acceptodds
Withdrawn from ICLR 2027

STEER: Steerable Recommendation System with User Memory and Think-Then-Recommend

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Users are no longer resigned to passive consumption of algorithmic recommendations; they increasingly demand active agency over their digital feeds. Recent product launches—Instagram's "Tune Your Algorithm", Spotify's "Taste Profile", Threads' "Dear Algorithm" —all point to a fundamental gap: mainstream recommendation systems lack a principled mechanism for users to steer recommendations through open, natural-language prompts. We identify a steerable recommendation trilemma: existing paradigms are optimized for at most two of steerability, stateful personalization, and scalability—while hybrid approaches can partially address all three, no prior framework provides a principled, unified solution —collaborative filtering is personalized and scalable but not steerable; text retrieval is steerable and scalable but stateless with limited personalization; LLM-as-recommender is steerable and personalized but not scalable. We present STEER, a framework that resolves this trilemma through two core innovations: (1) STEER User Memory, an LLM-powered user profile that makes the recommendation system's understanding of each user transparent, interpretable, and editable—serving as a shared representation between user and system; and (2) Think-Then-Recommend (TTR), a two-stage architecture that explicitly reasons over the user's natural-language steering prompt and memory, then executes a structured recommendation plan via scalable embedding-based retrieval. We show empirically that personalization (from user memory) and steering (from prompts) provide strong complementary signals: their combination exceeds their individual contributions with large margins. We evaluate STEER at multiple scales: a new human-annotated benchmark (Open-Steer, 2,000+ participants across four domains), large-scale public benchmarks, and online deployment to hundreds of millions of users. Against methods consuming the same steering signal, STEER demonstrates significant gains (e.g., +37.3 points absolute HR@10 from 18.9% to 56.1%, over prompt-only retrieval on our largest dataset), confirming that explicit reasoning over transparent user memory provides value beyond simple query-item matching. Online A/B tests show statistically significant improvements in time spent (+0.40%) and app visits (+0.21%), validating that the framework scales to latency-sensitive production environments.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Recommender systems, steerable recommendation, user modeling, large language models, natural language user memory, LLM reasoning, think then embed, personalization, benchmark

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_yj6YOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper combines an editable natural-language user memory with a Think-Then-Recommend pipeline that interprets a steering prompt and memory before embedding-based retrieval. It is evaluated on a new human-annotated benchmark, a public recommendation benchmark, proprietary production data, and an online holdout study.

Critical strengths
  • This paper studies very important research questions.

  • This paper is easy to follow.

Critical weaknesses
  • The motivation of this paper is quite unclear.

  • The novelty and technical contributions of this paper can be further improved. It's unclear to me what the new insights are from this paper. The writing of this paper can be further improved for better understanding.

  • The online experiment cannot attribute its effects to the proposed memory or reasoning components, the statistical reporting is insufficient to verify the claimed significance, the memory-quality metric is unvalidated, and the production-scale claim lacks the serving detail needed to assess it

Questions for authors

see the weaknesses.

Reviewer_jHTQOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper targets recommendation systems that users can steer with free-text prompts while retaining long-term personalization and production-scale latency, framed as a "trilemma" that collaborative filtering, text retrieval, and LLM-as-recommender each satisfy only two-thirds of (Sec. 1.2, Table 1). The proposed STEER has two parts: an LLM-generated natural-language "User Memory" that is readable and editable (Sec. 3), and Think-Then-Recommend, where an LLM produces a reasoning trace from memory plus prompt, and a Think-Then-Embed-style encoder maps trace, memory and prompt into one dense vector for ANN retrieval over an item index. Evaluation covers a new human-annotated benchmark, GISTBench, and a proprietary 6.5M-user / 22.8M-post dataset with real steering prompts Baselines are organized into history-only, prompt-only, and full-synthesis tiers (Table 3); the authors additionally report memory-generation quality for a distilled 3B model and a production A/B test.

Critical strengths
  1. Open-Steer offers a potentially valuable benchmark for open-ended steering. It includes human annotations from approximately 2,000 participants across four domains, pairing user histories and steering prompts with held-out matches. If publicly released as promised, it would provide a contribution independent of the proposed method and enable external evaluation of the paper’s central complementarity claim.

  2. Controlled ablation and deployment evidence strengthen the method’s empirical support. The “no think” ablation holds the encoder, item descriptors, training data, and objective fixed, with consistent gains across all three datasets (Table 3). This helps isolate the contribution of explicit reasoning traces. The reported production deployment, spanning over three months against a 5% holdout, further supports practical feasibility under production latency constraints, although these results cannot be independently verified.

Critical weaknesses

1. The baselines do not adequately test the comparative claims.

The prompt-only tier contains only "STEER (prompt only)," with no BM25, public dense retriever, or LLM recommender/reranker, so the LLM-as-recommender row of Tab. 1 and the implied quality–cost trade-off are not empirically tested. Sequential baselines cannot observe the steering prompt, and their training data and item coverage on Open-Steer are unclear. Thus, the 37-point gain on the internal dataset demonstrates improvement over the included prompt-only baseline rather than over strong alternatives for steerable retrieval. Public retrieval baselines and a budget-controlled LLM reranker would strengthen the comparison.

2. The internal and online results lack sufficient validation details.

The topicality labeler's inputs are unspecified, preventing assessment of its claimed independence from STEER's descriptors, and potential exposure bias in the retrieval positives is unaddressed. Online lifts (+0.40% time spent, +0.21% visits) have similarly sized uncertainties, but "±," sample sizes, the exact test window, and significance tests are not given. The control also lacks the entire steering feature, so the experiment does not isolate Think-Then-Recommend. The perceived-control result likewise needs a survey instrument and effect-size definition. Because the headline numbers come from this unreleasable data, these gaps leave statistical significance and method-specific benefits unverifiable.

3. Reporting inconsistencies obscure several central results.

  • Metrics and inputs: Tab. 2's F1 values do not match the harmonic mean of the listed precision and recall, and no averaging scheme is explained. The K = 50 memory setting also needs reconciliation with the substantially stronger full-history results in Tab. 4.
  • Efficiency and comparisons: The 5–10× cost, 6× latency, and 100× relative-cost figures need consistent reference conditions. The +108%/+70% gains over the "strongest baseline" exclude STEER w/o prompt; including it gives approximately +34%/+51%.
  • Method and presentation: The memory-quality result cites Cui et al. (2026), which does not address memory generation, and the related-work discussion describes output pathways absent from the evaluated method. Smaller issues include inconsistent verification-step counts, unnamed "three necessary functions," undefined "TYA," duplicated text, and an incorrect appendix reference. Clarifying these points is necessary to assess the memory, comparative-performance, and production claims reliably.
Questions for authors
  1. Which LLM implements the think stage (model, size, whether fine-tuned and on what data)? Is it the same 2B Qwen-3-VL used as the encoder ψθ, as the scaling ablation seems to imply?
  2. For Open-Steer, what is the retrieval corpus (size, source of item descriptions), and how were SASRec/HSTU trained given about 2k users with five items each? Were item IDs shared across participants?
  3. What does "±" denote in the A/B results (standard error, 95% CI, bootstrap SD)? Please report sample sizes, p-values, and the exact test window so "statistically significant" can be assessed.
  4. For the internal dataset, were the held-out 3-day engagements collected while STEER (or a precursor) was already serving recommendations to those users? If so, how is exposure bias toward STEER-retrieved items controlled in the evaluation positives themselves?
  5. What inputs does the 2B topicality labeler consume (raw content, different tags, embeddings)? Please describe them so the independence claim can be assessed.
  6. Tab. 2: how was the 40K-user "golden set" of primary interests constructed, and which LLM served as judge? Could the judge favor the STEER-trained model's output format? Also, how is F1 computed? The reported values do not match the harmonic mean of the listed precision and recall.
Reviewer_yjtBOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes STEER, a framework for steerable recommendation that aims to address the trade-off among steerability, stateful personalization, and scalability. The method introduces two main components. First, STEER User Memory uses an LLM to generate transparent and editable user profiles from historical interactions. Second, Think-Then-Recommend uses an LLM to reason over user memory and natural-language steering prompts, then encodes the results for embedding-based item retrieval. The paper also introduces the Open-Steer benchmark. Experiments on three datasets and online A/B tests show that STEER improves recommendation performance and user engagement.

Critical strengths
  1. The paper introduces Open-Steer, a new human-annotated benchmark with over 2,000 participants across four domains, providing a useful dataset for evaluating steerable recommendation systems.
  2. The paper conducts extensive experiments on three datasets and online A/B tests with hundreds of millions of users, demonstrating the practical utility and scalability of the proposed framework.
Critical weaknesses
  1. Insufficient comparison with related work. The two proposed components are closely related to existing LLM-based recommendation methods. First, generating transparent and editable user profiles with LLMs has already been explored in previous studies. Some recent methods, such as LangPTune and LettinGo, further optimize user profile generation based on downstream recommendation performance. However, STEER does not clearly demonstrate its technical advantages over these approaches. Second, using LLM-generated user information for embedding-based retrieval is also common in existing recommendation studies. Moreover, the experiments mainly compare against traditional sequential recommendation methods and simple retrieval variants, without sufficient comparisons with closely related LLM-based methods. This makes it difficult to assess the actual methodological contribution of STEER.
  2. Lack of recommendation-aware optimization and insufficient experimental validation. The proposed framework separates user memory generation, intent reasoning, and recommendation into different stages. However, these stages are not clearly optimized together for the final recommendation objective. Specifically, the recommendation loss is used to train the downstream embedding encoder, but the paper does not explain how recommendation results can directly improve the memory generation or intent reasoning models. As a result, the generated memory and reasoning text may not be optimal for the recommendation task. More importantly, the paper emphasizes editable user memory as a key advantage, but the experiments do not directly evaluate its effectiveness. Although the online study reports improvements in user engagement and perceived control, it does not show whether editing user profiles leads to more accurate or personalized recommendations. For example, the authors do not evaluate how well recommendations follow user edits or compare recommendation performance before and after editing. Therefore, the practical benefits of editable user memory and the effectiveness of the proposed design remain insufficiently validated.
Questions for authors
  1. How does STEER differ technically from existing LLM-based user profiling and retrieval methods, such as LangPTune, LettinGo, and personalized query-based recommendation methods?
  2. Are the memory generator and Think LLM optimized using downstream recommendation feedback, and if not, how do the authors ensure that their outputs are aligned with the recommendation objective?
  3. Can the authors provide experiments to evaluate the effectiveness of editable user memory, such as whether recommendations follow user edits and whether editing improves recommendation performance?
Reviewer_GAh5OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces STEER, a framework designed to support steerable, personalized, and scalable recommendation through natural-language user instructions. The authors identify a steerable recommendation trilemma: conventional collaborative filtering systems provide personalization and scalability but lack explicit natural-language control; text-based retrieval supports steering and scalability but lacks stateful personalization; and LLM-based recommendation supports steering and personalization but is computationally expensive at large scale. STEER addresses this challenge through two main components. First, STEER User Memory represents user preferences as an LLM-generated, human-readable profile derived from historical interactions. Unlike conventional latent user embeddings, this memory can be inspected and edited by users, providing an explicit interface for correcting the recommendation system's understanding of their interests. The memory is periodically refreshed and can be generated using a distilled smaller language model to reduce computational cost. Second, Think-Then-Recommend (TTR) separates recommendation into an intent-reasoning stage and a retrieval stage. The think stage uses an LLM to interpret the user's steering prompt in the context of their historical preferences and memory, producing a natural-language reasoning trace. The recommend stage then encodes the reasoning trace, user memory, and steering prompt into a dense embedding and retrieves relevant items through approximate nearest-neighbor search. The authors evaluate STEER on three settings: Open-Steer, a newly introduced human-annotated benchmark involving 2,059 participants across four content domains; GISTBench, a large-scale behavioral dataset; and an internal production dataset containing 6.5 million users and 22.8 million posts. The full framework substantially outperforms history-only, prompt-only, and no-think baselines. On the internal dataset, STEER achieves 56.124% HR@10, compared with 27.600% for the no-think variant and 18.858% for prompt-only retrieval. The authors also report production deployment to hundreds of millions of users, with online A/B testing showing improvements of 0.40% in app-level time spent, 0.21% in app visits, and approximately 4.2% in perceived user control. Overall, the paper demonstrates the potential of combining explicit user memory and language-based intent reasoning to support personalized, steerable recommendation at industrial scale.

Critical strengths

S1. Important and timely research problem with practical relevance. The paper addresses a meaningful limitation of current recommendation systems: users often lack direct and expressive control over the content they receive. The proposed framework supports natural-language steering while maintaining personalization and scalability, making the research relevant to modern large-scale recommendation platforms. The emphasis on user agency is particularly valuable because it shifts the focus from purely engagement-driven recommendation toward systems that can explicitly incorporate user preferences. S2. Coherent system architecture integrating memory, reasoning, and retrieval. STEER provides a conceptually clear decomposition of the recommendation process. The user-memory component maintains a readable representation of long-term preferences, the think stage interprets current intent, and the retrieval stage efficiently maps the resulting intent representation to relevant content. This separation allows the system to leverage LLM reasoning without requiring expensive per-item LLM inference. The architecture is practically appealing because it can integrate with existing embedding-based retrieval infrastructure. S3. Explicit and editable user memory as a mechanism for transparency. The use of natural-language user memory is an interesting design choice. Unlike opaque embedding-based representations, STEER memory allows users to inspect and correct the system's inferred preferences. The paper also describes a scalable implementation using cached memory updates and a distilled 3B language model. The memory-generation results are encouraging: the STEER-trained 3B model achieves an F1 score of 0.725, compared with 0.528 for the vanilla 405B model and 0.428 for the vanilla 3B model. This suggests that task-specific training can substantially improve memory quality while reducing computational cost.

Critical weaknesses

W1. The methodological novelty is relatively limited compared with existing LLM-based recommendation and retrieval frameworks. The two main components of STEER build on established techniques. LLM-generated user profiles have been explored in personalized recommendation and conversational systems, while reasoning-conditioned retrieval closely resembles existing approaches such as Think-Then-Embed, which the authors explicitly adopt. The main contribution is therefore the integration of user memory, natural-language intent reasoning, and dense retrieval into a scalable recommendation architecture. While this integration is practically meaningful, the paper does not clearly establish which aspects constitute fundamentally new methodological contributions. The authors should more explicitly distinguish their contributions from prior work on conversational recommendation, personalized retrieval, and reasoning-enhanced embedding models. W2. Baseline comparisons are not fully controlled for information access and model capacity. The paper compares STEER with SASRec and HSTU, which rely only on historical interaction sequences and do not consume explicit steering prompts. Although these baselines provide useful reference points, they are not direct competitors for the central task of steerable recommendation. The more relevant comparisons are prompt-only retrieval and STEER without the think stage. However, the paper does not include sufficiently strong alternatives that combine user history and steering prompts through other mechanisms, such as personalized query rewriting, prompt-conditioned retrieval, or learned fusion of user and query embeddings. Consequently, the experiments demonstrate the benefit of STEER over the selected baselines but do not fully establish its advantage over other plausible architectures for steerable recommendation.

Questions for authors

Q1. Methodological novelty Could the authors clarify the key methodological differences between STEER's Think-Then-Recommend architecture and existing reasoning-conditioned retrieval approaches such as Think-Then-Embed? Which components are newly proposed, and which are adaptations of existing techniques? Q2. Stronger steerable recommendation baselines Could the authors compare STEER against alternative architectures that consume both user history and steering prompts, such as personalized query rewriting, learned fusion of user and prompt embeddings, or prompt-conditioned retrieval without explicit reasoning? Q3. Necessity of explicit reasoning Could the authors compare the think stage with simpler alternatives, such as LLM-based query rewriting, relevant-memory extraction, or learned attention over user-memory attributes? Would the performance advantage remain under matched inference budgets?

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

Related papers

Discussion (0)

Sign in to comment.