acceptodds
Under review as a conference paper at ICLR 2027

OnePred: Next-Query Prediction via Recursive Intent Memory in Multi-Turn Conversations

Abstract

LLM-based conversational systems remain largely reactive: they respond after the user types a query. A key step toward proactive interaction is next-query prediction, which anticipates the user's subsequent query from the preceding dialogue. This task poses an efficiency–quality trade-off: concatenating the full history increases input-token consumption as the conversation grows, while using only the latest turn discards cross-turn context. Our key insight is to track the user's evolving intent trajectory across topics, unresolved needs, and interest shifts, rather than repeatedly re-read the raw history. We propose OnePred, which maintains a recursively updated textual memory as its sole cross-turn context. A two-stage reinforcement learning pipeline first teaches what to predict, then what to preserve, shaping the memory into a prediction-oriented intent chain. We introduce NQP-Bench, spanning three sources of conversations selected for context-grounded predictability, and complement it with deployment-log evaluation without predictability filtering. OnePred improves prediction quality over current-turn and full-history inputs across the evaluated model regimes, with smaller gains over matched learned-summary controls. Its input-token advantage reaches 22× at long turns; accounting for generated memory yields total-token ratios of 4.0–13.5× across dialogue-length groups.

Primary area:
foundation or frontier models, including LLMs
Keywords:
next-query prediction, multi-turn conversations, proactive dialogue, recursive memory, intent memory, intent prediction, conversational AI, reinforcement learning, benchmark datasets, LLM evaluation

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_Gz1POpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes OnePred, which predicts a user's next query using a compact memory updated after each conversational turn. It trains the prediction first with the full history and then with recursive memory. Experiments show improved prediction scores and lower token use, although gains over learned-summary baselines are small.

Critical strengths

The method reduces token use substantially, even after counting the cost of generating memory. This is useful for prediction over longer conversations.

Critical weaknesses

The improvement over learned-summary baselines is small, and the paper does not report paired significance tests or variation across training runs. It is unclear whether tracking intent provides a reliable advantage over ordinary recursive summarization.

Questions for authors

Are the gains over Summary-memory RL consistent across training runs, and can you provide paired confidence intervals?

How does OnePred compare with the closest next-query prediction and learned-memory methods discussed in the paper?

1 comment
AuthorsOpenReview ↗

Reliability across runs and comparison with the closest methods

Thank you for these questions. On NQP-Wild, OnePred scores higher than both learned-summary controls in each of three training runs, and the paired 95% intervals for the reported run lie above zero. The additional comparisons below also address the closest learned-memory and next-utterance methods.

Gains over learned-summary controls. On NQP-Wild, OnePred improves over Summary-memory RL by +0.87 points (paired 95% CI [0.28, 1.48]) and over Full-history Summary RL by +0.66 ([0.06, 1.24]). These are pointwise percentile intervals from 10,000 paired bootstrap resamples of conversations drawn with replacement, retaining all prediction points within each sampled conversation and preserving prediction-point-weighted averaging. Differences use the 0-100 alignment scale.

Across three training runs, the paired gains are:

BaselineRun ARun BRun CMean +/- SD
Summary-memory RL+0.87+0.60+0.70+0.72 +/- 0.14
Full-history Summary RL+0.66+0.39+0.42+0.49 +/- 0.15

SD is the sample standard deviation across runs; the bootstrap intervals concern Run A. The controls share the memory budget, candidate count, prediction reward, and training-step budget. These comparisons test the added predictive value of OnePred's representation and training procedure under matched conditions.

On the deployment logs, the gains over Summary-memory RL are +0.40 for Best-of-3 (95% CI [-0.28, 1.07]) and +0.45 for Top-1 ([-0.18, 1.06]); these intervals leave the differences uncertain.

Closest-method comparisons. We evaluated NQP adaptations of MemAgent and MEM1, which learn recurrent memory from task outcomes, and ProUtt, which uses intent trajectories for next-user-utterance prediction.

MethodNQP-Wild Best-of-3OnePred - methodPointwise 95% CI
MemAgent adapted to NQP45.06+0.94[0.27, 1.61]
MEM1 adapted to NQP45.22+0.78[0.11, 1.44]
ProUtt, three-candidate adaptation45.40+0.60[-0.07, 1.25]
OnePred46.00--

All rows use three candidates. ProUtt's published configuration generates four, so the result here concerns its three-candidate adaptation. The intervals favor OnePred over the MemAgent and MEM1 adaptations, while the ProUtt comparison remains uncertain.

OnePred directs its bounded state toward active topics, unresolved needs, and changes in user intent. Its two-stage schedule first learns full-history prediction, then learns to retain information useful for that prediction. The matched-summary results support a gain in prediction quality at the same memory and training-step budgets. This complements the paper's token-efficiency results against full-history prediction.

Reviewer_17HNOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

OnePred predicts the next user query from a multi-turn conversation using a recursively updated textual memory as its sole cross-turn context. The reported configuration retains at most 500 memory tokens and generates three candidate queries. A two-stage reinforcement-learning procedure first trains full-history prediction, then trains sequential memory updates using terminal prediction rewards without annotated memory states. The authors introduce NQP-Bench, covering three conversation sources selected for context-grounded predictability, and evaluate a separate deployment-log sample without predictability filtering. They report improvements over current-turn and full-history interfaces, smaller improvements over matched learned-summary controls, and reduced predictor token consumption.

Critical strengths
  1. Clear formulation and appropriately scoped claims. Equation (1) and Section 3.1 (p. 3) specify exactly what crosses turns and distinguish the bounded historical state from the variable-length current observation. Sections 2 and 4.1 also distinguish alignment with one recorded continuation from general query validity, and Best-of-3 coverage from first-candidate performance.

  2. Meaningful controls beyond basic history interfaces. Section 4.2 and Table 3 (pp. 5–6) include learned-summary baselines matched on memory budget, candidate count, prediction reward, and training steps. Table 5 also compares the two-stage schedule with 700 steps of full-history training. These controls substantially improve the interpretation of the results.

  3. Several complementary evaluation checks. Recursive memory outperforms full history in all nine model–dataset comparisons in Table 2 (p. 5), with consistent human rankings. Human agreement statistics, held-out judges, cross-source testing, and the unfiltered deployment-log evaluation provide useful corroboration (Appendices H, K, and M). The authors disclose the reuse of the primary judge ensemble for rewards.

  4. Efficiency accounting includes memory generation. Table 7 (p. 7) reports approximately 4.0–13.5-fold reductions in total predictor tokens across length groups, rather than relying solely on input savings. Appendix K additionally measures preliminary latency with and without prefix caching. These results support the practical motivation for bounded context.

Critical weaknesses
  1. The distinguishing improvement over learned compression is small, with unresolved uncertainty. In Table 3 (p. 6), OnePred scores 46.00 versus 45.34 for Full-history Summary RL and 45.13 for Summary-memory RL: margins of only 0.66 and 0.87 on the 0–100 alignment scale. In Table 6 (p. 7), the unfiltered-log margins over Summary-memory RL are 0.40 for Best-of-3 and 0.45 for Top-1. Appendix G (p. 18) explicitly provides neither paired uncertainty nor training-run variability, and its marginal intervals omit the summary controls. Positive human and held-out-judge results strengthen the broader memory-versus-history finding, but do not establish the robustness of this smaller incremental advantage. This distinction matters because recurrent outcome-trained memory already appears in MemAgent and MEM1, while ProUtt models intent trajectories for next-user-utterance prediction. Section 5 correctly cites and differentiates these works; my concern is the strength of the additional empirical contribution, not missing attribution. This is a substantial, potentially repairable evidence limitation. Paired difference intervals against the matched summary controls would be especially informative; repeated training runs would strengthen a revision. The existing intervals alone do not establish that the differences are nonsignificant.

  2. The released implementation does not clearly reproduce the reported memory constraint. Section 3.1, Equation (1) (p. 3), and Appendix A (p. 13) specify a hard 500-token stored-memory cap. However, the linked repository's recurrent loop sets MAX_MEMORY_TOKENS = 1500 and uses that constant for truncation; the system prompt likewise requests memory below 1500 tokens. The evaluation script imports and applies the same constant. This is a moderate reproducibility discrepancy affecting verification of the stated budget and matched comparisons. It could reflect an outdated release rather than the experiments themselves: it does not demonstrate that the reported results used 1500 tokens or that their numerical conclusions are wrong. Please identify the exact experimental version and reconcile its truncation settings with the release. Configuration records and existing memory-length logs could resolve the issue without new training; further evaluation would be needed only if the reported cap was not actually enforced.

Questions for authors
  1. For the Table 3 comparisons and OnePred versus Summary-memory RL in Table 6, can existing per-example scores provide paired confidence intervals for the differences, with conversations treated as clusters if multiple prediction points share a conversation? How were Appendix G's approximate intervals calculated? (W1)

  2. Are paired human scores already available for the summary-control comparisons? If so, do they support the same incremental ordering, with what uncertainty? Please distinguish these comparisons from the basic-interface results. (W1)

  3. Which code version and settings produced the reported 500-token experiments? How do they differ from the released 1500-token constant and prompt, and do existing logs verify the stored-memory cap? (W2)

1 comment
AuthorsOpenReview ↗

Paired evidence, human comparison, and the memory configuration

Thank you for identifying the missing paired and training-run evidence. On NQP-Wild, both paired 95% intervals favor OnePred over the learned-summary controls, and all three training runs give positive differences. We address your three questions below.

1. Paired uncertainty, training variability, and Appendix G. The paired comparisons requested for Tables 3 and 6 are:

EvaluationBaselineOnePred - baselinePointwise 95% CI
NQP-Wild, Best-of-3Summary-memory RL+0.87[0.28, 1.48]
NQP-Wild, Best-of-3Full-history Summary RL+0.66[0.06, 1.24]
Deployment logs, Best-of-3Summary-memory RL+0.40[-0.28, 1.07]
Deployment logs, Top-1Summary-memory RL+0.45[-0.18, 1.06]

Differences are points on the 0-100 alignment scale. The NQP-Wild intervals favor OnePred over both summary controls. The deployment-log intervals span zero, leaving the smaller differences on that sample uncertain.

For the paired intervals, we draw 10,000 bootstrap resamples of conversations with replacement, using shared sampling indices across methods. All prediction points within a sampled conversation remain together. We recompute the prediction-point-weighted mean difference from unrounded scores and report its 2.5th and 97.5th percentiles.

We also report three training runs:

MethodRun ARun BRun CMean +/- SD
Summary-memory RL45.1345.2244.9845.11 +/- 0.12
Full-history Summary RL45.3445.4345.2645.34 +/- 0.09
OnePred46.0045.8245.6845.83 +/- 0.16

Across these runs, the paired gains are 0.87, 0.60, and 0.70 over Summary-memory RL (mean 0.72 +/- 0.14), and 0.66, 0.39, and 0.42 over Full-history Summary RL (0.49 +/- 0.15). Here, SD is the sample standard deviation across runs. The bootstrap table concerns Run A; the run statistics capture training variability.

Appendix G originally used the same conversation-level resampling procedure to obtain marginal intervals for individual model means. The new paired analysis instead takes the percentiles of the differences between methods, directly addressing the incremental comparison in W1.

2. Paired human comparison. Yes, paired human scores are available for OnePred and Summary-memory RL on the same 250 examples. The mean difference is +0.50 points, with a paired 95% CI of [-2.10, 3.07]. The point estimate favors OnePred, while the interval leaves the human-rated difference uncertain.

This comparison addresses the learned-summary control. The paper's human comparisons against current-turn and full-history inputs assess the broader benefit of the memory interface.

3. The 500-token experimental configuration. The reported experiments use a hard 500-token stored-memory cap. The 1500-token constant and prompt identified in your review belong to an earlier snapshot in the anonymous repository.

SettingStored-memory cap
Experiments reported in the paper500 tokens
Earlier repository snapshot identified in the review1500 tokens

The cap applies to the state retained between turns; the current observation can vary in length. Reproducing the reported setting requires the recurrent-loop limit in onepred/recurrent_agent_loop.py and the evaluation limit in tools/evaluate.py to use 500 tokens, with the memory prompt in onepred/locale.py matched to that budget.

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

This paper presents a method for predicting the next user query in multi-turn conversations. Instead of using the entire conversation history, the framework maintains a short memory that tracks changes in the user's interests and intentions. The model is trained in two stages using reinforcement learning. First, predict future queries from full conversation histories. Second, update its memory based on how well it predicts the next query. The paper also introduces NQP-Bench for evaluating this task. Moreover, the comparison doesn't show much gap between baselines and the authors' framework without significance test.

Critical strengths
  • The two-stage training approach is intuitive and well motivated. By first training on full conversation histories and then switching to memory-based prediction, the model can learn the prediction task before adapting to the memory constraint.

  • The proposed memory mechanism is simple and does not require additional architectural components.

Critical weaknesses
  • The experiments are conducted on NQP-Bench and a separate set of deployment logs, both constructed by the authors. Although the paper includes cross-source evaluation, it does not evaluate OnePred on existing benchmarks for next-query or next-utterance prediction. This makes it difficult to assess how well the proposed method generalizes beyond the authors' evaluation settings.

  • The main components of OnePred, such as recursive textual memory and outcome-based reinforcement learning, are closely related to existing approaches like MemAgent and MEM1. While the two-stage training strategy is well motivated, the overall method appears to be a combination of existing ideas applied to next-query prediction. The paper could better clarify how its approach differs from prior memory-learning methods.

Questions for authors

The improvements over learned-summary baselines are relatively small. For example, in the section 4.2, OnePred outperforms summary-memory RL by only 0.87 points on NQP-Wild. Could you provide statistical significance tests or results across multiple random seeds to verify the reliability of these improvements? Also, how does OnePred compare with existing memory-learning methods such as MemAgent and MEM1?

3 comments
AuthorsOpenReview ↗

Reliability, comparison with prior methods, and evaluation coverage

Thank you for these questions. On NQP-Wild, the paired 95% intervals favor OnePred over both learned-summary controls, and OnePred scores higher in all three training runs. The comparisons below also evaluate NQP adaptations of MemAgent and MEM1.

Reliability of the gains. Scores and differences below use the 0-100 alignment scale.

MethodNQP-Wild Best-of-3OnePred - method95% paired CI
Summary-memory RL45.13+0.87[0.28, 1.48]
Full-history Summary RL45.34+0.66[0.06, 1.24]
MemAgent adapted to NQP45.06+0.94[0.27, 1.61]
MEM1 adapted to NQP45.22+0.78[0.11, 1.44]
ProUtt, three-candidate adaptation45.40+0.60[-0.07, 1.25]
OnePred46.00--

The pointwise 95% percentile intervals use 10,000 bootstrap resamples of conversations drawn with replacement. Each resample uses shared sampling indices across methods, retains all prediction points within sampled conversations, and preserves prediction-point-weighted averaging.

We also compared the learned-summary controls across three training runs:

MethodRun ARun BRun CMean +/- SD
Summary-memory RL45.1345.2244.9845.11 +/- 0.12
Full-history Summary RL45.3445.4345.2645.34 +/- 0.09
OnePred46.0045.8245.6845.83 +/- 0.16

The paired gains average 0.72 +/- 0.14 over Summary-memory RL and 0.49 +/- 0.15 over Full-history Summary RL; SD denotes sample standard deviation across runs. The bootstrap intervals concern Run A; the across-run statistics quantify training variability.

On the deployment logs, the gains over Summary-memory RL are +0.40 for Best-of-3 (95% CI [-0.28, 1.07]) and +0.45 for Top-1 ([-0.18, 1.06]); these intervals leave the smaller differences uncertain.

Relationship to prior methods. MemAgent and MEM1 provide precedents for learning recurrent memory from task outcomes. OnePred asks which dialogue information should survive compression to predict the user's next query. Its state tracks active topics, unresolved needs, and changes in intent. The first training stage learns prediction from full histories; the second trains memory updates through the prediction reward without annotated memory states.

The learned-summary controls share the memory budget, candidate count, prediction reward, and training-step budget. They test whether OnePred's intent-oriented state and training procedure add predictive value to task-trained compression under the same budget. The 700-step full-history control in Table 5 separately accounts for training duration in assessing the two-stage schedule.

The first table directly compares NQP adaptations of the prior methods. The paired intervals favor OnePred over the MemAgent and MEM1 adaptations. ProUtt is especially relevant to intent-based next-user-utterance prediction. Its published configuration generates four candidates; here we evaluate a three-candidate adaptation to match the NQP protocol. ProUtt's adaptation is the closest comparator in score; the +0.60 difference has an interval spanning zero.

Evaluation coverage. Two NQP-Bench subsets use public human-ChatGPT conversation corpora; our construction specifies the prediction examples, filtering, splits, and scoring protocol. The cross-source experiments test transfer across the sources in the study, and the deployment-log evaluation tests conversations without predictability filtering. These evaluations address source transfer and performance on unfiltered conversations. We have not evaluated an independently defined external benchmark split, so that aspect of generalization remains open.

Thank to the authors for the answers to my questions. However, I still believe my initial overall rating reflects my assessment, and I am inclined to keep it.

Thank you for considering our response. We summarize below the additional evidence and the scope of the conclusions it supports.

  1. Reliability of the gains. On NQP-Wild Best-of-3, OnePred improves over Summary-memory RL by +0.87 points (paired 95% CI [0.28, 1.48]) and over Full-history Summary RL by +0.66 ([0.06, 1.24]). These pointwise intervals use 10,000 paired bootstrap resamples at the conversation level, preserving all prediction points within each sampled conversation. OnePred also scores higher than both controls in each of three training runs. The mean gains are 0.72 +/- 0.14 and 0.49 +/- 0.15 points, respectively, where +/- denotes the sample standard deviation across runs. The paired intervals concern the reported run; the across-run statistics capture training variability.

  2. Comparisons and contribution. The learned-summary controls share the memory budget, candidate count, prediction reward, and training-step budget. These comparisons test the added predictive value of OnePred's intent-oriented memory representation and training procedure under matched conditions. We also compared OnePred with NQP adaptations of MemAgent and MEM1 using the same three-candidate evaluation protocol. The respective gains on NQP-Wild Best-of-3 are +0.94 (paired 95% CI [0.27, 1.61]) and +0.78 ([0.11, 1.44]). These results strengthen the empirical basis for the task-specific design, alongside the paper's efficiency results against full-history prediction.

  3. Evaluation scope. NQP-Bench includes two subsets constructed from public human-ChatGPT conversation corpora. Cross-source tests assess transfer across the sources in the study, and deployment logs assess conversations without predictability filtering. These evaluations use our NQP protocol; an independently defined external benchmark split has not been evaluated. The intervals for the gains over Summary-memory RL on deployment logs and in the paired human comparison include zero, as does the interval for the gain over the three-candidate ProUtt adaptation. Our strongest additional evidence therefore concerns the incremental benefit over matched learned summaries on NQP-Wild.

We would appreciate clarification of any specific concern that remains about the experimental comparisons, methodological contribution, or evaluation scope, so that we can address it directly.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.