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

Atahan Dokme; Larry Heck

cs.CLSep 27, 2026

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

Authors: Atahan Dokme, Larry Heck

Organizations: Georgia Institute of Technology

Abstract

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.

Figures & tables

Method	TS	Exact	DB-eq	Slot F1	Preserv.	Repair	Destr. $\downarrow$	Extra calls $\downarrow$
Base LM	.272	.318	.379	.626	1.000	.000	.000	0
Prompted DST (0-shot)	.328	.353	.445	.701	.503	.209	.731	5.02
Prompted DST ( $k{=}5$ )	.348	.384	.493	.727	.522	.238	.767	5.04
Prompted DST ( $k{=}10$ )	.363	.393	.504	.730	.523	.251	.768	4.99
T5-DST	.372	.549	.575	.798	.654	.402	.650	4.98
SAC-C	.349	.534	.569	.784	.995	.268	.010	0

Table 1: Held-out closed-loop MultiWOZ evaluation, five-checkpoint mean over Qwen-7B-IT, Llama-8B-IT, Mistral-7B-IT, OLMo-7B-IT, and OLMo-32B-IT,

1{,}000

goals per model and arm (per-checkpoint TS/Exact values in Table 25 ). TS: terminal task success. Exact, DB-eq: exact-query accuracy and database equivalence of the executed query, averaged over each goal’s SEARCH turns. Slot F1: slot-level fidelity of the executed query against the reference state. Preserv.: fraction of already-correct Base LM drafts that remain correct after the arm acts. Repair: fraction of incorrect Base LM drafts that are corrected. Destr.: destructive corrections per goal. Extra calls: model generations per goal beyond the Base LM’s own query.

\downarrow

: lower is better.

Appendix figures & tables29 assets

Supplementary material from the paper’s appendix.

Appendix

Model	Abbrev.	Hugging Face checkpoint	Layers	Analysis layer
Qwen-7B-IT	Q7	Qwen/Qwen2.5-7B-Instruct	28	20
Qwen-72B-IT	Q72	Qwen/Qwen2.5-72B-Instruct	80	57
Llama-8B-IT	L8	meta-llama/Llama-3.1-8B-Instruct	32	23
Llama-70B-IT	L70	meta-llama/Llama-3.1-70B-Instruct	80	57
Mistral-7B-IT	M7	mistralai/Mistral-7B-Instruct-v0.3	32	23
Mistral-24B-IT	M24	mistralai/Mistral-Small-24B-Instruct-2501	40	28

Appendix

Table 2: Instruction-tuned models used in the paper. Families: Qwen2.5 ( Qwen Team et al., 2025 ) , Llama 3.1 ( Grattafiori et al., 2024 ) , Mistral ( Jiang et al., 2023 ) , and OLMo 2 ( OLMo et al., 2025 ) . Model gives the name used in the text and Abbrev. the label used in the figures. The final column gives the frozen model-specific layer used as the primary analysis layer. Exact checkpoint revisions corresponding to the executed runs are recorded in the reproducibility package.

Model	Readouts	Causal	Profiles	Forcing	Closed loop
Qwen-7B-IT	Yes	Yes	Yes	Yes	Yes
Llama-8B-IT	Yes	Yes	Yes	Yes	Yes
Mistral-7B-IT	Yes	Yes	Yes	Yes	Yes
OLMo-7B-IT	Yes	Yes	Yes	Yes	Yes
OLMo-32B-IT	Yes	Yes	Yes	Yes	Yes
Mistral-24B-IT	Yes	Yes	No	No	No

Appendix

Table 3: Model coverage across experimental components. Columns: representation readouts, causal source analyses, natural-failure profiles, slot forcing, and the complete closed-loop comparison, which requires the Base LM, Prompted DST, T5-DST, and all three State–Action Controller operating points. “Partial” indicates that some of these arms are unavailable; no missing arm is imputed or included in the five-model macro averages.

Model	Supervised	PCA	Shuffled	Full	Random mean [min,max]
Qwen-7B-IT	.425	.303	.265	.794	.125 [.095,.161]
Llama-8B-IT	.449	.304	.286	.775	.151 [.140,.167]
Mistral-7B-IT	.408	.247	.229	.802	.117 [.108,.129]
OLMo-7B-IT	.445	.300	.283	.808	.163 [.141,.197]
Mistral-24B-IT	.394	.251	.220	.752	.113 [.101,.129]
OLMo-32B-IT	.444	.300	.278	.789	.165 [.148,.180]

Appendix

Table 4: Structural state accessibility across eight checkpoints. All entries are micro-F1 on the MultiWOZ development split at the fixed analysis layer. Supervised, PCA, and shuffled-supervision readouts use rank 32; Full uses the full-dimensional representation. Random reports the mean over eight rank-32 random subspaces; ranges are included where available.

Model	Layer	Mention	Last	Decision site ( $k=3$ )	Gap	95% CI
Qwen-7B-IT	20	.9696	.9241	.1819	$+.7877$	$[+.773,+.802]$
Llama-8B-IT	23	.9788	.9730	.2399	$+.7389$	$[+.722,+.756]$
Mistral-7B-IT	23	.9920	.9429	.1837	$+.8083$	$[+.794,+.823]$
OLMo-7B-IT	23	.9871	.9548	.2147	$+.7724$	$[+.757,+.787]$
Mistral-24B-IT	28	.9926	.9806	.1862	$+.8065$	$[+.792,+.821]$
OLMo-32B-IT	45	.9849	.9662	.2055	$+.7794$	$[+.765,+.794]$

Appendix

Table 5: Paired exact-value localization at frozen analysis layers. The same

3{,}255

held-out instances are scored at both positions, joined by dialogue, slot, and canonical value. Mention uses the mean-pooled historical value span and Last its final token; Decision site reads the representation at the decision position, which occurs exactly three turns after the mention (

k=3

). Gap is Mention

-

Decision site , with a dialogue-level bootstrap (

2{,}000

resamples,

922

dialogues).

Model	Decision site	Emission position	Emission $-$ decision (non-default)
Qwen-7B-IT	.295 / .182	.708 / .618	$+.436$ $[+.417,+.453]$
Llama-8B-IT	.363 / .240	.926 / .908	$+.668$ $[+.651,+.685]$
Mistral-7B-IT	.308 / .184	.903 / .872	$+.688$ $[+.672,+.705]$
OLMo-7B-IT	.334 / .215	.900 / .873	$+.658$ $[+.638,+.676]$
Mistral-24B-IT	.299 / .186	.901 / .872	$+.686$ $[+.669,+.702]$
OLMo-32B-IT	.328 / .206	.839 / .786	$+.580$ $[+.562,+.597]$

Appendix

Table 6: Exact-value probe at the decision site and the emission position on identical instances. All / non-default accuracy on the

4{,}887

3{,}255

held-out instances of Table 5 , same probe, layer, and training discipline. The last column is the paired non-default difference with a dialogue-level bootstrap (

2{,}000

resamples); every interval excludes zero.

Model	Old source	New source	Decision-site current binding
Qwen-7B-IT	.868 [.829,.906]	.888 [.853,.924]	.569 [.515,.625]
Llama-8B-IT	.888 [.852,.923]	.918 [.886,.948]	.477 [.422,.537]
Mistral-7B-IT	.885 [.848,.920]	.905 [.870,.938]	.507 [.448,.564]
OLMo-7B-IT	.898 [.862,.933]	.911 [.878,.942]	.507 [.446,.563]
Mistral-24B-IT	.895 [.860,.927]	.941 [.913,.965]	.566 [.514,.622]
OLMo-32B-IT	.888 [.854,.922]	.885 [.846,.919]	.480 [.423,.538]

Appendix

Table 7: Historical-value accessibility on SGD slot-update events. Forced-choice accuracies on 304 events from 258 held-out dialogues, with 95% bootstrap intervals. Old and new source columns measure identification of the value at the corresponding mention. Decision-site current binding is the fraction of events on which the readout favors the updated value, not its mean predicted probability.

Model	L	Rank	WD	Overall	Congruent	Incongruent	Correction	Corr. $\times$ Incong.
Qwen-7B-IT	20	8	$10^{-4}$	.510 [.479,.542]	.527 [.482,.571]	.494 [.449,.538]	.488 [.436,.541]	.474 [.400,.548]
Llama-8B-IT	23	16	$10^{-4}$	.557 [.526,.588]	.538 [.493,.582]	.577 [.532,.620]	.474 [.421,.527]	.456 [.383,.531]
Mistral-7B-IT	23	2	$10^{-4}$	.534 [.503,.566]	.429 [.386,.474]	.640 [.596,.681]	.503 [.450,.556]	.602 [.528,.673]
OLMo-7B-IT	23	16	$10^{-4}$	.532 [.501,.564]	.494 [.449,.538]	.571 [.526,.614]	.506 [.453,.558]	.503 [.429,.577]
Mistral-24B-IT	28	512	$10^{-3}$	.517 [.485,.548]	.544 [.499,.588]	.490 [.445,.534]	.500 [.447,.553]	.515 [.440,.588]
OLMo-32B-IT	45	8	$10^{-3}$	.591 [.559,.621]	.556 [.512,.600]	.625 [.581,.667]	.515 [.462,.567]	.550 [.475,.622]

Appendix

Table 8: Decision-site current-binding readout. Item-level accuracy with Wilson 95% intervals. Analysis layer, rank, and weight decay (WD) are fixed on development data before test evaluation. The full test has 960 items, each recency stratum has 480, the held-out correction family has 342, and its recency-incongruent subset has 171.

Model	All	Incong.	All	Incong.	Width	All	Incong.	Width
	Linear		MLP-H			MLP-J
Qwen-7B-IT	.510	.494	.503	.500	64	.498	.496	256
Llama-8B-IT	.557	.577	.559	.537	256	.523	.537	1024
Mistral-7B-IT	.534	.640	.494	.502	64	.528	.571	1024
OLMo-7B-IT	.532	.571	.500	.500	64	.514	.515	64

Appendix

Table 9: Linear versus nonlinear current-binding readouts on the same 960 test items. Overall and recency-incongruent item-level accuracy; the linear readout is the frozen one of Table 8 , scored on identically extracted states. Width is the development-selected hidden size.

Model	Pair acc. [95% CI]	Item acc.	RD $\checkmark$ /MB $\checkmark$	RD $\checkmark$ /MB $\times$	RD $\times$ /MB $\checkmark$	RD $\times$ /MB $\times$	$P(\mathrm{MB}\checkmark\mid\mathrm{RD}\times)$
Qwen-7B-IT	.027 [.015,.042]	.434	227	263	190	280	.404
Llama-8B-IT	.210 [.173,.248]	.463	277	258	167	258	.393
Mistral-7B-IT	.258 [.219,.298]	.580	273	240	284	163	.635
OLMo-7B-IT	.090 [.065,.115]	.467	243	268	205	244	.457
Mistral-24B-IT	.588 [.542,.629]	.784	394	102	359	105	.774
OLMo-32B-IT	.410 [.367,.456]	.689	407	160	254	139	.646

Appendix

Table 10: Model behavior and readout correctness on the same counterfactual items. Pair accuracy (480 pairs) requires both members to select their current values; item accuracy is over the 960 items. RD denotes the fixed decision-site readout and MB the unmodified model’s current-value behavior; counts cross-tabulate item-level correctness.

Model	$n$	Dial.	Unmodified	Blocked	vs. same turn	vs. other turn	vs. random
Qwen-7B-IT	1187	533	.473	.202	-.269 [-.298,-.239]	-.270 [-.299,-.241]	-.276 [-.306,-.247]
Llama-8B-IT	1194	534	.569	.225	-.343 [-.375,-.311]	-.346 [-.379,-.317]	-.343 [-.374,-.313]
Mistral-7B-IT	1186	532	.498	.206	-.292 [-.321,-.261]	-.293 [-.325,-.263]	-.295 [-.325,-.266]
OLMo-7B-IT	1192	533	.534	.205	-.323 [-.352,-.293]	-.326 [-.358,-.296]	-.325 [-.358,-.293]
Mistral-24B-IT	1186	531	.605	.239	-.358 [-.388,-.327]	-.365 [-.394,-.334]	-.362 [-.392,-.330]
OLMo-32B-IT	1195	534	.577	.234	-.341 [-.374,-.308]	-.346 [-.379,-.315]	-.342 [-.374,-.312]

Appendix

Table 11: Broad blocking of historical value sources. Unmodified is value-selection accuracy without intervention. The final three columns report broad-block accuracy minus the corresponding matched-control accuracy, with 95% bootstrap intervals.

Model	$n$	Dial.	Unmod.	Patched	Shift [95% CI]	Says $v_{b}$	Says $v_{a}$
Qwen-7B-IT	798	424	-4.170	+4.384	+8.55 [+7.95,+9.16]	.123 $\rightarrow$ .600	.526 $\rightarrow$ .089
Llama-8B-IT	836	443	-4.914	+3.798	+8.71 [+8.13,+9.27]	.085 $\rightarrow$ .624	.620 $\rightarrow$ .086
Mistral-7B-IT	678	390	-4.972	+4.426	+9.40 [+8.68,+10.11]	.118 $\rightarrow$ .575	.544 $\rightarrow$ .102
OLMo-7B-IT	838	444	-4.943	+4.088	+9.03 [+8.44,+9.62]	.099 $\rightarrow$ .619	.585 $\rightarrow$ .086
Mistral-24B-IT	685	389	-3.465	+3.083	+6.55 [+6.04,+7.04]	.114 $\rightarrow$ .666	.634 $\rightarrow$ .091
OLMo-32B-IT	839	444	-5.962	+4.723	+10.69 [+9.99,+11.40]	.094 $\rightarrow$ .614	.616 $\rightarrow$ .113

Appendix

Table 12: Historical-value substitution. Margin is

\log P(v_{b})-\log P(v_{a})

. The final two columns report the fraction of generations that express the substituted and original value, respectively, before and after patching.

Model	Current $\Delta_{\mathrm{mask}}$	Superseded $\Delta_{\mathrm{mask}}$	Superseded $-$ Current
Qwen-7B-IT	-19.009 [-19.34,-18.67]	-6.591 [-7.16,-6.01]	+12.418 [+11.82,+12.99]
Llama-8B-IT	-10.458 [-10.91,-9.98]	-5.163 [-5.40,-4.92]	+5.295 [+4.87,+5.70]
Mistral-7B-IT	-14.092 [-14.72,-13.47]	-7.154 [-7.75,-6.57]	+6.938 [+6.22,+7.72]
OLMo-7B-IT	-18.982 [-19.58,-18.38]	-11.034 [-11.79,-10.29]	+7.948 [+7.30,+8.64]
Mistral-24B-IT	-9.015 [-9.48,-8.57]	-6.275 [-6.50,-6.05]	+2.739 [+2.33,+3.16]
OLMo-32B-IT	-12.111 [-12.64,-11.58]	-4.371 [-4.74,-4.06]	+7.741 [+7.19,+8.32]

Appendix

Table 13: Matched current-versus-superseded source influence. The same historical source is current in one twin and superseded in the other. The paired difference is Superseded

-

Current, so positive values indicate stronger causal influence while the source is current.

Model	$D_{\mathrm{validity}}$	95% CI
Qwen-7B-IT	$+.8598$	$[+.8084,+.9065]$
Llama-8B-IT	$+.5607$	$[+.4907,+.6308]$
Mistral-7B-IT	$+.6121$	$[+.5421,+.6822]$
OLMo-7B-IT	$+.0514$	$[+.0187,+.0888]$
Mistral-24B-IT	$+.6916$	$[+.6262,+.7523]$
OLMo-32B-IT	$+.7009$	$[+.6355,+.7617]$

Appendix

Table 14: Control-adjusted current-versus-superseded effect on the binary current-value decision. Difference-in-differences against a token-length-matched random non-value span, on the same

214

matched twin pairs per checkpoint. Intervals are dialogue-level bootstraps (

2{,}000

resamples, seed 954100); the twin pair is the dialogue.

Model	$\tau=0$	$+1$	$+2$	$+3$	$+4$
Qwen-7B-IT	+16.62 [+14.5,+17.7]	+1.50 [+.2,+2.7]	+.85 [-.3,+2.0]	+.00 [-1.2,+1.3]	-.32 [-1.9,+1.1]
Llama-8B-IT	+8.58 [+7.8,+9.4]	-.23 [-.8,+.4]	-.85 [-1.5,-.2]	-1.13 [-1.8,-.4]	-1.72 [-2.6,-.9]
Mistral-7B-IT	+14.33 [+13.0,+15.6]	+.60 [-.3,+1.5]	-.76 [-1.6,+.1]	-1.22 [-2.1,-.3]	-1.81 [-2.8,-.8]
OLMo-7B-IT	+8.42 [+7.4,+9.4]	-1.68 [-2.4,-1.0]	-2.02 [-2.8,-1.2]	-2.86 [-3.7,-2.0]	-2.99 [-4.0,-2.0]
Mistral-24B-IT	+9.07 [+8.4,+9.7]	+1.07 [+.6,+1.5]	+.49 [+.0,+1.0]	+.04 [-.5,+.6]	-.14 [-.8,+.5]
OLMo-32B-IT	+8.72 [+7.9,+9.6]	-.14 [-.8,+.5]	-.41 [-1.0,+.2]	-.86 [-1.6,-.2]	-1.06 [-1.8,-.2]

Appendix

Table 15: Relative dependence on original new and old source mentions.

D=C_{\mathrm{new}}-C_{\mathrm{old}}

with 95% bootstrap intervals. Positive values indicate stronger dependence on the newly introduced source.

Model	SEARCH	Omissions	Structure	Binding	Deployment	Ambiguous
Qwen-7B-IT	2379	687	230 (.335)	38 (.055)	197 (.287)	222 (.323)
Llama-8B-IT	2869	365	152 (.416)	28 (.077)	72 (.197)	113 (.310)
Mistral-7B-IT	2631	173	61 (.353)	11 (.064)	19 (.110)	82 (.474)
OLMo-7B-IT	2760	314	58 (.185)	8 (.025)	173 (.551)	75 (.239)
OLMo-32B-IT	2165	133	55 (.414)	5 (.038)	21 (.158)	52 (.391)
Qwen-72B-IT	2527	424	149 (.351)	36 (.085)	131 (.309)	108 (.255)

Appendix

Table 16: Natural omission profiles. Counts are per offending slot; fractions are relative to the omission count for that checkpoint. Raw totals should not be compared across models because the number of SEARCH decisions differs.

Model	Profile	$n$	Goals	Repair [95% CI]	$\Delta$ Exact	$\Delta$ DB
Qwen-7B-IT	Deployment	197	81	.934 [.884,.970]	+.579	+.421
	Structure	230	64	.478 [.271,.672]	+.187	+.070
	Binding	38	20	.132 [.033,.237]	+.053	-.079
	Ambiguous	222	99	.554 [.420,.667]	+.302	+.135
Llama-8B-IT	Deployment	72	46	.486 [.348,.618]	+.333	+.125
	Structure	152	49	.125 [.034,.214]	+.020	-.092

Appendix

Table 17: Slot forcing on Qwen-7B-IT and Llama-8B-IT. Repair requires both insertion of the omitted slot and generation of the current value.

\Delta

Exact and

\Delta

DB are relative to the Base LM query. Collateral damage and new-extra counts are exactly zero in every displayed slot-forcing cell.

Model	Deployment $-$ Structure [95% CI]	$n_{D},n_{S}$
Qwen-7B-IT	+.456 [+.275,+.681]	197, 230
Llama-8B-IT	+.361 [+.224,+.508]	72, 152
Mistral-7B-IT	+.602 [+.330,+.848]	19, 61
OLMo-7B-IT	+.445 [+.337,+.547]	173, 58
OLMo-32B-IT	+.807 [+.560,+.952]	21, 55
Qwen-72B-IT	+.669 [+.478,+.811]	131, 149

Appendix

Table 18: Slot-forcing Deployment–Structure contrast. The 95% interval for the repair-rate difference excludes zero in all six evaluated checkpoints, favoring Deployment over Structure .

Model	Stale source	Current source	Unrelated	Random
Qwen-7B-IT	17/29 (.586)	0/29 (.000)	3/28 (.107)	3/29 (.103)
Llama-8B-IT	47/121 (.388)	1/121 (.008)	9/121 (.074)	8/121 (.066)

Appendix

Table 19: Natural stale-value intervention on the prespecified Qwen-7B-IT and Llama-8B-IT populations. Repair means that the regenerated query emits the current value for the updated slot. Qwen-7B-IT: 17/29 on all stale errors; the main text’s 17/28 uses the common paired population, where one unrelated-control span is unavailable.

Model	Protection	Turns	Success	Exact	DB-eq	Preserve	Repair	Destructive	Deletions	Vetoes
Qwen-7B-IT	Off	2325	.502	.694	.701	.970	.334	35	764	0
	On	2306	.502	.703	.711	1.000	.344	0	638	108
Llama-8B-IT	Off	2875	.206	.602	.619	.928	.314	83	1448	0
	On	2901	.202	.612	.625	1.000	.294	0	1216	287

Appendix

Table 20: Evidence protection at the 90/30 operating point. Each arm evaluates 500 development goals. Destructive corrections, deletions, and vetoes are counts; preservation and repair are within-turn conditional rates defined in Appendix E.8 . Exact and DB-eq are goal-weighted.

	Qwen-7B-IT		Llama-8B-IT
Selector	Exact	SAC $-$ arm [95% CI]	Exact	SAC $-$ arm [95% CI]
Base LM draft	.518	–	.430	–
Full SAC $.70/.50$	.708	–	.640	–
Resolver only ( $p_{s}=.5$ )	.519	$+.190$ $[+.170,+.209]$	.435	$+.205$ $[+.179,+.230]$
Observable text selector	.501	$+.208$ $[+.183,+.235]$	.406	$+.234$ $[+.207,+.262]$
Shuffled readout	.586	$+.127$ $[+.106,+.148]$	.509	$+.130$ $[+.108,+.153]$

Appendix

Table 21: Selector attribution: goal-weighted exact-query accuracy under alternative selection signals. Offline fixed-prefix replay on the logged development SEARCH turns; all arms share the evidence resolver, safeguards, compiler, and fallback. SAC

-

arm is the paired difference with a goal-clustered bootstrap interval (

2{,}000

resamples, seed 954100); for the shuffled and rate-matched arms it uses the single seed-954100 draw, so it differs slightly from the difference of the

1{,}000

-draw mean accuracies.

Model	Edits	Deletions	Illegal del.	Additions	Replacements	Turns changed	Exact	DB-eq
	SAC-P edits					Legality-only rescoring
Qwen-7B-IT	2,284	1,365	371	897	22	.166	.487 $\to$ .545	.567 $\to$ .656
Llama-8B-IT	3,065	1,999	947	938	128	.400	.389 $\to$ .501	.442 $\to$ .617
Mistral-7B-IT	4,507	3,891	2,007	504	112	.752	.165 $\to$ .449	.205 $\to$ .581
OLMo-7B-IT	9,107	3,091	756	5,960	56	.462	.068 $\to$ .118	.096 $\to$ .167
OLMo-32B-IT	2,628	2,295	475	296	37	.200	.482 $\to$ .611	.586 $\to$ .741

Appendix

Table 22: Controller edit composition and legality-only rescoring on TEST1000. Edit counts are pooled over each checkpoint’s SAC-P SEARCH turns. Exact and DB-eq are goal-weighted over the logged Base LM SEARCH turns before and after removing illegal slots.

Model	$n$	Rescue	$n$	Rescue
	Binding		Deployment
Qwen-7B-IT	38	.158 [.053,.289]	197	.863 [.817,.909]
Llama-8B-IT	28	.143 [.036,.286]	72	.583 [.472,.694]
Mistral-7B-IT	11	.273 [.000,.545]	19	.947 [.842,1.00]
OLMo-7B-IT	8	.125 [.000,.375]	173	.988 [.971,1.00]
OLMo-32B-IT	5	.600 [.200,1.00]	21	.714 [.524,.905]

Appendix

Table 23: Controller rescue by natural-omission profile. Offline fixed-prefix replay on the profiled omissions; rescue is the fraction of omissions whose controller query contains the slot with its current value. Intervals are 95% bootstrap intervals over omissions.

Model	Setting	Turns	Success	Exact	DB-eq	Preserve	Repair	Destr./goal	Rescue/goal	Coverage
Qwen-7B-IT	Base LM	2405	.486	.518	.599	1.000	.000	.000	.000	.000
	90/30	2306	.502	.703	.711	1.000	.344	.000	.782	.302
	70/50	2280	.490	.753	.745	.955	.468	.108	1.010	.405
	50/50	2238	.490	.741	.731	.926	.489	.176	1.028	.445
Llama-8B-IT	Base LM	2949	.156	.430	.466	1.000	.000	.000	.000	.000
	90/30	2901	.202	.612	.625	1.000	.294	.000	1.034	.498

Appendix

Table 24: Complete two-checkpoint development comparison. All rows use 500 goals. Success is goal-level; Exact and DB-eq average over SEARCH turns within each goal and then over goals. Preservation and repair condition on draft correctness within the same turn. Rescue and destructive counts are divided by goals; coverage is the fraction of SEARCH turns whose executed query text differs from the Base LM draft.

Model	Base	PDST-0	PDST-5	PDST-10	T5-DST	SAC-C	SAC-P	SAC-B	Oracle
Qwen-7B-IT	.415/.487	.437/.449	.446/.396	.454/.379	.456/.574	.431/.675	.449/.719	.433/.698	.503/.999
Llama-8B-IT	.151/.389	.172/.395	.180/.430	.194/.444	.207/.483	.184/.563	.210/.608	.180/.580	.223/.938
Mistral-7B-IT	.086/.165	.175/.403	.150/.378	.160/.382	.185/.546	.174/.447	.181/.540	.174/.514	.226/1.000
OLMo-7B-IT	.130/.068	.292/.190	.359/.258	.363/.277	.426/.580	.312/.301	.376/.491	.394/.569	.471/.977
OLMo-32B-IT	.577/.482	.563/.330	.606/.459	.644/.484	.584/.563	.644/.683	.640/.745	.620/.723	.690/1.000
Mean	.272/.318	.328/.353	.348/.384	.363/.393	.372/.549	.349/.534	.371/.621	.360/.617	.423/.983

Appendix

Table 25: Closed-loop task success and exact-query accuracy. Cells report Success/Exact on 1,000 goals per checkpoint and arm. Macro results give each checkpoint equal weight and are computed before rounding. All controller settings use the same evidence safeguards. PDST-

k

is Prompted DST with

k

exemplars; Oracle is the privileged Oracle Query reference of Table 1 .

	TS, SAC-P $-$ Base LM		TS, SAC-P $-$ Prompted DST		Exact, $-$ Prompted
Model	$\Delta$ [95% CI]	$p$	$\Delta$ [95% CI]	$p$	$\Delta$ [95% CI]
Qwen-7B-IT	$+.034$ $[+.014,+.054]$	$.0012^{*}$	$+.012$ $[-.010,+.033]$	$.3112$	$+.270$ $[+.250,+.290]$
Llama-8B-IT	$+.059$ $[+.038,+.080]$	$<.0001^{*}$	$+.038$ $[+.016,+.059]$	$.0007^{*}$	$+.213$ $[+.192,+.235]$
Mistral-7B-IT	$+.095$ $[+.073,+.117]$	$<.0001^{*}$	$+.006$ $[-.015,+.027]$	$.6427$	$+.137$ $[+.113,+.162]$
OLMo-7B-IT	$+.246$ $[+.216,+.277]$	$<.0001^{*}$	$+.084$ $[+.051,+.114]$	$<.0001^{*}$	$+.301$ $[+.277,+.323]$
OLMo-32B-IT	$+.063$ $[+.038,+.089]$	$<.0001^{*}$	$+.077$ $[+.047,+.107]$	$<.0001^{*}$	$+.416$ $[+.396,+.436]$

Appendix

Table 26: Paired contrasts for the primary controller. Success differences are goal-paired on the shared

1{,}000

initial goals per checkpoint; intervals use

2{,}000

paired bootstrap resamples (seed 954100) and

p

-values are raw exact two-sided McNemar tests on discordant goals. ∗ marks comparisons that remain significant at

.05

after Holm correction across the five checkpoints within that comparator family; the two families are corrected separately. Exact query is goal-weighted turn-level exact-query accuracy: each goal contributes the mean over its own SEARCH turns, and goals are averaged with equal weight. Because edited queries change downstream turns, the two arms’ SEARCH turns are not matched, so this column is paired only by initial goal and no McNemar test is reported for it. The macro row is a fixed-panel contrast over the five checkpoints.

	SAC-P $-$ Prompted DST ( $k{=}5$ )			SAC-P $-$ Prompted DST ( $k{=}10$ )
Model	TS $\Delta$ [95% CI]	$p$	Exact $\Delta$ [95% CI]	TS $\Delta$ [95% CI]	$p$	Exact $\Delta$ [95% CI]
Qwen-7B-IT	$+.003$ $[-.019,+.025]$	$.856$	$+.323$ $[+.303,+.343]$	$-.005$ $[-.028,+.017]$	$.723$	$+.340$ $[+.318,+.359]$
Llama-8B-IT	$+.030$ $[+.008,+.052]$	$.0079^{*}$	$+.179$ $[+.159,+.199]$	$+.016$ $[-.006,+.036]$	$.171$	$+.164$ $[+.144,+.184]$
Mistral-7B-IT	$+.031$ $[+.008,+.052]$	$.0080^{*}$	$+.162$ $[+.137,+.185]$	$+.021$ $[-.001,+.042]$	$.078$	$+.158$ $[+.132,+.181]$
OLMo-7B-IT	$+.017$ $[-.011,+.046]$	$.273$	$+.233$ $[+.213,+.254]$	$+.013$ $[-.014,+.040]$	$.395$	$+.214$ $[+.194,+.237]$
OLMo-32B-IT	$+.034$ $[+.008,+.061]$	$.018$	$+.286$ $[+.266,+.306]$	$-.004$ $[-.030,+.021]$	$.821$	$+.261$ $[+.240,+.281]$

Appendix

Table 27: Paired contrasts against few-shot Prompted DST. Same protocol as Table 26 : success is goal-paired on the shared

1{,}000

initial goals per checkpoint with

2{,}000

paired bootstrap resamples (seed 954100) and raw exact two-sided McNemar

p

-values; ∗ marks comparisons significant at

.05

after Holm correction across the five checkpoints within that comparator family (OLMo-32B-IT at

k{=}5

has Holm

p=.053

). Exact-query differences are goal-weighted and paired only by initial goal.

	Qwen-7B-IT		Llama-8B-IT
Arm	median	p95	median	p95	Added memory
Base LM draft (reference)	655	1,148	1,202	2,002	1.9 / 4.5 GB
Prompted DST (0-shot)	813	1,224	1,222	2,171	0.8 / 1.9 GB
Prompted DST ( $k{=}5$ )	922	1,487	1,156	2,219	0.9 / 2.0 GB
Prompted DST ( $k{=}10$ )	983	1,596	1,249	2,228	1.0 / 2.1 GB
T5-DST	95	173	101	168	40 MB

Appendix

Table 28: Measured per-decision latency (ms) and added peak memory. One NVIDIA A40, batch size 1, bfloat16 backbone (T5-DST in float32),

300

identical logged SEARCH decisions from the TEST1000 Base LM trajectories per checkpoint; wall time between device synchronizations. Added memory is peak allocated memory beyond the resident backbone (

14.8

15.6

GiB). The Base LM draft is the shared reference that every arm starts from; the other rows are additional work. The controller row includes the readout, evidence resolution, and compilation. On an H200 all rows are

2

–

4\times

faster with the same ordering.

Metric	Base LM	SAC-P	SAC-P $-$ Base
Exact-query accuracy	.168	.326	$+.1575$ $[+.1453,+.1698]$
Database equivalence	.267	.429	$+.1623$ $[+.1499,+.1750]$
Initially correct Base query broken	–	.0097 $[.0072,.0126]$	–
Fraction of turns edited	–	.5944 $[.5818,.6068]$	–

Appendix

Table 29: Transfer to human-authored MultiWOZ dialogue. Base LM and the frozen primary controller are evaluated on 4,490 matched turns from 911 test dialogues. Intervals use 2,000 dialogue-clustered bootstrap resamples (seed 954100). This is fixed-history query evaluation; terminal task success is undefined.

Model	TS/Exact	TS/Exact	Preserv.	TS $\Delta$ [95% CI]	Exact $\Delta$	TS $\Delta$ [95% CI]	rate
	Base LM	SAC-P	SAC-P	SAC-P $-$ Base	SAC-P $-$ Base	Nat. $-$ templ.	Fallback
Qwen-7B-IT	.399/.446	.440/.706	.955	$+.041$ $[+.020,+.063]$	$+.260$	$-.009$ $[-.033,+.015]$	.002
Llama-8B-IT	.167/.354	.186/.553	.928	$+.019$ $[-.003,+.041]$	$+.198$	$-.024$ $[-.050,+.001]$	.002
Mistral-7B-IT	.081/.149	.189/.558	.986	$+.108$ $[+.087,+.132]$	$+.409$	$+.008$ $[-.016,+.032]$	.002
OLMo-7B-IT	.128/.058	.370/.518	.972	$+.242$ $[+.208,+.275]$	$+.461$	$-.006$ $[-.036,+.024]$	.004
OLMo-32B-IT	.530/.426	.619/.739	.977	$+.089$ $[+.061,+.118]$	$+.314$	$-.021$ $[-.051,+.008]$	.002

Appendix

Table 30: Closed-loop results with naturalized user language. Both arms run on the

1{,}000

TEST goals per checkpoint with rewritten user utterances. Contrasts are goal-paired with

2{,}000

bootstrap resamples (seed 954100); the Nat.

-

templ. contrast compares SAC-P under naturalized and template language on the same goals, and is not significant on any checkpoint after Holm correction. Fallback rate is the larger of the two arms’ template fallback rates.

Explore similar work

When Should Models Change Their Minds? Contextual Belief Management in Large Language Models
May 28, 2026Haoming Xu, Weihong Xu, Zongrui Li +6Latent Dynamics
♠Zhejiang University · ♣HomologyAI
Clarification Is Not Correction: LLMs Fail to Let Go
Sep 21, 2026Jianzhe Lin, Xiaolin Li, Fei Wang +3Large Language Models Fail Clarification
MetaAI
ReacTOD: Bounded Neuro-Symbolic Agentic NLU for Zero-Shot Dialogue State Tracking
May 18, 2026Yanjun Lin, Zimo Xiao, Kartik Natarajan +7Dialogue Benchmarks Neuro-Symbolic Framework
Amazon

May 28, 2026, cs.AIJ/K move · Enter open · S save

When Should Models Change Their Minds? Contextual Belief Management in Large Language Models

Haoming Xu, Weihong Xu, Zongrui Li, Mengru Wang+5

♠Zhejiang University · ♣HomologyAI

Latent Dynamics

Open paper AskPDF arXiv

Long-horizon interactions require language models to manage accumulating information: when to update their state, when to preserve their state, and what to ignore. We study this challenge as Contextual Belief Management (CBM): maintaining a predicted belief state aligned with formal evidence while isolating task-irrelevant noise. To make CBM measurable, we introduce BeliefTrack, a closed-world benchmark spanning Rule Discovery and Circuit Diagnosis, where a finite belief space and symbolic verifiers enable exact turn-level evaluation. BeliefTrack diagnoses three failures: Failed Stay, Failed Update, and Failed Isolation. Across multiple LLMs, vanilla models exhibit severe CBM failures, while explicit belief-tracking prompts provide limited gains. In contrast, reinforcement learning with belief-state rewards reduces failure rates by 70.9% on average. Further probing reveals latent belief-state dynamics behind these failures, and representation-level steering reduces failure rates by 46.1% across two tasks (Code is available at https://github.com/zjunlp/CBM).

Abstract

Explore similar work

When Should Models Change Their Minds? Contextual Belief Management in Large Language Models

Clarification Is Not Correction: LLMs Fail to Let Go

ReacTOD: Bounded Neuro-Symbolic Agentic NLU for Zero-Shot Dialogue State Tracking