acceptodds
Under review as a conference paper at ICLR 2027

Candidate-Conditioned Implicit Debate For Recommendation Agents

Abstract

Recommendation agents can refine candidate lists through multi-role discussion, but shared summaries may obscure candidate-specific evidence, ranking changes may misalign states and items, and previous experience must be reused without interfering with the current interaction. We propose Candidate-Conditioned Implicit Debate (CCID), a recurrent framework that addresses these challenges through continuous candidate states, persistent candidate identities, and conditional reflection memory. A proposer initializes the candidates and their states; utility and risk roles assess candidate fit and the consequences of replacing the current leader; and an adjudicator updates both rankings and states for further assessment. Before discussion, relevant lessons from completed offline trajectories are retrieved from a frozen memory snapshot and remain fixed throughout the request. Experiments on four recommendation benchmarks show that CCID consistently outperforms single-round discussion. Ablations support the contributions of candidate correspondence, communication capacity, and reflection memory. Paired top-1 analysis shows that recurrent discussion corrects more errors than it introduces on all four datasets. Additional rounds improve recommendation quality with diminishing gains, while adaptive stopping achieves quality close to fixed four-round discussion with fewer model calls. Together, these results support iterative candidate-level refinement that preserves evidence, incorporates complementary assessments, and reuses applicable experience.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Recommender Systems, LLM Agents, Multi-Agent Collaboration, Implicit Communication, Reflection Memory

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 3 reviews, above 27% of the others.
Reviewer_vo1ROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes CCID, a multi-round reranking method for recommendation agents. A frozen OneRec-1.7B model proposes candidates, and its hidden state for each candidate is shared as a vector tied to that candidate's ID. A utility role checks if a candidate fits, a risk role checks if moving it up could push out a correct top item, and an adjudicator updates the ranking over several rounds, stopping early when it is stable. Lessons from past runs are pulled from a fixed memory.

Critical strengths

S1. The controls on message content are well designed. Score-only, pooled, within-user shuffled and cross-user states separate how much information is passed from whether it is attached to the right candidate, and the ordering is the same on all four datasets.

S2. Adaptive stopping is a useful, practical result. Reporting harmful revisions next to corrections is also good practice.

Critical weaknesses

W1. The paper combines utility and risk judgments over multiple rounds to reach a final decision. However, the risk role seems closer to judging whether the current top-ranked item is correct. Simply evaluating how well the current top item fits the user might achieve similar results.

W2. The final recommendation requires multiple rounds of utility and risk assessments, which can be costly. Given the same time and compute budget, it is unclear whether this approach is better than having a single model make a more careful judgment. The method averages 2-3 rounds, compared with one round for the single-round variant.

W3. The paper shows improvements over its own variants, but it is unclear whether the complex multi-agent debate structure is necessary. There is no ablation removing each role individually, nor a comparison with a single model under the same compute budget.

W4. The paper does not directly compare with existing methods under the same evaluation conditions. Since the candidate sets and evaluation protocols differ, it is difficult to judge whether the proposed method offers meaningful improvements over existing approaches.

Questions for authors

Q1. Given the same time and compute budget, does the proposed method outperform a single model making a more careful judgment?

Q2. Computational cost is important for recommender systems deployed in real-world services. Could the authors report the actual per-request latency measured on their experimental server and compare it with OneRec alone and the single-round variant?

2 comments
AuthorsOpenReview ↗

Response to Reviewer vo1R

Thank you for focusing on the quality obtained for the additional computation and the cost of real deployment. We respond to W/Q as follows.

  1. Response to W2/Q1, also covering W4 (quality at equal compute and under the current protocol). See shared Table S1. Under the same receiver-evaluation and token budgets, CCID outperforms both the recurrent single model and Re2LLM. The recurrent single model benefits from additional computation, but CCID improves further. Table S2 also shows that continuous candidate states are more effective than scalar states.
  2. Response to W1/W3 (whether risk and learned adjudication are necessary). Shared Table S2 shows that replacing risk with either the leader-fit proxy or a second utility head reduces performance. A validation-tuned fixed rule also underperforms learned adjudication, even when the learned state update is retained.
  3. Response to Q2 (measured latency and quality–cost). Shared Table S3 reports quality, tokens, calls, rounds, and four-domain p50/p95. Mean p50 is 1118ms for the recurrent single model, 1094ms for fixed-three CCID, 1461ms for fixed-four CCID, and 1002ms for adaptive CCID. Adaptive CCID retains fixed-four quality within 0.0003 NDCG@5 while reducing mean p50 by 31.4%, so it preserves nearly the same quality with less computation on already-stable requests.
  4. Response to W2/W3 (both directions of top-1 revision). Table 1 below compares fixed-three CCID and the recurrent single model with the same single-round starting point and equal budget. For each method, we average the three seed-specific candidate scores per request and then select top-1; the single-round starting predictions use the same aggregation. The analysis uses 10,000 requests each for Movies, Toys, and Yelp, and all 9,159 Beauty requests.

Table 1: Final top-1 changes from the single-round starting point

ChangeMoviesBeautyToysYelp
Recurrent single: incorrect→correct0.42% (42)1.507% (138)0.34% (34)0.12% (12)
Recurrent single: correct→incorrect0.32% (32)0.895% (82)0.22% (22)0.10% (10)
Recurrent single: net change+0.10pp+0.611pp+0.12pp+0.02pp
CCID: incorrect→correct0.48% (48)2.140% (196)0.42% (42)0.21% (21)
CCID: correct→incorrect0.30% (30)0.972% (89)0.20% (20)0.16% (16)
CCID: net change+0.18pp+1.168pp+0.22pp+0.05pp

We can observe that CCID gives the larger net correction in every domain.

Thank you for the response. While they address some of my questions, I am keeping my score for the following reasons:

  • Necessity of the roles (W1, W3). Gains over simpler alternatives (a second utility head, scalar states, rule fusion) are small and lack variance. Moreover, given its supervision target, the risk role effectively just predicts whether the current leader is correct, leaving the need for a multi-role debate unclear.
  • Equal budget (W2, Q1). All baselines split the budget into multiple short evaluations rather than testing a single, more thorough pass (e.g., one longer call or a stronger model), which was the main point of Q1. Gains over the recurrent single model are also modest.
  • Cost (Q2). Latency is only compared against variants of similar or higher cost. Relative to OneRec alone or the single-round variant, CCID remains substantially slower for only a moderate quality gain.
  • Comparison with existing methods (W4). Only Re2LLM was added under the current protocol. Since OneRec's numbers in Table 6 match Table 2—which supposedly uses a different protocol—it remains unclear why other reference methods could not be compared under the same setup.
Reviewer_7qU3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Candidate-Conditioned Implicit Debate (CCID), a framework for improving recommendation through multi-role discussion. The method aims to address information loss in text communication, candidate misalignment during reranking, and the reuse of historical experience. CCID maintains continuous candidate states and uses utility and risk roles for recurrent assessment, while a candidate-wise adjudicator updates rankings and states and conditional reflection memory provides relevant historical lessons. Experiments on four recommendation datasets show that CCID improves recommendation performance over text communication and single-round discussion, while adaptive stopping reduces model calls.

Critical strengths
  1. The paper proposes a candidate-conditioned communication framework that maintains candidate-specific continuous states and preserves candidate identities during recurrent discussion.
  2. The paper evaluates CCID on four recommendation datasets and provides ablation studies on communication, reflection memory, and discussion rounds, showing consistent improvements and reduced model calls through adaptive stopping.
Critical weaknesses
  1. Overstated agent-based framing. The paper presents CCID as a multi-agent debate framework, but its implementation does not appear to involve actual agent-based reasoning. The utility and risk roles are essentially supervised prediction modules trained with predefined targets. They evaluate candidate representations and provide outputs for the adjudicator to update rankings and states. This process is more similar to a recurrent neural reranking model than a multi-agent debate system. The paper does not demonstrate any autonomous reasoning or decision-making by these roles beyond their predefined prediction tasks. Therefore, describing these modules as agents and their interactions as debate seems overstated and does not accurately reflect the proposed method.
  2. Insufficient comparisons with existing iterative reranking and reflection-based methods. The paper lacks comparisons with existing methods that address similar problems. First, iterative refinement and candidate reranking have already been explored in recommendation systems. However, the paper mainly compares recurrent discussion with its own single-round variant, without showing its advantages over existing reranking methods. Second, reflection-based recommendation methods, such as Re2LLM and SAGER, have explored learning from historical decisions to improve future recommendations. The proposed conditional reflection memory follows a similar idea, but the experiments only compare different memory and retrieval strategies within CCID. Without comparisons with these existing approaches, it is difficult to determine whether the proposed recurrent discussion and reflection memory provide meaningful improvements over prior methods.
Questions for authors
  1. What makes the utility and risk roles agents rather than supervised prediction modules, and what distinguishes their interactions from a standard recurrent reranking model?
  2. Could the authors compare CCID with existing iterative reranking and reflection-based recommendation methods to demonstrate the advantages of its recurrent discussion and conditional reflection memory?
2 comments
AuthorsOpenReview ↗

Response to Reviewer 7qU3

Thank you for your suggestions on framework positioning and comparisons with existing methods. We hope the following clarifications and additional results address your questions and resolve any confusion.

  1. Module positioning (W1/Q1). Utility measures how well a candidate fits the user, while risk estimates whether promoting it could displace a correct leader. Both are learned with explicit supervision. A learned adjudicator combines their outputs to update each candidate’s score and state. We now call this process recurrent candidate-wise reranking and use “debate” only for the exchange of complementary candidate assessments.

  2. Value of the structure (W1/Q1). Table S1 shows that the complete CCID model outperforms the recurrent single model under the same budget. In Table S2, variants using scalar states, copied utility scores, or rule-based fusion also perform worse. These results support the value of both continuous candidate states and learned adjudication.

  3. Comparison with reflection methods (W2/Q2). Table S1 directly compares CCID with Re2LLM using the same candidates, interaction rounds, and evaluation budget. CCID achieves better performance, providing a numerical rather than purely conceptual comparison. We also expanded the related-work discussion of Re2LLM, SAGER, and iterative reranking.

  4. Memory-length control (W2/Q2). We padded the no-memory and random-lesson inputs to match the token length of conditional memory. Their performance remains essentially unchanged, while conditional memory still performs best. This confirms that its improvement comes from useful lessons rather than longer inputs.

Thank you for the additional experiments and clarifications.

Regarding W1, I think my original concern may not have been fully understood. I understand the explanation and the additional results showing the effectiveness of the proposed structure. However, my main concern is the positioning of the framework as Multi-Agent Debate. I still think which is somewhat overstated.

Regarding W2, I appreciate the additional comparison with Re2LLM. However, my main question is how the proposed Conditional Reflection Memory differs from the reflection mechanism in Re2LLM, rather than whether the complete CCID framework performs better. Reflection has already been explored in many agent-based recommendation methods, such as AgentCF and InteRecAgent. I do not require additional experiments with all these methods, I would like to see a clearer explanation of what is new in the proposed reflection mechanism compared with these existing methods.

Therefore, I will maintain my original rating. I hope the AC will take these concerns into account when making the final decision.

Reviewer_8CU9OpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

In multi-agent recommendation, fine-grained information about candidate items is often lost during textual communication; furthermore, changes in ranking can lead to misalignments in candidate data, and historical decision-making experience is difficult to reuse effectively. To address these challenges, this paper proposes the Candidate-Aware Implicit Debate (CCID) framework. This approach maintains an independent, continuous state representation for each candidate item and iteratively optimizes the candidate list through utility assessment, risk assessment, and multi-round ranking updates. Additionally, it incorporates a "conditional reflection memory" mechanism to facilitate the reuse of historical decision-making experience. Experiments conducted on four recommendation datasets demonstrate that the proposed method improves recommendation accuracy compared to baselines relying on textual communication, while an adaptive stopping mechanism helps reduce the computational overhead associated with iterative updates.

Critical strengths
  1. This paper focuses on the problems of candidate information preservation, ranking update, and cross-task experience reuse in multi-agent recommendation. The research scenario is clear, and the overall framework has certain application significance.

  2. By integrating continuous candidate representations, multi-round utility and risk assessment, reflective memory, and an adaptive stopping mechanism into a unified framework, the approach enables continuous optimization of candidate rankings.

  3. The experiment was evaluated on four datasets: Movies, Beauty, Toys, and Yelp. The effects of factors such as continuous states, candidate identity correspondence, memory retrieval, and number of debate rounds were analyzed, providing relatively rich experimental results.

Critical weaknesses
  1. The core methodology exhibits limited innovation, and there is a lack of comparison against simpler candidate re-ranking models. While Figure 3 demonstrates the importance of identity correspondence by randomly shuffling candidate states, the inclusion of mismatched candidate features would inevitably degrade predictive performance; thus, this does not effectively prove that the proposed communication mechanism offers any unique advantage.

  2. The article defines the utility role as judging whether a candidate item is the real target, and the risk role as judging whether replacing the current correct candidate will cause an error. However, based on the definition of supervision labels provided on page 5, the "risk" labels can be directly derived from the "utility" labels of the current top-ranked candidate and the candidate under evaluation;

  3. Unfair comparison with existing recommendation methods makes it difficult to demonstrate overall performance advantages. Although Table 2 lists various baselines such as SASRec, OneRec, HLLM, and STARec, these historical results seem to use different candidate construction methods and evaluation samples, making direct comparison with the paper's model impossible.

  4. The actual performance gains are limited, and the computational overhead of the complex framework is not sufficiently justified. A single round of discussion requires four model invocations; the adaptive approach averages approximately 8–10 invocations, while a fixed four-round setup requires 13. The paper provides only estimated computation times, failing to adequately validate real-world deployment latency or the performance-to-cost trade-off when compared to simpler re-ranking models.

Questions for authors

Please see the weaknesses.

1 comment
AuthorsOpenReview ↗

Response to Reviewer 8CU9

Thank you for your careful suggestions on fair comparisons and role definitions. Shuffling shows the harm of incorrect candidate–state correspondence, while a simpler reranker is needed to measure the value of the full interface. We respond to your four points as follows.

  1. Response to W1 (simple rerankers under matched budgets). See shared Table S1. Under matched interaction and computation budgets, CCID consistently outperforms the recurrent single model, the scalar recurrent reranker, and Re2LLM across all reported metrics. These results demonstrate that the complete CCID structure provides clear improvements even when simpler methods receive the same budget.

  2. Response to W2 (whether risk can be derived from utility). See shared Table S2. We directly implemented your suggested derivation as the leader-fit proxy, using the current leader and candidate utility. The complete model outperforms both this fixed derivation and the second utility head across all reported metrics. The learned risk readout therefore provides additional value that cannot be achieved by simply deriving risk from utility or repeating the utility prediction.

  3. Response to W3 (direct comparisons with existing methods). Shared Table S1 also reports direct comparisons with text communication and Re2LLM under the current protocol. Adaptive CCID achieves consistent improvements over both methods across all reported metrics. In the revised main table, these direct comparisons appear in a current-protocol panel. Earlier SASRec, HLLM, STARec, and related results appear in a historical-reference panel, with candidate construction and evaluation populations stated in the caption.

  4. Response to W4 (measured quality–cost trade-off). Shared Table S3 reports quality, calls, rounds, tokens, and latency together. Adaptive CCID maintains comparable recommendation quality to the fixed-round variant while substantially reducing mean p50 latency. These results demonstrate a more favorable quality–cost trade-off and make the additional computation and its benefits directly comparable.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Shared supplementary results

We thank the reviewers for their careful reading and constructive suggestions. The following shared protocol and tables answer the questions that recur across the three reviews.

Common protocol

All added comparisons use the same frozen OneRec-1.7B proposer, initial candidates, request histories, receiver checkpoint, and complete evaluation denominator. Every three-round condition uses one proposal plus nine receiver evaluations, with a 512-token input cap and a 64-token output cap per evaluation. Across the three-round methods, the actual mean totals per request are 4,790–4,925 input tokens and 276–305 output tokens. The controls are defined as follows:

  • Independent evaluations: nine stochastic evaluations of the unchanged initial input at temperature 0.7, using nine preregistered decoding seeds. Candidate scores are averaged before the final sort.
  • Recurrent single model: three rounds with three receiver evaluations per round. The three scores are averaged before one recurrent update. It carries a request-level recurrent hidden summary of previous scalar scores and rankings, with the same 2.74M-parameter adapter budget as the other internal controls. It has no candidate-specific continuous state board or separate role replies.
  • Scalar recurrent reranker: the same evaluation schedule and adapter budget, but only each candidate's previous scalar score and rank are retained; there is no learned hidden summary or continuous state board.
  • Re2LLM fixed-candidate reflection: three reflection rounds with three evaluations per round, using the same proposer, candidates, receiver, reflection-token cap, and score aggregation.
  • CCID: utility, risk, and adjudicator evaluations in each of three rounds, with a separate continuous state for every candidate carried across rounds. Training seeds 17, 43, and 91 are shared across learned variants. We report sample SD across training runs. For paired CIs, we average seed-specific candidate scores by request before ranking, and then run 10,000 paired request-level bootstrap resamples.

Table S1. Equal-budget single-model and recurrent-ranking comparisons

(NDCG@5, mean ± sample SD over three training seeds)

MethodMoviesBeautyToysYelp
Single round, submitted reference0.03750.17340.02640.0193
Text communication, submitted reference0.03810.17460.02710.0192
Single model, 9 independent evaluations0.0378 ± 0.00040.1759 ± 0.00060.0269 ± 0.00030.0195 ± 0.0002
Scalar recurrent reranker, 3 rounds0.0384 ± 0.00040.1787 ± 0.00070.0275 ± 0.00030.0199 ± 0.0002
Re2LLM fixed-candidate reflection, 3 rounds0.0386 ± 0.00040.1795 ± 0.00070.0276 ± 0.00030.0200 ± 0.0003
Single model, 3 recurrent rounds0.0388 ± 0.000370.1802 ± 0.000620.0278 ± 0.000240.0202 ± 0.00016
CCID, fixed 3 rounds0.0394 ± 0.000420.1848 ± 0.001350.0286 ± 0.000360.0210 ± 0.00034
CCID, adaptive0.0398 ± 0.00040.1861 ± 0.00060.0290 ± 0.00030.0214 ± 0.0002

At equal three-round cost, every paired CI for CCID minus the recurrent single model excludes zero. Across the three paired training seeds, two-sided paired t-tests give p=0.035, 0.022, 0.036, and 0.044, respectively. Repeated decoding has a smaller benefit because nine stochastic runs sample different plausible rankings, and averaging reduces the influence of occasional poor outputs. Ordinary recurrence improves further, and the complete CCID structure adds a consistent gain. Adaptive CCID can exceed fixed three because already-stable requests stop early, while unsettled requests continue to rounds four or five. This agrees with the submitted round analysis: adaptive is slightly higher on Beauty and Yelp, lower on Movies, and equal on Toys. CCID also exceeds same-protocol Re2LLM in every domain.

AuthorsOpenReview ↗

Continued: Shared supplementary results

Table S2. Role, state, and adjudicator ablations

Each variant is trained independently with the same receiver, initial candidates, memory snapshot, data, three-round schedule, and call/token budgets. The leader-fit risk proxy retains candidate utility u_j and replaces learned risk with r_hat_j = u_leader × (1−u_j). Utility copy uses a second independently trained utility head of equal size. Scalar candidate state keeps utility, risk, and learned adjudication but replaces each candidate's continuous state with its previous scalar score and rank. Rule fusion keeps the learned state-update network and replaces only score adjudication with a validation-tuned linear rule.

VariantMoviesBeautyToysYelp
Utility only0.03860.18090.02780.0202
Risk only0.03810.18100.02740.0198
Utility + leader-fit risk proxy0.03850.18010.02810.0201
Utility + utility copy0.03880.18170.02800.0204
Utility + risk + scalar candidate state0.03870.18200.02790.0203
Utility + risk, rule fusion0.03900.18320.02820.0206
Utility + risk, learned adjudicator (CCID)0.03940.18480.02860.0210

Values are means over the same three seeds. Utility and risk predictions have a mean within-request Pearson correlation of 0.43 [0.39, 0.47]. The scalar-state ablation isolates the continuous-state contribution. Other variants show small local reversals across domains, while the complete model remains best in every domain. The leader-fit and utility-copy controls test risk, and rule fusion tests learned adjudication with the state update retained.

Table S3. Measured quality, tokens, and end-to-end latency

Measurements use one A100 80GB GPU, BF16 precision, batch size 1, warmed caches, and the same request order. Latency cells report p50 / p95 milliseconds and include OneRec proposal, receiver calls, state updates, memory retrieval, and stopping decisions. Token cells report mean input/output tokens per request. Slash-separated calls, rounds, and NDCG values follow Movies / Beauty / Toys / Yelp.

MethodCalls/requestRoundsInput/output tokensNDCG@5, M/B/T/YMovies p50/p95 (ms)Beauty p50/p95 (ms)Toys p50/p95 (ms)Yelp p50/p95 (ms)
OneRec proposal1.00—1,024 / 32.0241/.1547/.0211/.0149118 / 176114 / 168115 / 171117 / 177
Single round4.0011,760 / 96.0375/.1734/.0264/.0193464 / 680477 / 697473 / 688470 / 703
Single-model recurrent, 3 rounds10.0034,912 / 301.0388/.1802/.0278/.02021130 / 16481106 / 16181114 / 16261122 / 1644
CCID fixed three10.0034,904 / 296.0394/.1848/.0286/.02101104 / 16021082 / 15701088 / 15781102 / 1598
CCID fixed four13.0046,512 / 392.0401/.1860/.0290/.02131482 / 21671464 / 21361458 / 21291440 / 2108
CCID adaptive9.52 / 8.32 / 8.98 / 9.132.84 / 2.44 / 2.66 / 2.714,846 / 289.0398/.1861/.0290/.02141080 / 1572928 / 1368990 / 14501010 / 1514

Adaptive CCID retains nearly the same quality as fixed four while reducing computation on requests whose rankings have already stabilized. The table reports quality, token usage, calls, and latency together, using measured rather than estimated time.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.