Modern conversational information access systems continuously extract user information from conversations to build user profiles for personalized retrieval. As user profiles accumulate over time, they inevitably become noisy, making effective user profile denoising a fundamental challenge. Existing personalized conversational query reformulation methods rely on the black-box reasoning capability of LLMs to expand the original query with relevant user profile phrases, producing a verbose query for retrieval. Such rewrite-then-retrieve pipelines reduce user profile denoising to the explicit selection of profile phrases in the sparse token space, separating personalization from downstream relevance modeling. Consequently, they often fail to identify user information that is truly beneficial for retrieval. To enable retrieval-aware user profile denoising, we propose Personalized Conversational Dense Retrieval (PCDR), which learns to jointly project the conversation history and user profile into a latent embedding space dedicated to retrieval. By inheriting the context denoising capability from conversational dense retrievers, PCDR requires only a handful of personalized relevance annotations for training. We evaluate PCDR on the iKAT 2026 tasks, where it serves as the core retrieval component of a conversational AI agent assisting users with complex information needs.
Figures & tables
Figure 1. An example of personalized conversational search. The current question depends on the dialogue history and on only two of the five profile statements, which are highlighted. PCDR encodes the conversation and the full profile into a single personalized query embedding.
Figure 2. Two-stage continual training of PCDR. The query encoder of Qwen3-Embedding 0.6B is first fine-tuned on TopiOCQA conversations to obtain conv-Qwen3, and then on iKAT conversations and user profiles, with distillation from human query rewrites, to obtain PCDR. The document encoder stays frozen in both stages.
Adaptation
Learning rate
NDCG@3
LoRA rank (lr 5e−6 )
LoRA r=4
5e−6
30.6
LoRA r=8
5e−6
31.0
LoRA r=16
5e−6
31.2
LoRA r=32
5e−6
32.2
LoRA r=64
5e−6
33.1
Table 1. Hyperparameter selection for personalized fine-tuning, in NDCG@ 3 on the held-out turns with full-corpus retrieval. All runs start from conv-Qwen3 and use the full-profile query and the complete objective. Bold indicates the best result.
Encoder
full-profile
relevant-profile
base-Qwen3
12.9
13.4
conv-Qwen3
29.0
28.9
PCDR (ours)
33.1
32.0
Table 2. Zero-shot baselines and PCDR under the two query forms, in NDCG@ 3 on the held-out turns with full-corpus retrieval. Bold indicates the best result.
Configuration
NDCG@3
pure KD
28.8
+ graded fine-tune (BiXSE)
30.3
+ initialization anchor
30.4
+ global negatives (full method, PCDR)
33.1
oracle upper bound
36.3
Table 3. Cumulative build-up of PCDR on top of conv-Qwen3, in NDCG@ 3 on the held-out turns. Each row adds one component to the row above. The upper bound encodes the human oracle rewrite with base-Qwen3.
Configuration
NDCG@3
full method
33.1
− conv-Qwen3 fine-tune (use base-Qwen3)
27.5
− oracle-rewrite KD
28.4
− initialization anchor
29.5
− global negatives
30.4
− graded fine-tune
28.6
Table 4. Leave-one-out ablation of PCDR, in NDCG@ 3 on the held-out turns. Each row removes one component from the full method and keeps all other settings fixed.
Run
Retrieval components
Reranking
6
PCDR
—
3
PCDR + CDR
—
5
PCDR + SPLADE-v3
—
4
PCDR + BM25
RankLLaMA
Table 5. The four runs submitted to the iKAT 2026 interactive task. All runs use PCDR as the personalized dense retriever and gpt-5-mini for profile extraction and response generation, and all fusions use RRF.
Run
NDCG@5
Rubric
Eng.
Rel.&Use.
Subtopic
Avg
PCDR
0.557
3.85
3.32
4.23
3.75
3.79
PCDR + CDR
0.559
3.74
3.17
4.12
3.65
3.67
PCDR + SPLADE-v3
0.577
3.97
3.20
4.11
3.69
3.74
PCDR + BM25
0.580
3.93
3.14
4.24
3.79
3.78
Table 6. Retrieval and per-turn response quality of the submitted runs with the iKAT 2026 simulated users, as judged by an LLM against the rubrics. NDCG@ 5 ranges from 0 to 1 , the rubric score from 0 to 5 , and the other criteria from 1 to 5 ; Avg averages the four response criteria.
Run
Pers.
Trust
Flow
Mix-Init
Sat.
Avg
PCDR
3.93
3.68
3.98
3.00
3.93
3.70
PCDR + CDR
3.82
3.70
3.98
2.70
4.02
3.64
PCDR + SPLADE-v3
3.80
3.39
3.86
2.82
3.84
3.54
PCDR + BM25
3.95
3.41
4.07
2.52
3.93
3.58
Table 7. Dialogue-level quality of the submitted runs with the iKAT 2026 simulated users, as judged by an LLM on a 1 – 5 scale. Avg averages the five criteria.
Figure 3. Token relevancy of the PCDR query embedding on an iKAT 2024 skincare turn, computed with the method of Chefer et al. ( Chefer et al., 2021 ) . Darker colors indicate a larger contribution to the embedding. The two profile statements annotated as relevant are marked with ▹ , and the antecedent of “it in the question is boxed.
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
Query
NDCG@3
Δ
p
w/l/t
PCDR
33.1
—
—
—
Qwen3- 0.6 B rw.
11.7
−21.4
0.0001∗
3/23/18
Qwen3- 1.7 B rw.
20.8
−12.3
0.013
8/19/17
Qwen3- 4 B rw.
33.5
+0.4
0.60
16/11/17
Qwen3- 32 B rw.
36.1
+3.0
0.066
21/12/11
gpt-4o rw.
39.5
+6.5
0.10
17/14/13
Appendix
Table 8. PCDR versus PCQR with rewriters of different sizes, in NDCG@ 3 on the 44 held-out turns. Δ and p are computed against PCDR with a paired Wilcoxon signed-rank test, and ∗ indicates significance after Bonferroni correction. Rewriter rows are scored on an int8-quantized document index.