acceptodds
Under review as a conference paper at ICLR 2027

ADAR: Abductive Distance-Aware Retrieval for Implicit-Preference Dialogue Memory

Abstract

Retrieving a user's implicit preferences requires finding supporting dialogue turns that may share few words with the current request. We study how evidence annotations available during training can teach a compact model to search for such indirect support without annotations at inference. We introduce ADAR, a dialogue search agent trained on restricted-oracle demonstrations: a teacher uses annotated evidence to explore search trajectories, filters reject paths with overt leakage, incorrect answers, or missing evidence, and the student learns from the resulting observable traces. During search, an Abductive Distance Score (ADS) summarizes lexical coverage and semantic competition as interpretable feedback alongside retrieved text. Under a shared backbone, ADAR improves accuracy over the strongest evaluated search baseline by 14.94 percentage points on PersonaMem-v2, 4.38 on held-out LongMemEval, and 2.63 on LoCoMo. Withholding numerical ADS feedback from the same checkpoint while preserving autonomous stopping lowers accuracy on all three datasets. These results support converting source evidence annotations into search supervision that improves preference answering and transfers to new dialogue histories. Our code is available at https://anonymous.4open.science/r/adar27-E72B.

Primary area:
foundation or frontier models, including LLMs
Keywords:
retrieval, knowledge augmented, conversational modeling, evaluation and metrics

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_3KDkOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes an evidence-supervised iterative retrieval framework, ADAR, for identifying implicit user preferences from dialogue histories. The method uses a teacher model with access to annotated evidence and reference answers to generate search trajectories, which are filtered based on information leakage, answer correctness, and evidence presence before being used to fine-tune a smaller student model. It also introduces an Abductive Distance Score (ADS) combining lexical coverage and semantic competition to provide feedback during iterative retrieval. Experiments on selected subsets of benchmark datasets demonstrate improvements over several retrieval and memory-based baselines.

Critical strengths
  • The paper addresses an important and practical problem of retrieving implicit user preferences from long-term dialogue histories (when supporting evidence may not be directly related to the current query).

  • The proposed framework provides a reasonable approach to leveraging annotated evidence for learning iterative search behavior through teacher-generated trajectories and supervised fine-tuning of small models.

  • Experiments on three benchmarks demonstrate improvements over several retrieval- and memory-based baselines, with ablation studies supporting the effectiveness of trajectory supervision and online feedback.

Critical weaknesses
  • The evaluation of retrieval effectiveness has more room for improvement. Although the proposed framework primarily aims to improve the retrieval of implicit supporting evidence, the experimental evaluation mainly focuses on downstream answer accuracy. Direct comparisons of evidence retrieval performance would better support the effectiveness of the learned search policy.

  • The quality and effectiveness of teacher-generated search trajectories are not sufficiently validated. Although the authors report trajectory filtering statistics and demonstrate the benefits of student SFT, these results may not verify the contributions of individual filtering strategies and the quality of intermediate search behavior. Additional ablations comparing filtered and unfiltered trajectories, along with qualitative case studies illustrating successful and failed teacher/student search trajectories, would help justify the proposed trajectory supervision pipeline.

  • The comparison and discussion of preference-aware memory construction approaches needs to be enhanced more comprehensively. Although the paper includes several memory-system baselines, the area is evolving rapidly and more recent papers (e.g., Mem-PAL, PGMem) explicitly organize user preferences and supporting information through structured memory representations. Further comparisons or discussions would help clarify the advantages and trade-offs of learning an iterative search policy over constructing preference-aware memories from various perspectives (e.g., retrieval effectiveness, computational efficiency, and adaptation to evolving user preferences).

Questions for authors

Please see weaknesses.

2 comments

Thank you for the constructive review. Accuracy and coverage values are percentages.

1. Direct evidence retrieval. Quantitative retrieval evaluation on the personalization benchmark PM shows that ADAR finds more complete support and reaches the first supporting source earlier: coverage improves by 5.02pp over PGMem and MRR@25 rises from 0.59 to 0.66 (Table 1). This directly supports improved evidence search beyond final-answer accuracy.

Methods share a 25-source and visible-token budget. Memory outputs are linked to original sources. MRR uses ordered, deduplicated sources, with misses scored zero. Direct retrieval embeds the original question and returns top25 without reformulation or iteration.

Table 1. PM retrieval.

MethodComplete support coverageMRR@25
Direct dense retrieval45.120.54
IRCoT38.940.49
PGMem47.160.59
Mem-PAL45.380.56
ADAR52.180.66

Two cases show how reformulation bridges the query-to-evidence gap. In pm_test_945, support discusses viewing the same play from front-row versus upstairs seats. Direct retrieval ranks it 100th (cosine 0.1591). ADAR's second-round query “theater performances art experiences” finds row33. In pm_test_1403, support discusses the conductor's and concertmaster's roles. It ranks 103rd (cosine 0.0243). The query “classical music concerts orchestras” finds row17 in round two. Both connect concrete dialogue details to broader interests.

2. Filtering and intermediate behavior. Evidence filtering improves both trace grounding and student performance. Answer-only filtering retains 22.80% of traces without annotated support. Full reduces this to 0% and improves PM by 3.08pp. Removing answer/evidence filtering lowers PM by 5.95pp (Table 2). Thus, answer correctness alone is insufficient to select evidence-supported search demonstrations.

Controls share the 4B initialization, 4,960 traces and optimization budget. All retain anti-leakage/validity checks. The last removes only answer/evidence thresholds. Zero denotes passing the corresponding data check.

Table 2. Selected-trace quality and student accuracy.

FilteringWrong answersMissing annotated supportPM accuracy
Full0.000.0047.31±2.42
Answer-only0.0022.8044.23±1.94
No answer/evidence filtering21.6025.4041.36±2.61

Success/failure cases distinguish source localization from integration. For c19f7a0b (arrival time, reference 6:30 pm), the student covers the answer session in five rounds and answers correctly. The teacher fails to cover it fully in four rounds and abstains. For c6853660 (coffee limit: one to two cups), the student reaches the answer session in one round but says “decreased”. The teacher answers “increased to two cups” after three rounds. Search supervision addresses reaching useful history, while the latter case shows why answer correctness is also checked.

3. Preference-aware memory and cost. ADAR's higher online cost comes from iterative search: 4.18 calls per question, versus 1.00 for PGMem/Mem-PAL. This additional computation accompanies higher PM accuracy and support coverage. Memory methods organize preferences before answering, enabling fewer online calls. Construction and later updates incur additional costs outside the online QA figures.

Table 3. PM accuracy and online QA cost per question.

MethodAccuracyInput tokensOutput tokensTotal tokensCalls
Native RAG29.32±1.382106.1283.572189.691.00
IRCoT32.37±3.512212.45154.272366.722.64
PGMem40.82±2.141662.13122.451784.581.00
Mem-PAL36.27±1.891004.3193.111097.421.00
ADAR47.31±2.426833.56512.117345.674.18

ADAR's online tokens are 4.12/6.69 times those of PGMem/Mem-PAL. The included Native RAG/IRCoT comparisons also show the accuracy–cost trade-off against search baselines. These figures exclude judging and offline memory construction. For changing preferences, memory systems update stored representations, while ADAR searches source history at answer time. We will clarify where these approaches perform adaptation and separate online answering cost from memory construction and maintenance.

Thank you again for your review. We followed up on the points you raised with a direct evaluation of evidence retrieval, an analysis of trajectory filtering and search behavior, and comparisons with preference-aware memory methods.

We’d welcome your feedback on whether our response addresses your concerns.

Reviewer_Gq63OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents ADAR, a search agent that retrieves implicit user preferences from dialogue history. A 27B teacher with access to the gold evidence generates search trajectories. Only trajectories that avoid leaking that evidence, answer correctly and retrieve the annotated evidence are kept, and these are used to fine-tune a 4B student that needs no annotations at inference. During search, the agent also receives the Abductive Distance Score (ADS), which combines how many query words the retrieved pool misses with how well the pool’s best match beats the closest competing turn. ADAR beats the strongest search baseline by 14.94 points on PersonaMem-v2, 4.38 on held-out LongMemEval and 2.63 on LoCoMo, and removing ADS at test time lowers accuracy on all three.

Critical strengths
  1. Well-motivated problem. Retrieving implicit-preference evidence that shares few words with the request is a real failure mode for memory-augmented assistants.
  2. Sensible supervision idea. Filtering teacher trajectories on whether they retrieve the annotated evidence, not just on the final answer, is reasonable. The funnel analysis shows that about 23% of correct-answer trajectories never retrieved the evidence.
  3. Transfer and reproducibility. LongMemEval is fully held out from training, and the prompts, hyperparameters and code are all provided.
Critical weaknesses
  1. The baselines look broken. After GRPO training, Search-R1 and DeepResearcher fall below chance on PersonaMem-v2 (16.6% and 21.6% on a four-option task). On LongMemEval they score 14.7% and 19.4%, far below training-free single-shot RAG (44.0%). The largest gains are measured against these baselines.
  2. The core claim is not isolated. The only training ablation compares against the untrained model. Two controls are missing: SFT on trajectories filtered by answer only, and SFT on trajectories from a teacher without privileged inputs. Without them, the gain may simply come from distilling a larger model.
  3. The ADS ablation is confounded. The model was trained with ADS in every observation, so removing it only at test time puts the model out of distribution. A student trained without ADS is needed, and Table 2 has no error bars.
  4. ADS is weak on the target task. On PersonaMem-v2, its correlation with baseline errors is only 0.04, weaker than plain cosine distance or dense rank. The label “abductive” overstates what the score measures.
  5. Evaluation is non-standard. LongMemEval and LoCoMo are scored with a “does the response honor the preference” judge, not each benchmark’s own judging. On PersonaMem-v2, a separate LLM maps free-form answers to option letters. Results are not comparable with the literature.
  6. Several gains are within noise. Gains of 2.13 points over LightMem on LongMemEval and 1.43 over AssoMem on LoCoMo are smaller than the reported standard deviations, and no significance tests are given.
Questions for authors
  1. Why do the GRPO-trained Search-R1 and DeepResearcher score below chance on PersonaMem-v2 and far below Native RAG on LongMemEval? Please report their reward curves and format-error rates.
  2. How does ADAR compare with SFT on trajectories filtered by answer correctness only, and with SFT on trajectories from a teacher without privileged inputs?
  3. How does a student trained without ADS in its context compare with removing ADS only at test time?
  4. What accuracy does the 27B teacher reach when it runs the same search loop without privileged inputs?
2 comments

Thank you for the detailed review. PM denotes PersonaMem-v2. Accuracy values are percentages.

1. RL baselines. Reward growth and low format-error rates indicate normal training behavior under the paper's 500-step GRPO setup (Table 1).

Table 1. RL diagnostics.

ModelReward, steps 1–50Reward, steps 451–500Format errors
Search-R10.180.461.40%
DeepResearcher0.170.441.80%

To examine how missing or irrelevant history affects answering, we test only Base 4B without running search. It receives either question+options alone or the same input plus history randomly sampled outside annotated support (Table 2). Accuracy is below the 25% chance level without history and falls further with distractors. This shows that irrelevant history can harm the backbone's answers. We consider retrieved distractors a possible contributor to the low RL scores.

Table 2. Base 4B accuracy under different input contexts.

ModelNo historyDistractor history
Base 4B18.8215.27

2. Controlled SFT. Restricted-oracle trajectories and evidence filtering both contribute beyond ordinary teacher distillation: Full exceeds nonprivileged-teacher and answer-only SFT by 4.91/3.08pp on PM, with the same direction on LME/LoCoMo. Controls share the 4B initialization, 4,960 traces, 2 epochs and 1,240 optimizer steps. All converge with final format errors below 2%.

3. ADS ablation. ADS contributes beyond matching train/test conditions: Full exceeds the student trained and evaluated without ADS by 2.61pp on PM. Retraining without ADS improves over inference-only removal, but retains a gap to Full. Table 3 also adds STD for every original ablation.

Table 3. Accuracy, mean±STD over three decoding seeds.

ConditionPMLMELoCoMo
Full47.31±2.4248.40±3.5735.94±0.57
Nonprivileged-teacher traces42.40±2.1646.28±2.9532.89±0.87
Answer-only filtering44.23±1.9447.00±2.7132.51±0.83
No answer/evidence filtering41.36±2.6145.96±3.1131.05±0.97
No ADS in training/inference44.70±2.0547.08±2.8433.88±0.71
No ADS at inference only40.06±2.5445.43±3.2131.76±0.83
No student SFT30.33±2.2142.92±3.0928.67±0.92
Lexical feedback only44.52±2.1547.71±3.4433.56±0.71
Semantic feedback only42.16±2.2846.63±3.1834.46±0.65

4. ADS interpretation. ADS provides an interpretable view of lexical coverage and semantic competition in evidence retrieval. The offline correlation of 0.043 measures association with baseline errors. Online ADS describes the current search pool, with utility tested separately by ablation. On 200 preference questions, negative-ADS/cosine/dense-rank correctness AUROC is 0.64/0.58/0.60. For 200 factual questions, the corresponding values are 0.57/0.61/0.63. This view is useful particularly for implicit preferences. Correlation and AUROC use different labels/definitions. “Abductive” describes the indirect query-to-evidence relation. We will clarify this meaning. Our evidence-supervised search contribution is supported by controlled training and downstream gains.

5. Evaluation protocol. ADAR's advantage persists under semantic judging without option mapping. For a consistent evaluation framework, we extend LLM-based semantic judging to PM. Within each benchmark, all methods share the rubric. The PM judge sees only the question, reference preference and response. Correct preference use passes. Generic or conflicting responses fail, and different compatible recommendations are accepted.

Table 4. Open-form PM accuracy.

MethodMean±STD
ADAR49.27±1.64
Native RAG28.11±1.17
IRCoT30.36±1.54
PGMem40.18±1.80
Mem-PAL36.52±1.29
Nonprivileged 27B teacher43.76±1.76

6. Factual-task gains. Our main contribution targets implicit personalization. Existing memory baselines are strong on factual recall. Our objective on LME/LoCoMo is to retain that capability after preference-search training. ADAR matches their performance level, with mean gains of 2.13/1.43pp, supporting maintained factual memory alongside improved preference retrieval.

7. Nonprivileged 27B teacher. The trained 4B student surpasses the teacher under the same search interface and evidence budget, supporting the value of learning the search policy at smaller deployment capacity. With gold preferences, answers and evidence hidden, teacher PM/LME/LoCoMo accuracy is 42.26±2.33/46.47±2.86/33.82±0.82. The student in Table 3 is higher by 5.05/1.93/2.12pp.

Thank you for the additional experiments. The controlled SFT comparisons and the nonprivileged-teacher result address my main concern, so I have raised my score. In the revision, please (1) report the retrained no-ADS student as the main ADS ablation and tone down the ADS and transfer claims, (2) state whether the PersonaMem-v2 train and test splits are disjoint at the user level, and (3) add results under LongMemEval’s official judge.

Reviewer_GmohOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes a dialogue search agent framework, called ADAR -Abductive Distance-Aware Retrieval- to account for implicit personalization in conversational AI, that outputs answers supported by retrieved turns given a query text and a turn-index dialogue history. ADAR uses a teacher-student model that converting source evidence annotations into search behavior that retrieves supporting dialogue turns, where aa 27B teacher uses privileged source annotations to explore shared prefix trajectories and filter out unsupported trajectories, while a 4B student model learns from the filtered traces. The model also uses Abductive Distance Score (ADS), an interpretable numerical feedback metric combining lexical anchor coverage and semantic competition to summarize lexical coverage and semantic competition during search rounds.
ADAR has been tested on PersonaMem-v2, LoCoMo, and held-out LongMemEval datasets where it showed superior accuracy compared to several training-free baselines (Search-R1, DeepResearcher, Native RAG, IRCOT).

Critical strengths

Framing implicit preference retrieval as an abductive search problem, then supervising a student model through filtered multi-turn teacher trajectories paired with the suggested interpretable addictive distance score, brings new novelty and alternatives to standard RAG or reinforcement learning models, providing a high-impact template for agentic dialogue memory systems, and demonstrating that a small 4B student model can match or exceed performance of highly demanding LLM agents..

The testing and validation have been performed at high quality standard with comparison with relevant baseline models.

Critical weaknesses
  • Many ideas of the paper have already been reported in unpublished report https://openreview.net/forum?id=fB6kthZyW2#discussion, which appears to be a rejected paper from ACL conference.

  • ADAR's reported multi-hop accuracy on LoCoMo benchmarks for instance is quite low (9.34%), which raises the question on the usefulness of the model for multi-turn dataset.

  • Open-form evaluations rely mainly on a LLM-as-judge paradigm -Qwen3.5-122B-. This raises questions about its compatibility with human judgments.

Questions for authors
  • Should explain the nature of contribution with respect to an existing archive in https://openreview.net/forum?id=fB6kthZyW2#discussion that contains more than 80% similarity to the current version.

  • Low accuracy in LoCoMo benchmark should trigger discussions about the scope and scalability of the ADAR framework to tackle random multi-turn dataset.

  • Comment on the compatibility between LLM-as-a-judge and human judgments where possible errors can have high impact on model performance

3 comments

Thank you for the constructive review.

1. Changes from the earlier manuscript. This is an expanded revision of the earlier work. The method framework is retained. The additions are comparisons with AssoMem and LightMem, broader component ablations, revised method and ADS analysis figures, and explicit definitions of offline diagnostics versus online feedback. We will state this relationship and the additions clearly in the revised paper.

2. Scope and multi-hop retrieval. ADAR focuses on locating implicit support whose surface wording is weakly related to the question. Learned query reformulation and iterative search help recover these easily missed dialogue turns. On LoCoMo multi-hop questions, ADAR achieves the highest mean among the compared methods: 9.34%, versus 8.33% for AssoMem and 4.17% for LightMem.

Multi-hop answering additionally requires collecting several complementary pieces of evidence and connecting their relations. Recovering one difficult clue helps this process. Completing the answer also requires the other clues. We will clarify the method's focus and scope in the revision, distinguishing implicit-support localization from complete multi-hop evidence collection and integration.

3. Judge reliability and score impact. We consider the judge reliable for this evaluation: it agrees with adjudicated human labels on 94% of the audited answers. All observed disagreements conservatively lower accuracy. No human-incorrect answer was credited as correct. The audit therefore supports the reported scores as reflecting correct responses rather than false-positive credit in this sample.

We sampled 50 anonymized answers across the three benchmarks. Two independent annotators used the same rubric, with a third adjudicating disagreements. They saw only the question, reference preference/answer and candidate response. Model identities, search traces and machine labels were hidden. The rubric checks factual completeness or correct preference use and accepts semantically valid alternatives to reference wording.

Table 1. Human versus Qwen3.5-122B judgments.

Human / JudgeYESNO
YES243
NO023
Reviewer_GmohOpenReview ↗

Feedback

Thanks for the clarification. The human-LLM judge study is very limited.. and therefore maintain my score unchanged

Thank you for the further feedback. We expanded human review from 50 to 100 responses, with 93 judgments agreeing with the LLM Judge. In addition, we conducted a more detailed analysis of task requirements, answer content and disagreements. We also ran supplementary tests with a prompt that explicitly distinguished required conditions from replaceable examples, observing closer agreement with human review on some advice responses.

1. Question requirements and answer coverage.

Using questions, correct options or reference answers, and gold evidence, we examined three aspects: factual content, personalization, and actionable advice. The 100 responses in the supplementary analysis were distributed as follows:

BenchmarkResponsesFactual contentPersonalizationActionable advice
PM3452633
LME332944
LoCoMo333300
Total100673037

Requirements can overlap. LME and LoCoMo mainly test factual content in this sample, while PM emphasizes personalization and actionable advice. Human review checked whether answers included the facts, preferences and advice required by each question.

We classified answer content as fully preserved, missing/altered content, or no clear omission but debatable sufficiency. Full preservation includes paraphrases and alternative examples meeting the same requirements:

BenchmarkResponsesFully preservedNot fully preservedUncertain coverage
PM345236
LME3311202
LoCoMo338250
Total10024688

Human and LLM judgments mostly agreed on clear omissions. Sufficient personalization and suitable alternatives required closer comparison.

2. Types of disagreement.

Human review and the LLM Judge disagreed on 7 of 100 responses. We categorized each by its main disagreement:

TypeCasesJudgment focus
Personalization4Whether advice uses personal clues and alternatives meet personal constraints
Actionable advice2Whether the answer provides a usable approach
Factual relations1Whether similar wording preserves the requested subject and relation
Total7One category per response

For example, a question asks for fun, eye-catching clothes for a weekend gathering. The reference suggests a bright T-shirt featuring the user's favorite cartoon characters. The answer recommends only bright colors and bold graphics. The LLM Judge accepted it as fun and eye-catching. Human review rejected it for omitting the specific cartoon preference.

3. Improvements from clearer scoring rules.

We revised the prompt to specify three checks and require supporting content from the answer:

  • Factual content: Verify subjects, numbers, times and relations, rather than accepting shared keywords or topics.
  • Personalization: Check whether the answer implements specific preferences or constraints. Accept different recommendations meeting the same requirements.
  • Actionable advice: Require usable suggestions rather than explanations alone, without requiring every reference example.

Background absent from the reference is not automatically a factual contradiction. Below are seven responses selected from the supplementary retest. Model, inputs and inference settings remained fixed:

PromptResponsesAgreed with human reviewStill disagreed
Original707
Revised725

Two advice responses changed from incorrect to correct, matching human review. The LLM Judge recognized usable advice meeting the request, rather than rejecting it solely for differing from reference examples or adding background absent from the reference. The remaining five disagreements involved four personalization cases and one factual relation, concerning whether advice adequately reflected personal preferences and whether similar wording preserved the same factual relation. We will include this analysis in the revision to help readers interpret LLM Judge decisions across answer types and their agreement with human review.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.