER AN UPDATE

  • After an update, dependence initially favors the new source, but by four turns the difference is indistinguishable from zero in four models; the old-source effect remains +4.73 to +11.59 log-probability units.In four named models, dependence on the original source is stronger by that point, showing that supersession does not sever behavioral influence.Source: §4.3 · p. 5, Figure 2 · p. 6

5 NATURAL QUERY FAILURES AND THEIR INTERVENTION RESPONSES

Natural query omissions reflect different combinations of weak structural support, incorrect value resolution, and failure to deploy supported constraints. Targeted interventions produce profile-dependent repairs, while masking superseded evidence repairs a subset of stale decisions.

5 NATURAL QUERY FAILURES AND THEIR INTERVENTION RESPONSES

  • The operational profiles distinguish inactive slot structure, supported slots with a competing conditional value, and supported constraints absent from the executable query.Cases without a decisive assignment are AMBIGUOUS; these profiles are diagnostics over measured signals, not claims about separate internal modules or an exhaustive failure taxonomy.Source: §5 · p. 6
  • Qwen-7B-IT’s 687 omissions span 230 structure, 38 binding, 197 deployment, and 222 ambiguous cases, with multiple profiles also present in Llama-8B-IT.The AMBIGUOUS class remains large, while BINDING can be relatively rare on individual models, so binding-specific rates are descriptive.Source: §5 · p. 6
  • Slot forcing repairs deployment-profile omissions more often than structure, binding, or ambiguous omissions in both reported models.Rates are .934, .478, .132, and .554 for Qwen-7B-IT, and .486, .125, .036, and .133 for Llama-8B-IT, respectively; the deployment-minus-structure contrast is positive across six models with runs, but is predictive rather than causal evidence of separate mechanisms.Source: §5 · p. 6
  • Masking the superseded source repairs 17/28 = .607 Qwen-7B-IT errors and 47/121 = .388 Llama-8B-IT errors, versus .07–.11 under random or unrelated-source masking.Masking the currently valid source instead breaks previously correct prefixes, while masking the superseded source rarely does; this oracle-targeted diagnostic shows superseded evidence causally drives a subset of natural stale decisions.Source: §5 · p. 7

6 STATE–ACTION CONTROL

The State–Action Controller selectively edits the Base LM’s query using structural slot readouts and visible dialogue evidence, without constructing a complete belief state. On held-out closed-loop MultiWOZ, it improves query fidelity and task success while adding little inference cost.

6 STATE–ACTION CONTROL

  • SAC starts from the Base LM query, proposes targeted slot edits from a structural readout, and resolves values from visible user evidence rather than carrying a complete predicted slot–value state.Evidence protection, value-resolution rules, and validation safeguard edits; intermediate thresholds retain the Base LM’s decision, and an unparseable query is executed unchanged.Source: §6 · p. 7, §6.1 · p. 7

6.2 CLOSED-LOOP PERFORMANCE

  • SAC-P raises five-checkpoint mean exact-query accuracy from .318 to .621 and task success from .272 to .371, while database equivalence rises from .379 to .640.Task success increases by +.099 (95% CI [+.089,+.111]); the evaluation averages 1,000 goals per model and arm across five checkpoints.Source: Table 1 · p. 8, §6.2 · pp. 8–9

6.1 CONTROLLER AND OPERATING POINTS

  • The intermediate operating point (.70, .50) was selected on development data, and offline results indicate that dialogue-specific readout targeting is needed for SAC’s gains.On Qwen-7B-IT/Llama-8B-IT, SAC reaches .708/.640 exact-query accuracy, versus at most .586/.509 for the listed selector controls; removing illegal slots reaches .445 against SAC’s .621.Source: §6.1 · pp. 7–8

6.2 CLOSED-LOOP PERFORMANCE

  • Against T5-DST, SAC-P matches task success (.371 vs .372) and repair (.401 vs .402), while improving query fidelity and preserving more correct drafts with fewer destructive corrections.SAC-P has +.072 exact-query accuracy, +.065 database equivalence, slot F1 .832 vs .798, preservation .965 vs .654, and .074 vs .650 destructive corrections per goal; it adds no generation. SAC’s advantage over the Base LM persists with rewritten users and on human-authored query-level turns.Source: §6.2 · pp. 8–9
  • SAC adds no model generation and its edit takes 5–7 ms and 8 MB per decision on Qwen-7B-IT/Llama-8B-IT.The comparison excludes methods that alter the deployed backbone’s weights, focusing instead on recovering behavior from a fixed model.Source: §6.2 · p. 9

7 DISCUSSION AND CONCLUSION

The paper concludes that conversational information can remain accessible and behaviorally consequential without reliably governing action, and presents selective editing as a way to improve action without replacing autoregressive generation. Its claims are scoped to the readout interfaces, interventions, operational diagnostics, and schema-aware task-oriented dialogue setting.

7 DISCUSSION AND CONCLUSION

  • SAC uses measured latent structure to target edits while preserving the Base LM’s action otherwise, improving query fidelity and task success without complete state reconstruction or an extra generation.The paper positions SAC as a test of whether interpretability findings support behavioral control, not as a state-of-the-art tracking architecture.Source: §7 · p. 9
  • The conclusions are limited to the readout interfaces, interventions, operational diagnostics, and schema-aware interaction setting.The work studies existing task-oriented dialogue benchmarks and simulated interaction, not a safety mechanism for high-stakes deployment.Source: §7 · p. 9, ETHICS STATEMENT · p. 10

REPRODUCIBILITY STATEMENT

  • The supplement includes model registries, analysis and controller code, frozen manifests with hashes, and evaluation scripts for reproducing the reported analyses and results.The appendix also documents checkpoints, protocols, interventions, and statistical procedures.Source: REPRODUCIBILITY STATEMENT · p. 10

AI USE STATEMENT

  • Generative AI assisted writing revisions, and gpt-4o-mini rewrote user utterances for the naturalized-language evaluation.The authors state they developed and reviewed the research and take responsibility for the final content.Source: AI USE STATEMENT · p. 10

A.1 MODELS

  • Experiments use eight instruction-tuned models from four families, with model-specific frozen analysis layers and no test-population selection of analysis layers or readout hyperparameters.The exact checkpoint revisions are recorded in the reproducibility package.Source: §A.1 · p. 16, Table 2 · p. 16

A.2 DATASETS AND EVALUATION POPULATIONS

The appendix documents the datasets, evaluation populations, protocols, baselines, readouts, and statistical procedures used across the paper. It also reports that structural state has a compact, supervised low-rank component across the evaluated model–task combinations.

A.2 DATASETS AND EVALUATION POPULATIONS

  • Representation analyses use dialogue-disjoint MultiWOZ 2.2 data with 35 slots across seven domains and 46 request keys, alongside SGD with 345,464 turns across 45 services.The update analysis includes 762 eligible events, with a 304-event held-out evaluation population.Source: §A.2 · pp. 16–17

A.3 CONTROLLER TRAINING AND TEST POPULATION REPRODUCIBILITY

  • SAC training and closed-loop testing use a single-executable-domain interface, with 22,083 eligible training decision sites and a frozen 1,000-goal test set.No test selection is used, and none of the 5,166 reference SEARCH decisions in the test set contains constraints from multiple domains.Source: §A.3 · p. 17

A.4 MULTIWOZ CLOSED-LOOP PROTOCOL

  • The closed-loop protocol executes parsed legal-slot queries against the MultiWOZ database and reports task success, exact-query accuracy, and database equivalence.Task success is terminal per-goal success; exact-query accuracy compares against the reference query, while database equivalence compares retrieved entity-name sets.Source: §A.4 · p. 17

A.5 PROMPTED BELIEF-STATE RECONSTRUCTION

  • Prompted DST generates a complete current slot–value state from dialogue history, then canonicalizes and validates it before compiling an executable query.Its few-shot variants prepend retrieved training-turn exemplars while retaining the same state target and compiler.Source: §A.5 · pp. 17–18

A.6 SUPERVISED T5-DST

  • T5-DST is a fully fine-tuned T5-small tracker that generates the complete current state from the same 22,083 training decision sites used for SAC’s structural readout.At inference, its predicted state is deterministically compiled into the shared executable-query format.Source: §A.6 · p. 18

A.7 REPRESENTATION READOUTS

  • The analyses fit linear readouts for structural state and exact values, and use a calibrated decision-layer slot-presence readout for failure profiles and SAC.The readout suite also examines current binding on matched counterfactual dialogues and evaluates paired counterfactual accuracy separately from item-level accuracy.Source: §A.7 · pp. 18–19

A.8 STATISTICAL ANALYSIS

  • Uncertainty intervals use bootstrap resampling aligned to each experiment’s statistical unit, and the five-checkpoint macro is a fixed-panel contrast rather than an estimate over checkpoints.Coverage is reported explicitly, and missing experimental arms are not imputed into macro averages.Source: §A.8 · p. 19, §A.9 · p. 19, Table 3 · p. 20

B.1 STRUCTURAL STATE READOUTS ACROSS EIGHT CHECKPOINTS

  • At rank 32, supervised structural projections outperform rank-matched PCA and shuffled-supervision controls in every evaluated model–task combination, retaining 51–58% of slot-state F1 and 33–59% of request-state F1.They also exceed every random-subspace result for slot and request prediction, while full-dimensional readouts remain stronger.Source: §B.1 · p. 20

B.2 EXACT HISTORICAL VALUES ARE MORE ACCESSIBLE AT THEIR MENTIONS

Across MultiWOZ and SGD, exact historical values are much more readable at their mentions than at later decision sites, while current-value binding is difficult to recover through tested decision-site readouts. The section also tests whether historical values remain behaviorally consequential and whether targeted interventions change value selection.

B.2 EXACT HISTORICAL VALUES ARE MORE ACCESSIBLE AT THEIR MENTIONS

  • Historical-value identity reaches .961–.993 accuracy at mentions but only .182–.313 at decision sites three turns later across eight models.The comparison uses the same 3,255 held-out instances; every paired gap interval excludes zero.Source: §B.2 · p. 21, Table 5 · p. 21

B.3 REPLICATION ON SCHEMA-GUIDED DIALOGUE

  • On SGD update events, cross-site generalization of .786–.878 is consistent with readouts recovering value content rather than only source-position signatures.The earlier position’s persistence cannot itself establish an updating mechanism.Source: §B.3 · pp. 22–23

B.4 COUNTERFACTUAL CURRENT-BINDING READOUTS

  • Decision-site current-binding accuracy ranges from .510 to .591 overall, and .474 to .538 on held-out correction items, so the tested linear readout does not robustly recover current binding.The conclusion is limited to the tested linear interface, not whether current-binding information exists in the model.Source: §B.4 · p. 23, Table 8 · p. 23
  • Nonlinear readouts do not materially improve on the linear readout: none exceeds .56 overall or .58 on recency-incongruent items.Across four evaluated checkpoints, nonlinear overall accuracy stays within .04 of linear accuracy; unmodified model item accuracy on the same items is .434–.580.Source: §B.4 · p. 23, Table 9 · p. 23

B.5 MODEL BEHAVIOR ON THE SAME COUNTERFACTUALS

  • On the same counterfactuals, model pair accuracy reaches .598 and .721 for Qwen-72B-IT and Llama-70B-IT, while Qwen-7B-IT has .027 pair accuracy and .434 item accuracy.Pair accuracy requires both members of a twin to select their respective currently valid values and is stricter than item accuracy.Source: §B.5 · p. 24, Table 10 · p. 24

C.1 BLOCKING HISTORICAL VALUE SOURCES

  • Blocking historical value sources lowers value-selection accuracy from .473–.605 to .202–.239, while matched controls change the unmodified baseline by at most .0067.Structural readout predictions remain highly stable under blocking, with slot-presence agreement of .981–.987.Source: §C.1 · pp. 24–25, Table 11 · p. 24

C.2 SUBSTITUTING HISTORICAL VALUE CONTENT

  • Substituting historical value content reverses the value margin in every checkpoint, with mean shifts of +6.55 to +14.22 and substituted-value production rising from .085–.123 to .575–.670.Controls support content- and location-sensitive influence, while structural readouts remain nearly unchanged and other active-slot values show some collateral effects.Source: §C.2 · p. 25, Table 12 · p. 25

C.3 MATCHED CURRENT-VERSUS-SUPERSEDED SOURCES

Matched interventions show that a historical source’s influence weakens but persists after supersession, while natural omissions fall into distinct structural, value-binding, deployment, or ambiguous profiles.

C.3 MATCHED CURRENT-VERSUS-SUPERSEDED SOURCES

  • Mean Δmask was −14.659 when the source was current and −6.306 after supersession; masking effects remained negative across all eight checkpoints.The paired difference excluded zero in all eight checkpoints, indicating weaker—but not eliminated—influence after supersession.Source: §C.3 · p. 25
  • Other-constraint masking removes another slot’s valid task content, and current-source masking is not an independent replication in the matched twin.The authors therefore use a length-matched random non-value span as the nonspecific control and treat other-constraint masking as secondary.Source: §C.3 · p. 26
  • The control-adjusted validity effect was positive on all eight checkpoints, with every interval excluding zero.It subtracts the effect of masking an equally long random non-value span from the source-masking effect within each validity condition.Source: §C.3 · p. 26, Table 14 · p. 26

C.4 TEMPORAL EVOLUTION OF ORIGINAL-SOURCE INFLUENCE

  • At the update turn, dependence on the original new-value source exceeded dependence on the old source in all eight checkpoints, but the advantage shrank at later lags.By τ = +4, old-source dependence was significantly stronger in four checkpoints; original old-source effects remained substantial, while later restatements stayed visible. Eligible populations changed across lags, so the trajectory is not a fixed-cohort decay estimate.Source: §C.4 · p. 27, Table 15 · p. 27

D.1 OPERATIONAL FAILURE PROFILES

  • Operational profiles distinguish weak structural support, scoring that favors a competing value, supported-but-omitted constraints, and unresolved cases.The value-margin threshold is selected on development data; the profile rule assigns structural inactivity first, then binding, deployment, or ambiguity.Source: §D.1 · p. 28

D.2 PROFILE FREQUENCIES ACROSS CHECKPOINTS

Across checkpoints, omission profiles vary substantially and predict different responses to slot forcing, while targeted masking shows that superseded sources drive a subset of natural stale decisions. The controller’s results depend on learned structural selection, evidence safeguards, and operating-point choices rather than value rules or legality checks alone.

D.2 PROFILE FREQUENCIES ACROSS CHECKPOINTS

  • Across seven checkpoints with complete profiles, DEPLOYMENT ranges from .110 to .551 and AMBIGUOUS from .239 to .474, so profiles are diagnostic rather than exhaustive.Stale-value-only cases are rare in most checkpoints, with 1–29 examples, and are treated descriptively below 20.Source: §D.2 · p. 28

D.3 FORCING AN OMITTED SLOT

  • Slot forcing repairs DEPLOYMENT omissions more often than STRUCTURE omissions in all six evaluated checkpoints, but the contrast is predictive stratification, not independent causal identification.The profile and intervention both condition on the slot key, though they use different query prefixes and generation procedures.Source: §D.3 · p. 29, Table 18 · p. 29

D.4 TARGETED HISTORICAL-SOURCE MASKING ON NATURAL FAILURES

  • On Qwen-7B-IT’s 28-example common paired population, stale-source masking repaired 17/28 (.607), versus 3/28 (.107) for both random and unrelated controls and 0/28 for current-source masking.This is an oracle-targeted, frozen-prefix intervention: oracle annotations identify offending spans, but no oracle value or state is inserted, and the trajectory does not continue in closed loop.Source: §D.4 · pp. 29–30

E.1 EVIDENCE PROTECTION ABLATION

  • Evidence protection removes destructive corrections in both audited development comparisons, from 35 to zero for Qwen-7B-IT and 83 to zero for Llama-8B-IT.Exact-query accuracy and database equivalence improve, but task success is unchanged for Qwen and changes from .206 to .202 for Llama; this is not a test-set ablation.Source: §E.1 · pp. 30–31

E.2 SELECTOR-ATTRIBUTION ABLATION

  • Full SAC exceeds every alternative selector on both checkpoints; disabling structural selection recovers little over the Base LM, and visible-text selection falls below it.Unsupported constraints have no matching user text span, so their absence cannot serve as evidence of inactivity; learned dialogue-specific targeting is consequential.Source: §E.2 · p. 31

E.3 EDIT COMPOSITION AND LEGALITY-ONLY RESCORING

  • Legality-only rescoring raises exact-query accuracy from .318 to .445 and database equivalence from .379 to .552, below the controller’s .621 and .640.The controller’s 21,591 edits include 12,641 deletions; 98.5% target slots absent from the reference state, while 4,556 remove database-illegal values.Source: §E.3 · pp. 31–32, Table 22 · p. 32

E.4 CONTROLLER RESCUE BY NATURAL-OMISSION PROFILE

  • Of 198 pooled BINDING omissions, 26 are rescued, 112 lack a usable value in visible user turns, and 60 receive a different value.The controller shares the calibrated slot readout with the profiles, so it cannot propose STRUCTURE omissions; only 6 of 973 AMBIGUOUS omissions receive one.Source: §E.4 · p. 32, Table 23 · p. 32

E.5 EXACT CONTROLLER PROCEDURE

  • The controller edits the Base LM query using visible user evidence and validation, without persistent complete slot–value state or additional generation; among three settings, 70/50 leads two-checkpoint development means: .7075 exact-query accuracy and .7065 database equivalence.The corresponding means are .6575/.6680 at 90/30 and .6930/.6895 at 50/50; 70/50 is selected for query fidelity, not development task success.Source: §E.5 · p. 33, §E.6 · p. 33

E.7 FIVE-CHECKPOINT CLOSED-LOOP RESULTS

On a five-checkpoint panel, the primary controller improves exact-query accuracy over both the Base LM and Prompted DST at every checkpoint, while task-success comparisons against T5-DST vary by checkpoint. Its operating points trade preservation against repair, and the controller adds less inference work than explicit-state alternatives.

E.7 FIVE-CHECKPOINT CLOSED-LOOP RESULTS

  • The primary controller improves exact-query accuracy over both the Base LM and Prompted DST in every checkpoint, but its task-success differences versus T5-DST are heterogeneous.SAC-P is higher than T5-DST on OLMo-32B-IT and lower on OLMo-7B-IT; intervals include zero for the other three checkpoints.Source: §E.7 · p. 34, Table 25 · p. 34

E.8 PRESERVATION, REPAIR, AND DESTRUCTIVE CORRECTIONS

  • Moving from 90/30 to 50/50 increases repair and destruction; 70/50 pairs higher preservation than 50/50 with higher repair than 90/30.These conditional rates describe deployed interfaces whose trajectories diverge, not an intervention effect on a shared fixed-prefix population.Source: §E.8 · pp. 34–35

E.9 SLOT-LEVEL FIDELITY METRIC

  • Slot-level fidelity counts a slot as correct only when the executable query and reference share it with an equivalent value; a wrong value can create both a false positive and false negative.These slot-level precision, recall, and F1 metrics are distinct from dialogue-level inform precision and recall.Source: §E.9 · p. 35

E.10 INFERENCE REQUIREMENTS, MODEL FOOTPRINT, AND SUPERVISION

  • The controller reuses the decision-layer representation and Base LM query without extra generation, while Prompted DST adds a state-generation call and T5-DST invokes a separate model.The controller’s main advantages are executable-query fidelity and preservation of correct drafts; explicit state reconstruction remains competitive for terminal task success.Source: §E.10 · pp. 35–36

E.11 TRANSFER TO HUMAN-AUTHORED DIALOGUE

A fixed-history replay tests the frozen primary controller on human-authored MultiWOZ dialogue, rather than measuring closed-loop task success. The controller improves query fidelity over its Base LM, especially on turns whose reference values are user-stated, while the one-domain query interface excludes multi-domain turns.

E.11 TRANSFER TO HUMAN-AUTHORED DIALOGUE

  • On 4,490 matched turns from 911 MultiWOZ 2.2 test dialogues, the controller improves exact-query accuracy by +.1575 and database equivalence by +.1623 over its Base LM.Only turns with one constrained executable domain are included, and terminal task success is undefined in this fixed-history replay.Source: §E.11 · p. 36, Table 29 · p. 37
  • Macro slot precision rises from .539 to .766 and recall from .538 to .641, with a larger reduction in unsupported constraints than in missing constraints.Unsupported constraints fall from 1.163 to .533 per turn, while missing constraints fall from 1.251 to .971.Source: §E.11 · p. 37
  • The resolver finds the correct latest value for 8,740 of 12,151 reference values (.719); 2,573 (.212) lack user-side evidence, limiting gains on mixed-evidence turns.The exact-query gain is +.275 when all reference values are user-stated, versus +.038 on mixed-evidence turns and +.079 when values are system-side.Source: §E.11 · p. 37
  • The replay excludes 1,686 of 7,363 user turns with constraints from multiple executable domains because the controller interface emits one domain per query.Source: §E.11 · pp. 36–37

E.12 CLOSED-LOOP EVALUATION WITH NATURALIZED USER LANGUAGE

With naturalized simulator language, the controller’s advantage over the Base LM remains, while the Base LM’s exact-query accuracy degrades more. The appendices specify the query format and a matched construction that isolates how an unchanged historical source’s influence varies with conversational validity.

E.12 CLOSED-LOOP EVALUATION WITH NATURALIZED USER LANGUAGE

  • The rewrite experiment preserves user-act semantics with a faithfulness gate, retrying a rewrite once before falling back to the template utterance.Across ten arm–model runs, 51,112 turns required 14,799 API calls and 84 turns (0.16%) used the fallback.Source: §E.12 · p. 37
  • Under naturalized user language, exact-query accuracy falls by .032 for the Base LM and .006 for SAC-P; the latter decline is not significant.Database equivalence improves by +.285 [+.275, +.294] over the Base LM under naturalized language, versus +.260 with template users; these are separate simulator controls.Source: §E.12 · p. 37, Table 30 · p. 38

F.1 EXACT MULTIWOZ PROMPTS AND QUERY FORMATTING

  • The MultiWOZ SEARCH prompt requests one-line domain-and-slot constraints using only what the user has told the model, with a fixed domain and slot vocabulary.Source: §F.1 · p. 38

F.1.1 QUERY SERIALIZATION

  • Queries are parsed independently of slot order, aliases are canonicalized, and syntactic validity does not guarantee that a value is legal for the domain.For example, a restaurant-area value can parse successfully yet fall outside the vocabulary and return no entities.Source: §F.1.1 · pp. 38–39

F.2 MATCHED HISTORICAL-SOURCE CONSTRUCTION

  • Matched historical-source twins keep the original value text and token positions fixed while changing only the final update instruction, isolating the source’s validity.Source: §F.2 · p. 39

Related papers

Loading 2609.33883v1…