acceptodds
Under review as a conference paper at ICLR 2027

Parametric Session States Recover Retrieval Intent without Replaying History

Abstract

Long-running assistants often interpret follow-up requests whose referents, constraints, and active goals were established many turns earlier. The central challenge is therefore not only context length, but retrieval intent: a current query such as “use the previous configuration” may omit the project, entity, and version terms needed to locate relevant evidence. Long-context models can recover those terms by replaying the interaction, but repeatedly pay the cost of processing history. Conventional retrieval-augmented generation (RAG) is cheaper, yet assumes an information-sufficient query; conversational rewriting either rereads the history or depends on an explicitly maintained textual memory. We introduce Parametric Session State (), a reusable session-memory layer that remembers what to search for while keeping factual knowledge external. An offline writer compiles the ordered interaction together with answer-free retrieval supervision into a fixed-shape, session-specific LoRA state. At query time, the state briefly produces a compact, answer-free retrieval clue; the system then retrieves inspectable source text, unloads the state, and lets the frozen base model answer from that evidence. Across controlled and sessionized public evaluations, correct PSS states improve target-evidence retrieval for underspecified follow-ups. On the registered controlled and public tracks, matched-wrong and shuffled controls attribute the gains to the corresponding session, and the retrieval benefit transfers across model families and heterogeneous backends. A three-seed carrier confirmation further isolates the contribution of ordered history beyond the supplied retrieval target and shows that the parameter carrier outperforms a compute-matched persistent text memory under dense semantic retrieval. These results establish as an effective retrieval-supervised memory carrier for RAG: it restores missing session scope, supports strong evidence access across diverse systems, and preserves an external, auditable knowledge store.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Parametric Session State, Retrieval-Augmented Generation, Long-Term Conversational Memory

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_qnY2OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper addresses the problem of missing retrieval intent in long-running conversations: users’ follow-up requests often omit project, entity, configuration, or version information established earlier, making it difficult for conventional RAG systems to retrieve the correct evidence from the current query alone. The paper proposes Parametric Session State (PSS), which compiles conversational information into a session-specific LoRA state through history language modeling and supervised readout of positive and negative retrieval cues. At inference time, this state is temporarily loaded to generate retrieval cues, which are passed to a frozen retriever to query external sources. The state is then unloaded, and the frozen base model produces an answer grounded in the retrieved evidence.

The method is evaluated on the synthetic PSS-320 dataset, conversational adaptations of public QA tasks, UltraDomain, and multiple base models and retrieval backends. Wrong-session, shuffled-state, and target-only controls are used to investigate the sources of improvement. The results support the effectiveness of PSS in improving evidence retrieval under the specified supervision conditions and indicate that conversational history contributes information beyond the explicitly provided retrieval targets.

Critical strengths

S1. Targeted methodological design. The method uses session state to recover missing retrieval information and decouples retrieval from answer generation, ensuring that the final answer remains grounded in inspectable external evidence.

S2. Substantial controlled experimentation. Comparisons involving incorrect states, shuffled states, and different training inputs examine whether the gains arise from information specific to the corresponding session, strengthening the credibility of the findings.

S3. Broad empirical coverage. The evaluation spans multiple datasets, language models, and retrieval backends. It also demonstrates that retrieval improvements can benefit downstream answering, providing evidence of applicability across different settings.

Critical weaknesses

W1. The practical task formulation and motivation for the method are insufficiently convincing.

Retrieval failures caused by underspecified follow-up queries are a legitimate problem. However, the paper does not sufficiently establish how its task formulation corresponds to the practical needs of long-running assistants. In the example shown in Figure 1, a single clarification exchange with the user appears sufficient to resolve the ambiguity.

The proposed offline training procedure requires some of the key information linking the query to its session context. Consequently, the evaluation primarily examines how to store and recover an already established retrieval scope, rather than fully addressing intent recognition in natural conversation. Under this assumption, directly storing the parsed information as text is a natural and efficient alternative. Why, then, is it necessary to encode this information in trained parameters? In Figure 3, Parsed reference achieves a retrieval score of 96.95%, compared with 96.56% for PSS. Substantial improvements over the information-deficient Raw query baseline therefore do not, by themselves, establish the need for session-specific LoRA training.

Although the introduction motivates the work through long-running assistants, the main paper primarily demonstrates readout from a fixed conversational history after it has been encoded. It does not clearly explain how the state is updated as the conversation evolves, how long a trained state remains reusable, or how the training and maintenance costs are amortized. The paper has not demonstrated that the benefits justify the additional parameter training and maintenance overhead introduced by PSS.

W2. The method primarily combines existing components, and its technical novelty is insufficiently established.

The method combines existing LoRA adaptation, history language modeling, positive and negative cue supervision, and external retrieval. Its main distinction is maintaining a parameter state for each session and loading it only before retrieval. This workflow has some design merit, but the paper does not sufficiently clarify what distinctive capability or learning mechanism it introduces beyond existing approaches to parametric memory, memory-assisted cue generation, and conversational query rewriting.

In particular, the authors identify MemoRAG as the closest work in terms of functional decomposition. A clearer explanation is therefore needed as to whether PSS differs primarily in where information is stored and how components are invoked, or whether these differences enable a new capability. The existing mechanism controls can establish that the correct state contains useful session information, but they cannot independently establish that storing this information in parameters is preferable to storing it as text.

W3. The experimental comparisons with closely related baselines are insufficient.

The paper provides a substantial set of internal baselines and mechanism controls, but direct comparisons with the closest external methods remain limited. For example, MemoRAG is explicitly identified as the closest approach in terms of functional decomposition, yet no corresponding experimental results are reported. Related parametric memory methods are also discussed primarily at a conceptual level.

The external conversational query rewriting comparison in the main paper largely relies on ConvSearch-R1. Meanwhile, experiments across retrieval backends primarily establish interface compatibility and cannot substitute for competitive comparisons with complete memory-based systems. The paper should provide same-task comparisons with representative closely related methods, including relevant approaches discussed in Related Work, or explain concretely why a fair comparison is not feasible.

W4. The organization and presentation lack clarity, making the central contributions difficult to identify and understand efficiently.

  1. Figure 2 lacks an explicit reference and a step-by-step explanation in the main text. Although the figure has a caption, the method description does not walk the reader through its inputs, training stages, and online workflow.
  2. The method description emphasizes formal definitions without adequately explaining the core contribution. Although the losses are specified, the paper lacks a complete worked example showing the correspondence among conversational history, positive and negative cues, training prompts, and the online query.
  3. The experimental configurations are not explained consistently. For settings such as Card, Summary, Parsed reference, Gold query, and Guarded PSS, the information sources, budgets, and experimental roles are not presented in a unified manner, making their meanings and differences difficult to understand.
  4. The presentation switches frequently among datasets and protocols. The original PSS-320 dataset, public conversational datasets, UltraDomain, newly introduced carrier datasets, and multi-record experiments are interleaved. The two similarly named datasets containing 320 sessions are particularly easy to confuse.
  5. The research questions lack a clear organizing thread. RQ1 covers retrieval, question answering, domain transfer, retrieval-backend transfer, and online workload, while RQ3 covers hyperparameters, the contribution of history, carrier selection, and multi-record access. Moreover, nearly every paragraph in the experimental section provides only a brief statement followed by an appendix reference, making it difficult to obtain sufficient substantive information from the main text.
  6. The main paper is not sufficiently self-contained, and appendix references are fragmented. For example, §3.2 directs readers to Appendices B, I, J, L, M, and S for dataset, baseline, prompt, and statistical details, while information about the method and its costs is further distributed across Appendices J, R, and V.5.
Questions for authors

Q1. What is the concrete application scenario? Asking the user for clarification is a natural solution to the ambiguity illustrated in Figure 1. In which practical scenarios is such interaction infeasible or sufficiently costly to motivate automatic intent recovery through PSS?

Q2. What concrete benefit does parametric storage offer over textual storage? In Figure 3, Parsed reference achieves 96.95%, slightly exceeding PSS at 96.56%. Given that retrieval cues are already available during training, under what conditions is training a LoRA state more valuable than directly storing and using those cues as text?

Q3. How is the state updated as the conversation evolves, and how is cost efficiency maintained? When users introduce new information, revise their goals, or revoke previously specified configurations, how does PSS update its state? Could the authors provide a cost analysis that includes repeated updates and identifies the level of reuse required to obtain a practical benefit?

Q4. What is the core technical innovation relative to existing memory methods? Compared with MemoRAG, parametric memory, and conversational query rewriting approaches, what distinctive capability does PSS introduce beyond changes to the storage format and invocation workflow? Which experiments directly validate that capability?

Q5. Can the authors provide fair comparisons with the closest related methods? Experiments across retrieval backends primarily demonstrate compatibility. Could MemoRAG or other representative memory methods be evaluated under the same task, available information, and resource budget? If such comparisons are infeasible, please explain the specific reasons.

Q6. Can the main paper include the essential information needed to understand the method and experiments? A complete worked example linking conversational history, positive and negative cues, training, and online querying would be helpful. The paper should also consolidate the information sources and budgets of the baselines and clarify the relationships among datasets, protocols, and research questions, reducing dependence on scattered appendix sections.

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

This paper addresses missing retrieval intent in long LLM conversations. Standard RAG may embed only the current query, leaving out essential context such as project names or configurations, while retaining or repeatedly processing long histories can be costly. The authors propose a session-specific LoRA adapter trained offline with the base model frozen. Training has two stages: next-token prediction over the conversation history, followed by retrieval-clue training using positive and negative clues. Positive clues are constructed from controlled session fields in synthetic data or extracted from explicitly marked active requests in the public-data evaluation. At query time, the adapter generates a compact search clue for an external retriever. It is then unloaded, and the base model answers using the retrieved evidence.

Critical strengths
  1. The question is valuable: the Subsequent queries in a conversation often lack the context necessary for effective retrieval. And the author try to solve this in parameter perspective.
Critical weaknesses
  1. PSS addresses missing intent information in the current request, but its training already requires correct retrieval clues. In practice, identifying the intended request from complex conversations and constructing reliable clues are major challenges. The paper mainly evaluates how information is stored and used once this supervision is available, without sufficiently evaluating clue extraction or the impact of extraction errors.
  2. The added value of parametric memory remains unclear. Storing clues directly as text also avoids replaying the full history and performs better in some experiments. Although the authors acknowledge these limitations, the paper does not clearly establish when the benefits justify per-session training, storage, and online readout costs. The cost of updating states as conversations evolve is also not evaluated.
Questions for authors
  1. The core question or the contribution the author trying to make is unclear, is the "identification of unknown intents from natural conversations," or the "compilation of conversational information into parametric states"?
  2. Since the author thinks it’s a trade-off for the question of querying correctly, the answer is still not clear, when should we choose PSS, not other methods?
  3. Whether the benefits gained justify the storage and training costs incurred for each session. Is it truly necessary to perform such frequent storage and training updates?
  4. How would positive and negative retrieval clues be obtained for training in real-world conversations without explicit annotations, and how sensitive is PSS to errors in extracting these clues?
Reviewer_f7tNOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

Summary:

The paper proposes a reusable session-memory layer to handle underspecified queries. This layer encodes conversation history into session-specific low-rank updates to a frozen language model, enabling it to generate retrieval clues that recover missing context from the conversation. The authors demonstrate that these clues substantially improve retrieval quality compared with ambiguous queries and alternative memory baselines, potentially benefiting downstream tasks that rely on the retrieved evidence. They also introduce an evaluation framework to isolate the contribution of session-specific memory to retrieval performance.

Critical strengths

Strengths:

  1. I believe the presented idea is not practical yet and might not replace the current approaches. Yet I really like the idea and believe it might inspire new ideas.

  2. The experimental numbers are strong.

Critical weaknesses

Weakness/Questions:

  1. C_{theta} was used in Equation 1 before it was defined.

  2. The paper is a difficult to read, I don’t still understand what is C_{theta}, in particular what is writeable layer-projection pair means. I would like section 2.2 and 2.3 to be elaborated a bit more.

  3. When learning the low-rank update for each session, what are the queries, positive causes and negative clues used? Would you please elaborate on this step a bit more. How does the training data looks like? Does the training data include underspecified query, positive clues and negative clues? How do we have access to these clues during the training process.

  4. I still doubt the practicality of the proposed approach. Storing low rank updates for each session seems expensive, and looks like even in the best case can only be done for SLMs. Can authors comment about the feasibility of the approach for SLMs and beyond SLMs.

  5. Why didn’t you learn the low rank updates for the retrieval model itself \phi instead of the language model \theta? Can you please comment on this. Retrieval models are generally of small size, storing LoRA updates for retrieval model seems more feasible to me.

  6. Would you please provide some examples of underspecified queries that we hope to see in practice. I am interested to know the categories of underspecified queries according to authors.

  7. Would authors please comment on whether the proposed approach would replace the full-context reader approach? I am aware the full-content reader is expensive, but I am interested to know if the learned retrieval clues is as good as full-context reader approach in all the scenarios.

Questions for authors

See Weaknesses. I am willing to increase my score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.