acceptodds
Under review as a conference paper at ICLR 2027

Ground the Past to Predict the Future: Optimizing Natural-Language User Representations Without Downstream Supervision

Abstract

In search and recommendation systems, user behavioral logs such as search queries and item clicks have long served as the substrate for user modeling. Typically, these interaction histories are distilled into dense representations, such as user embeddings trained end-to-end for the downstream task they serve. More recently, advances in large language models (LLMs) have opened a second axis for user representation in the form of natural-language user profiles, textual summaries of a user's behaviors. However, when checked against the input behaviors they are meant to summarize, LLM-generated profiles often misrepresent the user, overstating or over-inferring beyond what the evidence shows. This same interpretability means a profile's quality can be measured by whether it faithfully abstracts past behavior, a training signal that requires no downstream task at all. It is also a natural hypothesis that such faithfulness to the past would help predict the future. Yet existing methods still optimize profiles using downstream supervision from future interactions, leaving this hypothesis untested. We study this question by operationalizing groundedness as a reward along three axes (cohesion, alignment, and truthfulness) and training a profile generator against it with groupwise Direct Preference Optimization (DPO), without any downstream supervision. Across three behavioral log datasets, including a proprietary industry dataset, and multiple personalized search and recommendation tasks, groundedness-optimized profiles consistently improve over the base model and perform comparably to those trained with downstream supervision, ranking best or second-best in 13 of 16 settings. Groundedness further complements downstream supervision when the two rewards are combined. These results suggest that faithfulness to past behavior can serve as an effective training signal for user representations.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
personalized search, recommendation systems, natural language user representation, user profile generation, implicit feedback, large language models, direct preference optimization, reward modeling, groundedness

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_W4XZOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper trains a natural-language user-profile generator using a groundedness reward based on cohesion, alignment, and truthfulness, together with cluster-size and coverage regularizers. Groupwise DPO uses only past behavior. Across three datasets and several retrieval, ranking, and personalized-search settings, the resulting profiles improve downstream performance and often match profiles trained with future-interaction supervision.

Critical strengths
  1. The problem formulation is useful. Optimizing user representations without requiring future-interaction labels could make profile learning more reusable across downstream tasks. The paper clearly separates the retrospective training signal from the future-facing evaluation.

  2. The comparisons control several important confounders. The methods share the backbone, sampled rollouts, and optimization pipeline, allowing a more focused comparison of reward functions. Evaluation on disjoint users and temporal windows, together with checkpoint selection based on groundedness, strengthens the claim that future labels are unnecessary for the proposed training procedure.

  3. The evaluation covers multiple consumers and provides some informative decomposition. Gains are tested across different retrievers, rerankers, and two generator families. The observation that truthfulness contributes more to retrieval while alignment contributes more to fine-grained ranking is potentially useful, as is the complementary benefit of combining groundedness with downstream supervision.

Critical weaknesses
  1. The central explanation is not sufficiently isolated. Groundedness optimization simultaneously changes profile faithfulness, profile count, evidence grouping, coverage, and total length. Removing reward components demonstrates their contribution to the complete procedure, but does not establish whether faithfulness itself explains the gains over simpler evidence-preserving summarization or improved compression.

  2. The groundedness measurement lacks independent validation after optimization. Agreement among three LLM judges on training rollouts is reassuring, but agreement is not equivalent to factual correctness. Since the policy is explicitly optimized against these scores, an independent audit of optimized test profiles is particularly important for assessing whether the reward improves grounding or favors judge-preferred phrasing.

  3. The comparison with downstream supervision supports a narrower conclusion than the framing suggests. RLPF uses four-way prediction with random negatives, while LettinGo uses sentiment classification with random negatives. These are specific proxy rewards and may be poorly aligned with large-pool retrieval or hard-negative ranking. Matching them does not establish that groundedness is comparable to stronger task-aligned downstream supervision.

Questions for authors
  1. How does the method compare with extractive or evidence-preserving summaries under matched profile-count and token budgets?

  2. Do independent human judgments confirm improved groundedness on optimized test profiles?

  3. Would the relative advantage persist against a downstream reward using hard negatives or the actual evaluation ranker?

4 comments
AuthorsOpenReview ↗

Author Response #1

Thank you for the thoughtful review. Since each weakness (W#) has a corresponding question (Q#), we address them together.

(1) Isolating groundedness from profile structure with budget-matched comparison (W1, Q1)

To isolate groundedness from profile structure (total length, coverage, and number of profiles), we re-analyze the MerRec results in Table 4. For each test instance, we compute the paired score difference between Ours and each method on the same user, input window, and target items, and apply ANCOVA (analysis of covariance) on these paired differences:

Δscore = β₀ + β₁·Δwords + β₂·Δcoverage + β₃·Δ#profiles + ε

The intercept β₀ is the expected advantage of Ours when the two profiles have the same length, coverage, and number of profiles. Collinearity among the covariates is low (VIF ≤ 2.7), so β₀ is reliably estimated.

vs.L1 (E5)L1 (BGE)L2 (E5)L2 (Q3-R)L2 (LRL)
Base+0.0191***+0.0257***+0.0053*+0.0089***+0.0099***
RLPF+0.0133***+0.0115***−0.0023+0.0026+0.0019
LettinGo+0.0115***+0.0093***−0.0009+0.0041+0.0038
Ours+RLPF+0.0010+0.0012−0.0045**+0.0007+0.0027

(* p < 0.05, ** p < 0.01, *** p < 0.001)

In addition, among these paired instances, we keep only the 25% where the two methods' profiles are already most similar in structure (within this subset, the profiles differ on average by only 13–20 words and 0.5 or fewer profiles, versus 46–65 words and 0.7–1.6 profiles over all instances). The table below reports the mean paired difference (Ours − method) on this subset.

vs.L1 (E5)L1 (BGE)L2 (E5)L2 (Q3-R)L2 (LRL)
Base+0.0235***+0.0285***+0.0074*+0.0111**+0.0133***
RLPF+0.0131**+0.0139**−0.0015+0.0000+0.0052
LettinGo+0.0089*+0.0107***+0.0007+0.0026+0.0028
Ours+RLPF−0.0044+0.0016−0.0034+0.0007+0.0005

(* p < 0.05, ** p < 0.01, *** p < 0.001)

Both analyses agree and follow the same pattern as the main results in Table 4. On L1, Ours remains ahead of Base and both downstream-supervised methods. On L2, Ours remains ahead of Base under all three rankers and is comparable to RLPF and LettinGo (no significant paired difference), without using any future labels. As expected, Ours+RLPF shows no significant difference from Ours in most settings, and is slightly better on L2 (E5) as it additionally uses future labels. These results suggest that the gains of Ours largely come from how its profiles describe the user, rather than from their structure. These analyses are included in Sec. 5 (Budget-Controlled Comparison, Table 7 Table 8) and App. A.8 (Table 11) App. A.9 (Table 14).

Regarding extractive or evidence-preserving summaries, Base is the closest control. It compresses the same behavioral units into a few profiles using a prompt that requires each profile to cite its supporting units and to avoid exaggerated or overly abstract descriptions (App. A.9.1 App. A.10.1). All methods share this prompt and pipeline, differing only in whether and how they are trained. However, prompting alone does not ensure faithful compression. As shown in App. A.1, the same prompt yields unsupported claims not only from our backbone but also from frontier models such as Claude Sonnet 5 and GPT-5.1. This suggests that directly training for groundedness is what keeps Ours ahead of Base even when matched in structure.

AuthorsOpenReview ↗

Author Response #2

(2) Independent validation on optimized test profiles (W2, Q2)

We evaluate groundedness of the optimized profiles on the MerRec test set in two ways with external judges, and a human audit is currently in progress a human audit is reported in Author Response #3 below.

(2-a) External judge on the same criteria. To separate the effect of training from any bias of the training judges, we score alignment and truthfulness against the behavioral units with GPT-5.4-mini, using the same judge prompts.

(2-b) Faithfulness against the raw log. We also score each profile directly against the raw interaction log. Raw logs are much larger than their behavioral units (aggregation compresses raw events into about 31.5% as many units), so we sample 500 test instances shared across all methods to keep the evaluation cost manageable. Two judges from different model families (GPT-5.4-mini and Claude Sonnet 5.5) score each profile as 1.0 (clearly supported), 0.5 (partially supported or overstated), or 0.0 (unsupported), and we average their scores.

Method(2-a) Alignment(2-a) Truthfulness(2-b) Raw-log faithfulness
Base0.9410.6280.722
RLPF0.9590.6660.758
LettinGo0.9600.6740.742
Ours0.9710.7550.824

Under the external judge, Ours scores highest on both criteria, with the largest margin on truthfulness (0.755 vs. 0.628 for Base), while RLPF and LettinGo improve much less (0.666 and 0.674). Against the raw log, our training improves faithfulness over Base by 0.102, compared to +0.036 for RLPF and +0.020 for LettinGo, and Ours is significantly higher than all other methods (paired t-test, p < 0.001). Both judges independently rank Ours first. Since these judges differ from the training judges and the raw-log criterion differs from the training reward, the gains are unlikely to reflect judge-preferred phrasing alone. These results are now in the revised manuscript (Sec. 5, Groundedness of Optimized Profiles, Table 5; judge prompt in App. A.9.3 App. A.10.3).

(2-c) Human audit (in progress). (done; results in Author Response #3 below) Two independent annotators with expertise in search and recommendation are reviewing a 100-instance subset of these 500 instances. Through a web interface we deployed, each annotator sees a user's raw input log alongside the profiles from each method, with method identities masked, and judges whether each profile goes beyond the evidence. We will share the results, along with human-LLM agreement, within the discussion period. We hope the LLM judge results in (2-a, 2-b) partially address your concern in the meantime. The results are now reported in Author Response #3 below.

We will address W3 (+ Q3) in a separate response shortly. W3 (+ Q3) is addressed in Author Response #3 below.

AuthorsOpenReview ↗

Author Response #3: Human audit results (W2, Q2) and hard-negative downstream reward (W3, Q3)

As promised in (2-c) and at the end of Author Response #2, we report the human audit results and the W3 (+ Q3) response below.

(2-c) Human audit results (W2, Q2)

Setup. Two human annotators with expertise in search and recommendation reviewed a 100-instance subset of the 500 raw-log evaluation instances. Through the web interface (App. A.5, Figure 8), each annotator sees the raw input log and the generated profile set Pu=p1,...,pkP_u = {p_1, ..., p_k} from each method, with method identities masked and presentation order randomized. Each profile pip_i is rated 1.0 (clearly supported), 0.5 (partially supported or overstated), or 0.0 (unsupported).

Results. Mean raw-log faithfulness, with the LLM judge from (2-b) for reference:

MethodLLM judge (500 inst.)Annotator 1Annotator 2Human mean
Base0.7220.6070.4300.518
RLPF0.7580.6540.4420.548
LettinGo0.7420.6770.4420.560
Ours0.8240.7650.5220.643

Both annotators rate Ours highest. The paired advantage of Ours is significant against every method for each annotator (paired t-test, p < 0.001):

Gain of Ours overAnnotator 1Annotator 2
Base+0.158+0.092
RLPF+0.111+0.079
LettinGo+0.088+0.080

Training with the groundedness reward improves faithfulness judged directly against raw logs by human annotators, over Base and over both downstream-supervised controls. The absolute levels differ between annotators (Annotator 2 is stricter; per-annotator rating distributions are in App. A.5, Table 11), but the ordering of methods is the same.

Human-LLM agreement. We measure pairwise preference agreement, i.e., how often two raters prefer the same method in a method pair (random agreement is 0.5).

Ratervs. other annotatorvs. Claude Sonnet 5.5vs. GPT-5.4-mini
Annotator 10.700.700.72
Annotator 20.700.740.72

Human-LLM agreement (0.70-0.74) is comparable to agreement between the two annotators (0.70). At the method level, all four raters rank Ours first and Base last or tied for last. The ordering of RLPF and LettinGo in the middle varies across raters, and their human means differ by only 0.012.

These results are now in the revised manuscript: Sec. 5 Groundedness of Optimized Profiles, App. A.5 Human Evaluation Details.

(3) Stronger downstream supervision (W3, Q3)

To examine whether a harder prediction task makes downstream supervision more effective, we additionally train RLPF on a variant of its four-way task in which the random distractors are replaced by hard negatives constructed in the same way as the L2 ranking slates (App. A.2.1). For each target item in the training set, we sample three negatives from the hard negative pool (BM25 top-50 and BGE-M3 top-50 retrieved with the positive item as the query). All other components are identical. We evaluate on MerRec and report the change relative to Base.

vs. BaseL1 (E5)L1 (BGE)L2 (E5)L2 (Q3-R)L2 (LRL)
RLPF-HN−0.0026−0.0034+0.0005+0.0027‡+0.0044‡

(‡: significant improvement over Base, paired t-test, p < 0.001; no mark: not significant.)

Relative to Base, hard-negative RLPF yields no improvement on L1 retrieval, and its gains on L2 ranking are small (not significant with E5). These gains are smaller than those of RLPF with random negatives and of Ours (Table 4). This is consistent with hard negatives emphasizing fine-grained discrimination among similar items, as in L2 ranking. Thus, making the downstream task harder with hard negatives (RLPF-HN) does not yield large gains. The experiment is in App. A.7 (Table 13).


All revisions are reflected in the manuscript and highlighted in blue. We hope these additions address your concerns, and we would be glad to discuss any remaining questions. Thank you again for your time and constructive feedback.

Thanks a lot for the detailed rebuttal, which addresses most of my previous concerns. I'll raise my score to 3.

Reviewer_aL5AOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a framework for optimizing natural-language user profiles without downstream supervision. It introduces a groundedness reward based on cohesion, alignment, and truthfulness, and uses groupwise DPO to improve profile generation. Experiments on three behavioral datasets demonstrate improvements in personalized retrieval and recommendation, achieving performance comparable to downstream-supervised methods.

Critical strengths

1.The paper treats the faithfulness of user profiles to historical behaviors as an independent optimization objective, improving personalization without downstream task supervision.

2.The method constructs a reward function based on cohesion, alignment, and truthfulness, and employs Groupwise DPO to optimize user profile generation.

3.Experiments across three behavioral log datasets and multiple downstream tasks show that the proposed method outperforms the base model and achieves performance comparable to downstream-supervised methods.

Critical weaknesses
  1. Cohesion is computed using semantic similarity, while alignment and truthfulness rely on LLM scores in [0, 1]. The paper directly averages these components without clearly explaining how their scales are calibrated, potentially leading to imbalanced contributions.

  2. The policy model and the alignment and truthfulness judges share the same backbone. Although the authors assess agreement across different LLM judges, independent human evaluation is lacking, leaving potential biases in the reward evaluation insufficiently addressed.

  3. The optimized profiles differ structurally from the baseline, with fewer profiles per user (2.74 vs. 3.93) and lower behavioral coverage (0.89 vs. 0.93). Without controlled experiments on profile number, length, and coverage, it remains unclear whether the gains result from groundedness optimization or changes in profile structure.

Questions for authors

Please refer to Weaknesses

3 comments
AuthorsOpenReview ↗

Author Response #1

Thank you for the helpful review. We address each weakness below.

(1) Scale of the reward components (W1)

Cohesion combines intra-cluster similarity, a penalty on its variance, and separation from units outside the cluster (Sec. 3, Groundedness Reward), so it is not bounded to [0,1][0, 1] by construction. In practice, however, it falls in a comparable range (0.47 to 1.01 on rollouts from Base, the initial policy). Beyond the range, what matters for the average is how much each component varies, since a component with a much larger spread would dominate it. On the same rollouts, the standard deviation is 0.073 for alignment, 0.071 for cohesion, and 0.085 for truthfulness. They are comparable in magnitude, so no single component dominates. Still, we agree that simple averaging does not guarantee balance in general, and that normalizing each component before aggregation, as in GDPO (Liu et al., 2026), is a natural extension.

(2) Shared backbone between policy and judges (W2)

To address potential bias from the shared backbone, we evaluate groundedness on the MerRec test set in two ways with external judges, and a human audit is currently in progress a human audit is reported in Author Response #3 below.

(2-a) External judge on the same criteria. To separate the effect of training from any bias of the training judges, we score alignment and truthfulness against the behavioral units with GPT-5.4-mini, using the same judge prompts.

(2-b) Faithfulness against the raw log. We also score each profile directly against the raw interaction log. Raw logs are much larger than their behavioral units (aggregation compresses raw events into about 31.5% as many units), so we sample 500 test instances shared across all methods to keep the evaluation cost manageable. Two judges from different model families (GPT-5.4-mini and Claude Sonnet 5.5) score each profile as 1.0 (clearly supported), 0.5 (partially supported or overstated), or 0.0 (unsupported), and we average their scores.

Method(2-a) Alignment(2-a) Truthfulness(2-b) Raw-log faithfulness
Base0.9410.6280.722
RLPF0.9590.6660.758
LettinGo0.9600.6740.742
Ours0.9710.7550.824

Under the external judge, Ours scores highest on both criteria, with the largest margin on truthfulness (0.755 vs. 0.628 for Base), while RLPF and LettinGo improve much less (0.666 and 0.674). Against the raw log, our training improves faithfulness over Base by 0.102, compared to +0.036 for RLPF and +0.020 for LettinGo, and Ours is significantly higher than all other methods (paired t-test, p < 0.001). Both judges independently rank Ours first. That is, the gains hold with judges outside the policy's model family and when measured against the raw log. These results are now in the revised manuscript (Sec. 5, Groundedness of Optimized Profiles, Table 5; judge prompt in App. A.9.3 A.10.3).

(2-c) Human audit (in progress). (done; results in Author Response #3 below) Two independent annotators with expertise in search and recommendation are reviewing a 100-instance subset of these 500 instances. Through a web interface we deployed, each annotator sees a user's raw input log alongside the profiles from each method, with method identities masked, and judges whether each profile goes beyond the evidence. We will share the results, along with human-LLM agreement, within the discussion period. We hope the LLM judge results in (2-a, 2-b) partially address your concern in the meantime. The results are now reported in Author Response #3 below.

AuthorsOpenReview ↗

Author Response #2

(3) Structural differences (W3)

We re-analyze the MerRec results in Table 4 while controlling for differences in profile number, length, and coverage (now in Sec. 5, Budget-Controlled Comparison, Table 7 Table 8; App. A.8, Table 11 App. A.9, Table 14). For each test instance, we compute the paired score difference between Ours and each method on the same user, input window, and target items, and apply ANCOVA (analysis of covariance) on these paired differences:

Δscore = β₀ + β₁·Δwords + β₂·Δcoverage + β₃·Δ#profiles + ε

The intercept β₀ is the expected advantage of Ours when the two profiles have the same length, coverage, and number of profiles. Collinearity among the covariates is low (VIF ≤ 2.7), so β₀ is reliably estimated.

vs.L1 (E5)L1 (BGE)L2 (E5)L2 (Q3-R)L2 (LRL)
Base+0.0191***+0.0257***+0.0053*+0.0089***+0.0099***
RLPF+0.0133***+0.0115***−0.0023+0.0026+0.0019
LettinGo+0.0115***+0.0093***−0.0009+0.0041+0.0038
Ours+RLPF+0.0010+0.0012−0.0045**+0.0007+0.0027

(* p < 0.05, ** p < 0.01, *** p < 0.001)

In addition, among these paired instances, we keep only the 25% where the two methods' profiles are already most similar in structure (within this subset, the profiles differ on average by only 13–20 words and 0.5 or fewer profiles, versus 46–65 words and 0.7–1.6 profiles over all instances). The table below reports the mean paired difference (Ours − method) on this subset.

vs.L1 (E5)L1 (BGE)L2 (E5)L2 (Q3-R)L2 (LRL)
Base+0.0235***+0.0285***+0.0074*+0.0111**+0.0133***
RLPF+0.0131**+0.0139**−0.0015+0.0000+0.0052
LettinGo+0.0089*+0.0107***+0.0007+0.0026+0.0028
Ours+RLPF−0.0044+0.0016−0.0034+0.0007+0.0005

(* p < 0.05, ** p < 0.01, *** p < 0.001)

Both analyses agree and follow the same pattern as the main results in Table 4. On L1, Ours remains ahead of Base and both downstream-supervised methods. On L2, Ours remains ahead of Base under all three rankers and is comparable to RLPF and LettinGo (no significant paired difference), without using any future labels. As expected, Ours+RLPF shows no significant difference from Ours in most settings, and is slightly better on L2 (E5) as it additionally uses future labels. These results suggest that the gains of Ours largely come from how its profiles describe the user, rather than from their structure.

AuthorsOpenReview ↗

Author Response #3: Human audit results (follow-up to (2-c))

As promised in (2-c) of Author Response #1, we report the human audit results below. They are also included in the revised manuscript (Sec. 5 Groundedness of Optimized Profiles, App. A.5 Human Evaluation Details).

Setup. Two human annotators with expertise in search and recommendation reviewed a 100-instance subset of the 500 raw-log evaluation instances. Through the web interface (App. A.5, Figure 8), each annotator sees the raw input log and the generated profile set Pu=p1,...,pkP_u = {p_1, ..., p_k} from each method, with method identities masked and presentation order randomized. Each profile pip_i is rated 1.0 (clearly supported), 0.5 (partially supported or overstated), or 0.0 (unsupported).

Results. Mean raw-log faithfulness, with the LLM judge from (2-b) for reference:

MethodLLM judge (500 inst.)Annotator 1Annotator 2Human mean
Base0.7220.6070.4300.518
RLPF0.7580.6540.4420.548
LettinGo0.7420.6770.4420.560
Ours0.8240.7650.5220.643

Both annotators rate Ours highest. The paired advantage of Ours is significant against every method for each annotator (paired t-test, p < 0.001):

Gain of Ours overAnnotator 1Annotator 2
Base+0.158+0.092
RLPF+0.111+0.079
LettinGo+0.088+0.080

Training with the groundedness reward improves faithfulness judged directly against raw logs by human annotators, over Base and over both downstream-supervised controls. The absolute levels differ between annotators (Annotator 2 is stricter; per-annotator rating distributions are in App. A.5, Table 11), but the ordering of methods is the same.

Human-LLM agreement. We measure pairwise preference agreement, i.e., how often two raters prefer the same method in a method pair (random agreement is 0.5).

Ratervs. other annotatorvs. Claude Sonnet 5.5vs. GPT-5.4-mini
Annotator 10.700.700.72
Annotator 20.700.740.72

Human-LLM agreement (0.70-0.74) is comparable to agreement between the two annotators (0.70). At the method level, all four raters rank Ours first and Base last or tied for last. The ordering of RLPF and LettinGo in the middle varies across raters, and their human means differ by only 0.012.


All revisions are reflected in the manuscript and highlighted in blue. We hope these additions address your concerns, and we would be glad to discuss any remaining questions. Thank you again for your time and constructive feedback.

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

The paper asks whether natural-language user profiles can be trained to faithfully describe past behavior without using future-interaction labels, yet still predict future behavior. A model groups past interactions into evidence clusters and writes profiles. Cohesion, evidence alignment, and truthfulness form a groundedness reward, optimized with groupwise DPO.

Critical strengths
  1. The reward uses only the input window, and the same sampled rollouts, model, and optimization pipeline are reused across groundedness and future-supervised controls (Sections 3–4; Appendix A.6). This makes the comparison more informative than a change of both model and objective.

  2. The study uses three different logs, held-out users, multiple retrieval/ranking models, reward ablations, and a second generator backbone (Tables 4 and 8). Groundedness improves over the base generator in 15 of 16 main settings; combining it with a future-looking reward further improves fine-grained ranking on MerRec (Figure 4). The appendix also checks agreement among three LLM judges on 500 profiles (Table 6).

Critical weaknesses
  1. The central 'faithfulness' reward is not independently validated against actual user intent or human factuality judgments. The judges compare a profile to behavioral units that an LLM has already summarized from the raw log (Appendix A.3–A.4). Three judges' agreement (Table 6) is useful but cannot show that the upstream aggregation retained the original evidence, nor that inferred stable interests are justified by clicks. A human audit of raw events, aggregated units, and final profiles would test whether the intended quality actually improves.

  2. The downstream advantage may partly reflect shorter, differently structured profiles rather than groundedness alone. On MerRec, training reduces mean profile length from 207 to 160 words and changes the number and size of clusters (Table 5). The regularizer-only ablation addresses one alternative, but a length-/cluster-matched profile control is not reported. The biggest reported relative gains are in full-catalog Recall@100, whose absolute values remain low (e.g., 0.0192 to 0.0344 on MerRec with E5); fine-grained ranking gains are smaller (Table 4). These points limit the strength of the claimed general downstream benefit without negating the measured improvements.

Questions for authors
  1. Can independent human annotators assess whether profiles are actually supported by the raw interactions, not only the LLM-aggregated units, and whether your training improves those ratings over the controls?

  2. When profile length and number of clusters are matched, do the gains over the base and downstream-supervised controls persist? Please include paired differences between methods, particularly for L2 ranking, rather than only significance against the untrained model.

4 comments
AuthorsOpenReview ↗

Author Response #1

Thank you for the constructive review. We address W1 (with Q1), W2 (with Q2), and the ethics concern separately.

(1) Independent validation of faithfulness from raw logs to final profiles (W1, Q1)

Following your suggestion, we evaluate whether profiles are actually supported by the raw interaction logs, scoring them directly against the raw logs rather than the aggregated behavioral units. Since raw logs are much larger than their behavioral units (aggregation compresses raw events into about 31.5% as many behavioral units), we sample 500 test instances from MerRec, shared across all methods, to keep the evaluation cost manageable.

(1-a) LLM judge. Two external judges from different model families (GPT-5.4-mini and Claude Sonnet 5.5) score each profile as 1.0 (clearly supported), 0.5 (partially supported or overstated), or 0.0 (unsupported), and we average their scores.

MethodFaithfulness (mean)
Base0.722
RLPF0.758
LettinGo0.742
Ours0.824

Our training improves faithfulness over Base by 0.102, compared to +0.036 for RLPF and +0.020 for LettinGo, and Ours is significantly higher than all other methods (paired t-test, p < 0.001). Both judges independently rank Ours first. That is, our training improves faithfulness even when judged directly against the raw logs. These results are now in the revised manuscript (Sec. 5 Groundedness of Optimized Profiles, Table 5(b); judge prompt in A.9.3 A.10.3), and we will add the human audit results to Table 5(b) and the main text once they are complete and the human audit results are now included in Table 5(b).

(1-b) Human audit (in progress) (done; results in Author Response #3 below). Two independent annotators with expertise in search and recommendation are reviewing a 100-instance subset of these 500 instances. Through a web interface we deployed, each annotator sees a raw input log alongside the profiles from each method, with method identities masked, and judges whether each profile goes beyond the evidence. We will share the results, along with human-LLM agreement, within the discussion period. We hope the LLM judge results in (1-a) partially address your concern in the meantime. The results are now reported in Author Response #3 below.

AuthorsOpenReview ↗

Author Response #2

(2) Budget-controlled comparison with paired differences (W2, Q2)

To test whether the downstream gains persist once differences in profile structure (total length, coverage, and number of profiles) are controlled for, we re-analyze the MerRec results in Table 4. For each test instance, we compute the paired score difference between Ours and each method on the same user, input window, and target items, and apply ANCOVA (analysis of covariance) on these paired differences:

Δscore = β₀ + β₁·Δwords + β₂·Δcoverage + β₃·Δ#profiles + ε

The intercept β₀ is the expected advantage of Ours when the two profiles have the same length, coverage, and number of profiles. Collinearity among the covariates is low (VIF ≤ 2.7), so β₀ is reliably estimated.

vs.L1 (E5)L1 (BGE)L2 (E5)L2 (Q3-R)L2 (LRL)
Base+0.0191***+0.0257***+0.0053*+0.0089***+0.0099***
RLPF+0.0133***+0.0115***−0.0023+0.0026+0.0019
LettinGo+0.0115***+0.0093***−0.0009+0.0041+0.0038
Ours+RLPF+0.0010+0.0012−0.0045**+0.0007+0.0027

(* p < 0.05, ** p < 0.01, *** p < 0.001)

In addition, among these paired instances, we keep only the 25% where the two methods' profiles are already most similar in structure (within this subset, the profiles differ on average by only 13–20 words and 0.5 or fewer profiles, versus 46–65 words and 0.7–1.6 profiles over all instances) The table below reports the mean paired difference (Ours − method) on this subset.

vs.L1 (E5)L1 (BGE)L2 (E5)L2 (Q3-R)L2 (LRL)
Base+0.0235***+0.0285***+0.0074*+0.0111**+0.0133***
RLPF+0.0131**+0.0139**−0.0015+0.0000+0.0052
LettinGo+0.0089*+0.0107***+0.0007+0.0026+0.0028
Ours+RLPF−0.0044+0.0016−0.0034+0.0007+0.0005

(* p < 0.05, ** p < 0.01, *** p < 0.001)

Both analyses agree and follow the same pattern as the main results in Table 4. On L1, Ours remains ahead of Base and both downstream-supervised methods. On L2, Ours remains ahead of Base under all three rankers and is comparable to RLPF and LettinGo (no significant paired difference), without using any future labels. As expected, Ours+RLPF shows no significant difference from Ours in most settings, and is slightly better on L2 (E5) as it additionally uses future labels. These results suggest that the gains of Ours largely come from how its profiles describe the user, rather than from their structure. Regarding the low absolute Recall@100, this applies to all methods since L1 retrieves from the full item pool (over 2M items on MerRec), and the L2 results above show that the gains over Base also hold for fine-grained ranking. These analyses are included in Sec. 5 (Budget-Controlled Comparison, Table 7 Table 8) and App. A.8 (Table 11) A.9 (Table 14).

(3) Ethics and data governance

We appreciate the request for clarification. To clarify our data, we use two publicly released research datasets (MerRec and AOL4PS), used in accordance with their release terms, and one proprietary log (Srv.), which was anonymized and filtered for sensitive information internally before being provided for research use. Srv. is not publicly released, and no user IDs or log examples from it appear in the paper. We address each point below.

  • User IDs in the appendix: These were included only to make examples traceable to the public datasets. Following your concern, they are now redacted in App. A.3 and replaced with letters (A–E) in App. A.7 App. A.8.
  • Linking back to individuals: We use only the anonymized IDs provided in each dataset and make no attempt to link profiles to individuals or external data.
  • Inferred traits: Our profiles summarize observable behavior (e.g., queries, clicks, viewed or purchased items) for retrieval and ranking, and are not designed to infer protected attributes.

These points are now summarized in an Ethics Statement in the revised manuscript.

AuthorsOpenReview ↗

Author Response #3: Human audit results (follow-up to (1-b))

As promised in (1-b) of Author Response #1, we report the human audit results below. They are also included in the revised manuscript (Sec. 5 Groundedness of Optimized Profiles, App. A.5 Human Evaluation Details).

Setup. Two human annotators with expertise in search and recommendation reviewed a 100-instance subset of the 500 raw-log evaluation instances. Through the web interface (App. A.5, Figure 8), each annotator sees the raw input log and the generated profile set Pu=p1,...,pkP_u = {p_1, ..., p_k} from each method, with method identities masked and presentation order randomized. Each profile pip_i is rated 1.0 (clearly supported), 0.5 (partially supported or overstated), or 0.0 (unsupported).

Results. Mean raw-log faithfulness, with the LLM judge from (1-a) for reference:

MethodLLM judge (500 inst.)Annotator 1Annotator 2Human mean
Base0.7220.6070.4300.518
RLPF0.7580.6540.4420.548
LettinGo0.7420.6770.4420.560
Ours0.8240.7650.5220.643

Both annotators rate Ours highest. The paired advantage of Ours is significant against every method for each annotator (paired t-test, p < 0.001):

Gain of Ours overAnnotator 1Annotator 2
Base+0.158+0.092
RLPF+0.111+0.079
LettinGo+0.088+0.080

Training with the groundedness reward improves faithfulness judged directly against raw logs by human annotators, over Base and over both downstream-supervised controls. The absolute levels differ between annotators (Annotator 2 is stricter; per-annotator rating distributions are in App. A.5, Table 11), but the ordering of methods is the same.

Human-LLM agreement. We measure pairwise preference agreement, i.e., how often two raters prefer the same method in a method pair (random agreement is 0.5).

Ratervs. other annotatorvs. Claude Sonnet 5.5vs. GPT-5.4-mini
Annotator 10.700.700.72
Annotator 20.700.740.72

Human-LLM agreement (0.70-0.74) is comparable to agreement between the two annotators (0.70). At the method level, all four raters rank Ours first and Base last or tied for last. The ordering of RLPF and LettinGo in the middle varies across raters, and their human means differ by only 0.012.


All revisions are reflected in the manuscript and highlighted in blue. We hope these additions address your concerns, and we would be glad to discuss any remaining questions. Thank you again for your time and constructive feedback.

Thank you for the additional experiments. The blinded human audit against raw logs addresses my main concern about faithfulness, and the structure-controlled analyses reduce my concern about profile length and count. I also appreciate the data-governance clarifications and redaction of identifiers. The audit remains limited to MerRec, and L2 gains are modest, but the new evidence is sufficient to change my assessment. I am updating my recommendation to positive.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.