ESAinsTOD: A Unified End-to-End Schema-Aware Instruction-Tuning Framework for Task-Oriented Dialog Modeling

Dechuan Teng; Chunlin Lu; Libo Qin; Wanxiang Che

cs.CLMar 10, 2026

ESAinsTOD: A Unified End-to-End Schema-Aware Instruction-Tuning Framework for Task-Oriented Dialog Modeling

Authors: Dechuan Teng, Chunlin Lu, Libo Qin, Wanxiang Che

Organizations: Research Center for Social Computing and Information Retrieval, Harbin Institute of Technology, Harbin, China. · School of Computer Science and Engineering, Central South University, Changsha, China.

Abstract

Existing end-to-end modeling methods for modular task-oriented dialog systems are typically tailored to specific datasets, making it challenging to adapt to new dialog scenarios. In this work, we propose ESAinsTOD, a unified End-to-end Schema-Aware Instruction-tuning framework for general Task-Oriented Dialog modeling. This framework introduces a structured methodology to go beyond simply fine-tuning Large Language Models (LLMs), enabling flexible adaptation to various dialogue task flows and schemas. Specifically, we leverage full-parameter fine-tuning of LLMs and introduce two alignment mechanisms to make the resulting system both instruction-aware and schema-aware: (i) instruction alignment, which ensures that the system faithfully follows task instructions to complete various task flows from heterogeneous TOD datasets; and (ii) schema alignment, which encourages the system to make predictions adhering to the specified schema. In addition, we employ session-level end-to-end modeling, which allows the system to access the results of previously executed task flows within the dialogue history, to bridge the gap between the instruction-tuning paradigm and the real-world application of TOD systems. Empirical results show that while a fine-tuned LLM serves as a strong baseline, our structured approach provides significant additional benefits. In particular, our findings indicate that: (i) ESAinsTOD outperforms state-of-the-art models by a significant margin on end-to-end task-oriented dialog modeling benchmarks: CamRest676, In-Car and MultiWOZ; (ii) more importantly, it exhibits superior generalization capabilities across various low-resource settings, with the proposed alignment mechanisms significantly enhancing zero-shot performance; and (iii) our instruction-tuning paradigm substantially improves the model's robustness against data noise and cascading errors.

Figures & tables

Panel of Figure 1 — Figure 1: Comparison between conventional PCMs and our instruction-tuned LLM for task-oriented dialog systems. In Figure (a), PCM A and PCM B are derived by fine-tuning pre-trained models on specific corpora A and B, respectively. Each PCM is tailored for a particular workflow and lacks the flexibility to apply to other dialogue scenarios. In contrast, as depicted in Figure (b), the TOD Instruction-tuned LLM is developed by fine-tuning a large language model on multiple corpora, employing both instruction-aware and schema-aware mechanisms. This design enables it to generalize effectively to rare or unseen dialogue scenarios.

Dataset	Annotations					# Sess.	# Turn	# Dom.	# Sch.
Dataset	NLU	DST	DB	POL	NLG	# Sess.	# Turn	# Dom.	# Sch.
CamRest676	$\times$	✓	✓	$\times$	✓	676	2,744	1	1 (0)
In-Car	$\times$	✓	✓	$\times$	✓	3,028	7,876	3	3 (0)
MultiWOZ 2.1	$\times$	✓	✓	✓	✓	10,433	71,484	7	7 (0)
SGD	✓	✓	✓	✓	✓	22,593	229,156	20	270 (528)
Frames	$\times$	✓	$\times$	$\times$	✓	1,367	9,579	1	1 (0)

Table 1: Detailed statistics of our end-to-end instruction-tuning TOD corpus 0 0 footnotetext: These datasets include either subsets or the complete set of annotations pertinent to task-oriented dialogue. In this context, “DB” denotes database query results, “Sess.” refers to the number of dialogue sessions, and “Dom.” indicates the number of domains. In the “Sch.” column, the figures outside the parentheses reflect the number of domain schemas, while those inside the parentheses represent the count of intent schemas.

Figure 2: Illustration of the instruction-tuning paradigm for task-oriented dialog systems

System Instruction
Please act as an AI assistant to interact with the user in a task-oriented dialogue scenario to meet his/her needs. For each message from the user, follow the instructions below to generate intermediate results until the assistant replies:
[1.5pt/3pt]
Full Task Instructions
DI	Please identify the domains involved in the user message from: { domain_list } .
Example : ["banking"] ; ["hotel", "train"]
ID	Select the correct intent(s) expressed in the user text among the provided intents.

Table 2: Instruction template variations across datasets. The Full Task Instructions part presents example output for each task instruction. Note that the example output is not included in the instruction template. 2 2 footnotetext: The numbers (①-⑥) denote the sequential order of the instructions and the corresponding tasks to be executed. 2 2 footnotetext: Except for BiToD, the dst_format of the datasets containing DST annotations is { "format": { " { slot_name } ": " { slot_value } " } , "examples": { "slot1": "val1", "slot2": "val2" }} .

Model	CamRest676				In-Car
Model	Match	SuccF1	BLEU	Comb	Match	SuccF1	BLEU	Comb
SEDST	92.70	75.40	23.60	107.65	84.50	82.90	19.30	103.00
TSCP	92.70	85.40	25.30	114.35	84.50	81.10	21.90	104.70
LABES	96.40	83.00	25.50	115.20	85.80	77.00	22.80	104.20
FSDM	93.50	86.20	25.80	115.65	84.80	82.10	21.50	104.95
$\text{SPACE}^{\star}$	97.74	88.24	23.68	116.67	85.26	83.16	22.92	107.13

Table 3: End-to-end dialog modeling results on CamRest676 and In-Car 3 3 footnotetext: Best results are in bold, and the second best results are underlined. ⋆ ⋆ footnotetext: Corresponding pre-trained conversational model (PCM) has been further fine-tuned on each downstream dataset individually.

Model	MultiWOZ 2.0				MultiWOZ 2.1
Model	Inform	Success	BLEU	Comb	Inform	Success	BLEU	Comb
SimpleTOD	84.40	70.10	15.01	92.26	85.00	70.50	15.23	92.98
$\text{UBAR}^{\dagger}$	85.10	71.02	16.21	94.27	86.20	70.32	16.48	94.74
SOLOIST	85.50	72.90	16.54	95.74	-	-	-	-
$\text{MinTL}^{\dagger}$	84.88	74.91	17.89	97.78	-	-	-	-
TOP+NOD	86.90	76.20	20.58	102.13	-	-	-	-

Table 4: End-to-end dialog modeling results on MultiWOZ † † footnotetext: During inference at current turn, corresponding methods require the orcale dialogue states of previous turns. ⋆ ⋆ footnotetext: Corresponding PCMs have been further fine-tuned on each dataset individually. 3 3 footnotetext: Llama 2 7B (FT) signifies that Llama 2 7B model was fine-tuned exclusively on the training set of each dataset, without cross-dataset training. 3 3 footnotetext: ESAinsTOD (GT) indicates that results are based on ground-truth system actions and responses from previous turns.

Model	MultiWOZ 2.0	MultiWOZ 2.1
Classification-based Approaches
TOD-BERT	-	48.00
$\text{DST-Picklist}^{\dagger}$	54.39	53.30
$\text{SST}^{\dagger}$	51.17	55.23
TripPy	-	55.29
$\text{CHAN}^{\dagger}$	52.68	58.55

Table 5: DST evaluation: Joint Goal Accuracy on MultiWOZ 3 3 footnotetext: The end-to-end methods tagged with (DST) include versions specifically trained for dialogue state tracking (DST) task. † † footnotetext: Corresponding methods necessitate a full ontology in which all possible values are predefined for each slot. ⋆ ⋆ footnotetext: Corresponding PCM has been further fine-tuned on each downstream dataset individually.

Model	BANKING77	CLINIC150	HWU64
Classification-based Approaches
ConvBERT	92.95	97.07	90.43
USE+ConvRT	93.36	97.16	92.62
Example+Observer	93.83	97.31	93.03
ConvFiT	94.16	97.34	92.42
$\text{SPACE}^{\star}$	94.94	97.89	94.14

Table 6: Comparison results on intent detection tasks 3 3 footnotetext: Best results are in bold, and the second best results are underlined. ⋆ ⋆ footnotetext: Corresponding PCMs have been further fine-tuned on each downstream dataset individually.

Model	Intent Acc	Slot F1	Overall Acc
BERT-Joint	99.00	96.20	91.60
Stack-Propgation+BERT	99.00	97.00	92.90
Co-Interactive transformer+BERT	98.80	97.10	93.10
ESAinsTOD ( Llama 2 7B )	99.43	96.76	92.14

Table 7: Comparison results on SNIPS dataset 3 3 footnotetext: Best results are in bold, and the second best results are underlined.

Model	Training Percentage (%)
Model	5	10	20
SimpleTOD	16.14	22.37	31.22
MinTL	21.28	30.32	35.96
SOLOIST	26.53	32.42	38.68
$\text{PPTOD}_{\text{base}}$	40.20	43.45	46.96
$\text{PPTOD}_{\text{large}}$	43.61	45.96	48.95

Table 9: Low-resource DST evaluation on MultiWOZ 2.0 3 3 footnotetext: Instruction-tuning data comprises the complete training sets of five end-to-end (E2E) dialog datasets, excluding CamRest676, along with corresponding proportions of the MultiWOZ 2.0 dataset.

Model	JGA	Inform	Success	BLEU	Combined
Turn-level evaluation on MultiWOZ 2.1 test set
ESAinsTOD ( Llama 2 7B )	63.00	91.60	73.50	6.60	89.15
w/o ia	56.17	87.80	67.00	6.61	84.01
w/o sa	48.67	69.20	39.30	6.88	61.13
w/o ia, sa	43.38	67.30	31.70	6.65	56.15
Session-level evaluation on MultiWOZ 2.1 test set

Table 10: Zero-shot evaluation results under different ablation settings 3 3 footnotetext: For turn-level evaluation, where oracle historical dialog states are included in the dialog context, the Joint Goal Accuracy (JGA) primarily measures the ability of the dialogue system to recognize the user’s needs in the current turn.

Panel of Figure 3 — Figure 3: Evaluation results of ESAinsTOD trained on varying percentages of our instruction-tuning corpus with or without schema information

Figure 4: Results of ESAinsTOD on MultiWOZ 2.1 with varying context window sizes

Schema	Context	Current	Inform	Success	BLEU	Combined
Schema	Belief	Belief	Inform	Success	BLEU	Combined
$\times$	GT	Gen	94.50	85.80	19.61	109.76
$\times$	Gen	Gen	93.40 ( $-$ 1.1)	84.70 ( $-$ 1.1)	19.54 ( $-$ 0.07)	108.59 ( $-$ 1.17)
✓	GT	Gen	94.50	87.80	21.33	112.48
✓	Gen	Gen	94.40 ( $-$ 0.1 )	87.50 ( $-$ 0.3 )	21.41 ( $+$ 0.08 )	112.36 ( $-$ 0.12 )

Table 11: Session-level cascading error analysis of ESAinsTOD on MultiWOZ 2.1 3 3 footnotetext: We report the inference results conditioned on dialogue history turns that contain either ground-truth or generated dialogue states. The numbers in parentheses denote the cascading impact of the model’s predictions from previous turns on the overall performance of dialogue modeling.

Schema	Context	Current	Inform	Success	BLEU	Combined
Schema	Belief	Belief	Inform	Success	BLEU	Combined
$\times$	Gen	GT	94.80	86.10	19.70	110.15
$\times$	Gen	Gen	93.40 ( $-$ 1.4)	84.70 ( $-$ 1.4)	19.54 ( $-$ 0.16)	108.59 ( $-$ 1.56)
✓	Gen	GT	94.10	86.90	21.52	112.02
✓	Gen	Gen	94.40 ( $+$ 0.3 )	87.50 ( $+$ 0.6 )	21.41 ( $-$ 0.11 )	112.36 ( $+$ 0.34 )
$\times$	GT	GT	94.70	86.40	19.76	110.31

Table 12: Task-level cascading error analysis of ESAinsTOD on MultiWOZ 2.1

Method	Session-Level	JGA	Inform	Success	BLEU	Comb
Backbone model: Qwen2.5-0.5B-Instruct
UBAR	✓	55.17	86.80	78.40	18.37	100.97
PPTOD	$\times$	51.22	89.90	81.30	19.78	105.38
[1.5pt/3pt]
ESAinsTOD	✓	56.43	93.40	85.30	19.56	108.91
w/o ia	✓	53.68	88.20	77.50	20.07	102.92

Table 13: Ablation study on MultiWOZ 2.1 using Qwen2.5 model family 3 3 footnotetext: Session-Level stands for session-level end-to-end task-oriented dialog modeling.

Schema Management	Samples $\downarrow$	Turns per Sample $\uparrow$	Tokens per Turn $\downarrow$	Total Tokens (Billion) $\downarrow$
$\times$	63,008	7.53	401.7	0.184
✓	47,489	8.63	252.8	0.103

Table 14: Impact of schema management on the statistics of our unified instruction-tuning dataset 3 3 footnotetext: Tokens are counted by the Qwen2.5 tokenizer.

\uparrow

: higher is better and

\downarrow

: lower is better. 3 3 footnotetext: The maximum token length per sample is 4096; dialogues exceeding this limit are split into multiple samples (see Section 3.3 for details).

Method	SM	End-to-End Dialogue Modeling					Inference Measurement
Method	SM	JGA	Inform	Success	BLEU	Comb	Latency (ms) $\downarrow$	Speedup $\uparrow$
PPTOD	-	51.22	89.90	81.30	19.78	105.38	20.94	1.00 $\times$
UBAR	-	55.17	86.80	78.40	18.37	100.97	25.96	0.81 $\times$
ESAinsTOD	$\times$	56.52	93.50	85.00	19.54	108.79	40.59	0.52 $\times$
ESAinsTOD	✓	56.43	93.40	85.30	19.56	108.91	30.65	0.68 $\times$

Table 15: Runtime efficiency comparison on MultiWOZ 2.1 using Qwen2.5-0.5B-Instruct as the backbone 3 3 footnotetext: SM stands for Schema Management.

\uparrow

: higher is better and

\downarrow

: lower is better. 3 3 footnotetext: The latency of each method is measured on a single Nvidia A100 80GB GPU using offline inference with vLLM and is calculated by dividing the total inference time by the total number of dialogue turns.

Input	Golden	Prediction
Similar Intent Group: { card_delivery_estimate (cde) , card_arrival (ca) }
Error Rate: 14/80=17.50%
User: I am waiting for my card to arrive.	cde	ca
User: How long will it take to arrive?	cde	ca
User: How long does a card delivery take?	ca	cde
Similar Intent Group: { compromised_card (cc) , card_payment_not_recognised (cpnr) , direct_debit_payment_not_recognised (ddpnr) }

Table 16: Failure cases from ESAinsTOD on BANKING77 test set 3 3 footnotetext: The reported error rates exclusively account for classification errors within intent groups.

Error Type: Incorrect Dialogue State
Dialogue: MUL2060
Turn 4	User: Great I also need to find a place to stay that doesn’t need to have free parking and is in the cheap price range.
Oracle	[State] hotel: { parking=dontcare , pricerange=cheap} [DB] hotel: 10
Ours	[State] hotel: { parking=dontcare , pricerange=cheap} [DB] hotel: 10
pptod	[State] hotel: { parking=no , pricerange=cheap} [DB] hotel: 0
[1.5pt/3pt]

Table 17: Case study on E2E dialog modeling from MultiWOZ 2.1 test set 3 3 footnotetext: Detailed database results and system actions are omitted for brevity.

Task-Oriented Dialogue Systems

Task-oriented dialogue systems use conversation to understand user goals, gather information, and complete actions across multiple turns.

See all 59 papers

Aug 31, 2026

Learning to Reason and Use Tools through Unsupervised Fine-Tuning in Task-Oriented Dialog Systems

Current dialogue systems struggle with dynamic information retrieval, often leading to hallucinations and lower response accuracy. We address this by adapting the ReAct framework for Task-Oriented Dialogue, enabling Large Language Models (LLMs) to access external knowledge and produce factual responses. Mainly, we propose an unsupervised fine-tuning pipeline that harvests reasoning trajectories via in-context learning inference. High-quality samples are filtered using an LLM-based judge to construct a robust training set. This is enhanced by a unsupervised self-improvement loop, where improved checkpoints generate increasingly better trajectories for subsequent fine-tuning iterations. Experiments on the SIMMC dataset demonstrate that ReAct-based systems outperform baselines due to superior reasoning and tool use. Notably, our fine-tuned 8B model surpasses a 70B in-context system. Finally, we present an error analysis, impact of scene complexity, and cross-domain generalization.

Markel Ferro, Oier Lopez de Lacalle

HiTZ Center - Ixa, University of the Basque Country UPV/EHU

Self-Training for Language Models LLM Fine-Tuning LLM Tool Use Task-Oriented Dialogue Systems

Ask PDF

May 18, 2026

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

Task-oriented dialogue systems -- handling transactions, reservations, and service requests -- require predictable behavior, yet the moderately-sized LLMs needed for practical latency are prone to hallucination and format errors that cascade into incorrect actions (e.g., a hotel booked for the wrong date). We propose ReacTOD, a bounded neuro-symbolic architecture that reformulates NLU as discrete tool calls within a self-correcting ReAct loop governed by deterministic validation. A bounded ReAct loop enables iterative self-correction, improving accuracy by up to 9.3 percentage points over single-pass inference on MultiWOZ. A symbolic validator enforces action compliance, schema conformance, and coreference consistency on every dialogue state update, achieving a 93.1% self-correction rate on intercepted errors and producing structured execution traces. Incremental state prediction and on-demand history retrieval keep prompts compact, empirically improving instruction adherence in parameter-constrained models. On MultiWOZ 2.1, ReacTOD achieves a new zero-shot state-of-the-art: gpt-oss-20B reaches 52.71% joint goal accuracy, surpassing the previous best by 14 percentage points, while Qwen3-8B achieves 47.34% with only 8B parameters. On the Schema-Guided Dialogue (SGD) benchmark, ReacTOD with Claude-Opus-4.6 achieves 80.68% JGA under fully end-to-end evaluation with predicted domains, and Qwen3-32B reaches 64.09% -- demonstrating cross-benchmark generalization without task-specific training data.

Yanjun Lin, Zimo Xiao, Kartik Natarajan +7

Amazon

Spoken Dialogue Systems LLM Agents State Tracking Neuro-Symbolic AI Task-Oriented Dialogue Systems Multi-Turn Dialogue Management

Ask PDF

Apr 15, 2026

Synthesizing Instruction-Tuning Datasets with Contrastive Decoding

Using responses generated by high-performing large language models (LLMs) for instruction tuning has become a widely adopted approach. However, the existing literature overlooks a property of LLM-generated responses: they conflate world knowledge acquired during pre-training with instruction-following capabilities acquired during post-training. We hypothesize that disentangling the instruction-following capabilities from pre-trained knowledge improves the effectiveness of instruction tuning. To this end, we propose CoDIT, a method that applies contrastive decoding between a post-trained model and its pre-trained counterpart during response generation. The method suppresses pre-trained knowledge shared between the two models while amplifying the instruction-following behavior acquired via post-training, resulting in responses that more purely reflect instruction-following capabilities. Experiment results demonstrate that models trained on datasets constructed via CoDIT consistently outperform those trained on directly generated responses. Training on our datasets also yields better performance than on existing publicly available instruction-tuning datasets across multiple benchmarks. Furthermore, we theoretically and empirically show that CoDIT can be interpreted as distilling the chat vector from parameter space to text space, enabling the transfer of instruction-tuning capabilities across models of different architectures.

Tatsuya Ichinose, Youmi Ma, Masanari Oi +2

Department of Computer Science, School of Computing, Institute of Science Tokyo · National Institute of Advanced Industrial Science and Technology · Research and Development Center for Large Language Models, NII

Instruction Tuning Synthetic Data Generation for Language Models Contrastive Decoding Language Model Distillation

Ask PDF

Abstract

Task-Oriented Dialogue Systems

Explore similar work

Learning to Reason and Use Tools through Unsupervised Fine-Tuning in Task-Oriented Dialog Systems

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

Synthesizing Instruction-Tuning Datasets with Contrastive Decoding