acceptodds
Under review as a conference paper at ICLR 2027

COMPUTED-KEY HANDOFFS: MULTI-HOP LOOKUP IN HYBRID LANGUAGE MODELS

Abstract

A two-hop question makes a language model retrieve twice: once to produce a bridge entity, and once to retrieve with it. Transformers resolve that entity in their early layers. Hybrid models replace most of the attention with a recurrent state, so where the key becomes usable, and where it is spent on an answer, is not known. Answering "Where is the ball?" from "the ball belongs to Alice" and "Alice is in the garden" turns on Alice, a name the question does not mention. A model could reach garden through Alice, the key it computed, or through where the fact sits in the prompt. Across twelve models, dense and hybrid, we move a hidden state from one story into another where the two routes lead to different places, and read off which place the model gives. An attention layer converts the key in every model we tested, and crossing that layer removes its usable effect, all of it where no attention follows. In sequential hybrids this makes the answer a handoff: recurrent layers carry the key forward, and attention spends it. Retrieval does not always stop there. Writing a different fact into the memory of the recurrent layers that follow the last attention layer can move the answer toward that fact, multiplying its odds by 1.3 to 2.7, and which hybrids do this is not settled by their architecture or training. That read is addressed by the key, not by position: recurrent state in a hybrid is not only a carrier, but a memory that later layers can query.

Primary area:
interpretability and explainable AI
Keywords:
Mechanistic Interpretability, Hybrid Models, Retrieval

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_6P66OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

The paper studies how two-hop questions are processed in hybrid LMs (sequential and parallel) and transformer-only LMs. Activations are patched from a donor to a recipient prompt at every layer, and the logits of three candidates are tracked: (1) the recipient's value for the donor's intermediate key, (2) the recipient's value at the position of the donor's relevant fact, and (3) the donor's final answer. Each effect is normalized by subtracting a reference candidate's effect. Tracking these effects across layers shows where the intermediate key becomes available and where it is converted into an answer. The authors find that recurrent layers can carry the intermediate key to attention layers, which use it to retrieve the answer; in some (but not all) hybrids, recurrent layers also retrieve from their own memory after the final attention layer.

Critical strengths

[S1] The paper extends prior interpretability methods to hybrid architectures, showing how the intermediate key is carried through recurrent layers and used by attention layers.

[S2] The methodology is closely related to Gur-Arieh et al. (2026), which uses counterfactual interventions to distinguish different retrieval mechanisms. However, this paper extends the approach to two-hop retrieval, where the intermediate key must first be computed from the context. By independently varying the intermediate key and the position of its associated fact, the proposed setup distinguishes retrieval based on the computed key from positional retrieval and direct transfer of the donor's answer. This provides a useful way to track the intermediate key's causal effect across model layers.

[S3] The paper highlights differences in how hybrid and transformer-only models perform two-hop retrieval, with hybrid models concentrating retrieval in fewer attention layers and some additionally using recurrent memory. These findings provide insights into different retrieval strategies across architectures, although they do not establish a clear advantage for either architecture.

Critical weaknesses

[W1] The motivation is unclear. While the application to hybrid models appears novel, novelty alone does not sufficiently establish the importance of a problem. In particular, the introduction and abstract could better explain why understanding two-hop retrieval is important and why studying this process specifically in hybrid LMs is relevant. The significance of identifying where a key is computed, carried, and used could also be made more explicit.

[W2] Certain terminology and the paper's organization make the paper difficult to follow. Several important concepts are introduced before they are sufficiently explained, which makes the paper difficult to follow, especially in the abstract and Sections 1–2:

  • Bridge entity vs. retrieval key. The abstract describes the model as producing a "bridge entity", but subsequently introduces the term "key" (line 014) without clearly explaining its meaning. The introduction clarifies that this is a retrieval key, but the relationship between the terms could be established earlier and used more consistently.
  • Donor and recipient. Section 2 introduces donor and recipient prompts before sufficiently describing the tasks and experimental setup. Consequently, the examples are difficult to interpret on first reading (e.g., it is unclear what "row 14" means in line 090). Introducing at least the task setup before describing the intervention could substantially improve readability.
  • Terminological overlap. Terms such as "reference" and the different answer candidates require several reads to be understood. An explicit definition to the notations would help (especially for reference).
  • More suggestions (not as critical) about terminology can be found below.

[W3] While the paper specifies that the recipient's earlier-token computations remain unchanged during hidden state transplantation (Section 2, lines 098–107), and later clarifies that recurrent layers combine the injected residual with memory computed from the recipient's unmodified prefix (Section 3.3, lines 255–258), the explanations feel a bit scattered across sections and could be clarified much earlier since this is the primary novelty of the paper. In particular, it would be helpful to clarify how the intervention operates in sequential versus parallel hybrid architectures, where attention and recurrence are combined within the same layer.

[W4] The first three paragraphs of Section 2 do not clearly distinguish established techniques, such as donor–recipient activation patching and logit differences, from the paper's novel contributions. Similarly, the addition of a reference candidate is presented as a contribution, although its novelty is unclear. While the later "Relation to Prior Work" paragraph is well written, clarifying these distinctions for more core methods earlier would help readers better understand the methodological contributions.

I would consider revisiting my score if the authors propose clear steps to address the critical weaknesses.

Additional comments and suggestions:

Figures/tables

  • Fig 1: See questions above.
  • Fig 2: Label models as sequential hybrid, parallel hybrid, or transformer (as in Fig 3). Define balanced, key-dominant, and position-dominant ideally in caption.
  • Table 1: State the units of the causal-scrubbing results (percentages?).

Writing

  • Line 047: "Assay" and "crossed assay" are unusual in ML. Define in footnote or use standard terms.
  • Line 049: The three predicted answers are not evident from Fig 1(b) as stated in Introduction. State them explicitly, along with the reference/recipient candidate, in ideally both text and figure.
  • Line 105: A positive logit difference need not mean a meaningful probability change. This may need to be discussed.
  • Line 115: The notation B_i, B_j, B_k, A is hard to remember; use more intuitive names or add a compact table.
  • Line 119: Give an example of the arrow task. It is hard to understand what it is from this description alone. Or maybe add a forward-pointer to examples in Appendix.
  • Lines 130–132: This reads as if prior work already tested hybrids and this paper makes minor improvements. To my knowledge it did not, so this undersells the contribution of this paper. Revisiting can help position the paper better.
  • Line 216: "Step" is unclear at first. Also, isn't the reported alignment expected by definition, given how the largest conversion step is defined?
  • Line 216: Define "profile" (e.g., "23 profiles"). This term is often used for different setups but it is not intuitive to the reader. Maybe use "setup" or "setting" instead?
Questions for authors

See clarifications requested in Weaknesses

3 comments

Thank you for the detailed review. You asked for clear steps on the critical weaknesses; each step below is already made in the revised paper, and we give its location.

W1. Motivation.

  • Why two hops. The Introduction now opens with the reason: reasoning over a context often means retrieving with something the model itself computed. This is a common pattern: question-answering benchmarks from bAbI to HotpotQA are built on it, and language models often answer each hop of such a question yet fail to compose them (Press et al., 2023). Following a chain of facts or tracing a variable through code repeats the step, and two hops is its smallest instance.
  • Why hybrids. The second paragraph states the design question. Most hybrids keep attention in a minority of their layers because a recurrent layer is cheaper to run, so how many attention layers to keep and where to place them is a choice. Prior work shows that retrieval in hybrids depends on the attention that remains, but not which step of a two-hop lookup needs attention and which a recurrent layer can perform.
  • Why locating the key matters. The same paragraph says what the division tells us: how much attention multi-step retrieval needs and where it must sit, and where inside the model an intermediate result can be read or edited. A new "Implications" paragraph in the Discussion draws three consequences: for hybrid design, for inspecting a model, and for compression (Nano-9B-v2 reads from recurrent memory where its parent does not, so a pruned or distilled hybrid should be checked for how it retrieves, not only whether it does).
  • Abstract. It now states the gap directly: both lookups have been located in transformers, and in hybrids, which replace most of the attention with a recurrent state, where the key becomes usable and where it is converted is not known.

W2. Terminology and organization.

  • Bridge entity versus key. "Bridge entity" is gone from the paper. The abstract and the Introduction introduce one term, the key, through the example: the first lookup finds Alice, and the second uses Alice as a key to find the garden.
  • Donor and recipient before the task. Section 2 now starts with the task and defines a fact's row as its line number, then describes the intervention, then the candidates. Both terms are also defined at first use in the Introduction.
  • Reference and the candidate answers. Section 2 ("Three accounts of the patched state") names each predicted answer on the running example: office if the state carries the donor's answer, garden if it carries the key, kitchen if it points to the row, and hall, predicted by none, as the reference. Figure 1b labels the same four, and Appendix A has a notation table.

W3. Where the intervention is explained. The explanation is now in one place, Section 2 ("Counterfactual patching"), before any result:

  • only the final token's state is replaced;
  • everything the recipient computed over its earlier tokens stays its own, which is the keys and values its attention layers read and the memory from which a recurrent layer computes its output;
  • in a sequential hybrid each later layer reads the recipient's context through attention or through recurrence, and in a parallel hybrid both branches of every layer read the same patched state, attention over the recipient's earlier tokens and recurrence from the recipient's memory;
  • the paragraph ends by listing what each section replaces: the residual state (Section 3), a layer's or a head's output (Section 4), and the recurrent memory after the last attention layer (Section 5, sequential hybrids only).

W4. Established techniques versus our contribution.

  • Section 2 now presents counterfactual patching as the standard tool of prior work on binding and retrieval, with citations, and maps our donor and recipient onto the counterfactual and original inputs of Gur-Arieh et al.
  • Scoring by a logit difference against a baseline candidate is introduced with "As is standard" and citations. The reference candidate is no longer presented as a contribution.
  • What is new is stated in the first contribution bullet and in "Relation to prior work": the counterfactual design is extended from a key the question names to a key the model computes, which requires varying the key and the position of its fact independently, and it is applied to hybrid models at every layer.

Figures and tables.

  • Figure 1. Both captions are rewritten. Panel (a) is described as left and right halves; panel (b) names what each tag predicts and which state moves.
  • Figure 2. Each panel is labelled sequential hybrid or dense transformer, and the caption defines key-dominant, position-dominant and balanced.
  • Table 1. The scrubbing rows are marked (x), and the caption says they are multiplicative factors, not percentages, with x1 meaning no change.

Writing.

  • Line 047, "assay". Removed from the main text; we say counterfactual patching and prompt pairs.
  • Line 049, the three predicted answers. Stated explicitly in Figure 1b's labels and caption and in Section 2; the Introduction points to the figure.
  • Line 105, logit difference versus probability. Section 2 now says what the number means: a candidate's logit minus the reference's is the log of their probability ratio, so an effect of d logits multiplies that ratio by e^d. Because a ratio can grow while both probabilities stay small, we also report how often the key's candidate is the model's top choice among the seven scored answers (Table 1: 86%, 71% and 67% in the rebinding test, against a chance level of 14%).
  • Line 115, notation. B_i, B_j, B_k are replaced by B_key, B_pos and B_ref, with A for the donor's answer.
  • Line 119, the arrow task. Section 2 gives an example (Key NOG -> Value CFG, Key CFG -> Value 39, and the question "Starting from NOG, follow mappings until reaching a number"), and Appendix A shows a complete prompt pair.
  • Lines 130–132, prior work and hybrids. "Relation to prior work" now says that both prior studies examine transformers and a key the question names, and that we bring their design to hybrid models and extend it to a computed key.
  • Line 216, "step", and whether the alignment holds by definition. The conversion step of a layer is defined as the increase in the donor-answer effect across it, and the resolving layer is the layer with the largest step. The alignment is not built in: the resolving layer is defined by the donor-answer effect and the peak by the key effect, and the rise spans three to twelve layers. A largest step placed at random within that span would meet the key peak in two to three of the nine model families; it does in all nine (Section 3.2 and Appendix C). The alignment is closest to a tie in Granite-H, whose key effect is nearly flat over the five layers before the resolving layer (margin 0.01 logits). In one of six new sweeps with other list lengths, run for Reviewer B9Na, its peak falls at the start of that plateau, statistically tied with the resolving layer's input (Appendix G).
  • Line 216, "profile". Removed from the main text; we say layer sweep.

Dear Reviewer 6P66,

A short followup since the discussion period is ending. You said you would revisit your score if we had clear steps for W1–W4. We made all four changes in the revised PDF and our earlier reply goes through them point by point.

In short:

  • Motivation. The introduction now says why this matters up front. Retrieving with a key the model computed itself is the core step of multi-hop reasoning. Hybrids keep only a few attention layers, so which step needs attention is a real design question.

  • Section 2. It is rebuilt: the task comes first, then the intervention, then what each candidate answer means. Donor, recipient and the reference are defined before they are used, and one paragraph says exactly what the patch replaces and what it leaves alone.

  • Contribution. We now say plainly what is standard (activation patching, logit differences) and what is ours (a key the model computes and hybrid models).

If you get a chance to look at the revision, we'd really appreciate an updated assessment. If anything is still unclear, tell us and we'll fix it.

Reviewer_B9NaOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The article examines sequential hybrid LLMs when responding to so‑called multi‑hop questions and how and where the intermediate key is calculated to answer the question.

  • Twelve models and 200 stories with 17 facts are considered.
  • Two stories are used for the study: a donor and a recepient — to analyze the appearance of the computed key.
  • Various effects are used: the key effect, the position effect, and the donor’s response effect. In particular, it is shown that retrieval occurs after attention.
Critical strengths
  • A methodology has been proposed for understanding the retrieval process in sequential hybrid LLMs.
Critical weaknesses
  • A small dataset: 200 prompts, no variability in the number of facts — only 17.
  • No examples of practical application of these results.
Questions for authors
  • What happens when the number of facts changes or when the input data format changes, for example, when the query length is increased several times?
  • How robust are the results to such changes?
  • Why were 17 facts specifically chosen?
  • Does the same pattern persist when the length of the input matrices is increased?
1 comment

Thank you for the review and the questions. We ran the experiment you ask about, the layer sweep with other numbers of facts, and report it below. References are to the revised paper.

W1. Size of the dataset. The 200 is the number of independent fact lists per experiment, not the number of measurements. Each fact list yields nine donor–recipient prompt pairs (three donor keys crossed with three positions of the donor's fact), so every model in every experiment is run on 1,800 pairs, and the 1,200 pairs in which the key and the position predict different answers are scored. The layer sweep repeats this at every layer of twelve models on two tasks. Confidence intervals resample whole fact lists, so the uncertainty we report is that of 200 independent lists: the key effect at the resolving layer is about 1.1 logits with an interval half-width of about 0.12 (Section 2, "Pairs"; Table 5 in Appendix A gives the sample of every experiment).

Q3. Why 17 facts. The design itself needs 14 facts per list:

  • nine for the three alternative keys: for each key, a first-hop fact, the key's own fact, and a fact that keeps the key's name in the recipient;
  • three that keep every candidate answer present in both prompts, so that none can win merely by being present;
  • two for the recipient's own two-hop question.

The other three are distractors that lead nowhere. Seventeen is the size of the fact-list generator the design was built on, and we did not tune it. The revision says this in Appendix A.

Q1, Q2, Q4. Changing the number of facts (new experiment). We repeated the layer sweep with 14, 26 and 34 facts per list (the paper uses 17) on Nemotron-H and Granite-H, with 200 new fact lists per length; only distractor facts are added or removed. The results are the same at every length: at the last attention layer the key effect is well above the position effect, it is gone after that layer, and the conversion is concentrated there (C = 0.80 to 0.98).

14 facts17 (paper)2634
Nemotron-H, key effect κ1.151.131.000.95
Nemotron-H, position effect π0.380.360.290.29
Granite-H, key effect κ1.061.071.181.08
Granite-H, position effect π0.350.250.250.17

Logits at the input of the last attention layer; interval half-widths 0.08 to 0.14. Details are in Appendix G (Table 34).

Q1, Q2. Changing the format and the length of the input (in the paper).

  • A longer input with the same facts. Filler sentences after every fact make the prompt 2.3 times as long in words. Nemotron-H stays key-dominant at its last attention layer (κ = 0.81 ± 0.12, π = 0.36 ± 0.09), and the conversion stays concentrated there (C = 0.90) (Appendix G, Table 33).
  • Other formats. A second prompt format (Appendix C), word-valued answers in place of numbers, a one-hop task with no key to compute, and a bAbI-style task with typed facts and location answers (Appendix G, Table 33). Every hybrid run that passes the accuracy gate is key-dominant at the last attention layer.
  • The memory experiment. Section 5 uses the word task with 14 and with 17 facts. Six hybrids were run on both, and the outcome is the same for each: four read from recurrent memory at both lengths and two at neither (Appendix F, Table 30).

The limits are these. Under filler sentences Granite-H's recipient accuracy is 59%, below our 60% gate, so we do not interpret that run. Among the dense models, Qwen2.5-7B is balanced between key and position under filler sentences, as it is on the main task. We have not tested lists longer than 34 facts or natural text, and Section 6 ("Scope") says so.

W2. Practical application. The paper is an analysis of how these models retrieve. The revision adds an "Implications" paragraph to the Discussion (Section 6) that says what the results can be used for:

  • Designing hybrids. In multi-step retrieval, one or two late attention layers convert the key through a few heads, and recurrent layers carry it to them. This is evidence on how many attention layers such a model needs for this step and where they must sit.
  • Inspecting or editing a model. The input of the resolving layer is where an intermediate result is strongest, so it is the place to read it or change it. Section 4.1 does this: replacing the output of recurrent layers before that layer changes which key the model looks up.
  • Compression. Nano-9B-v2 reads from recurrent memory where its parent model does not, so a pruned or distilled hybrid should be checked for how it retrieves, not only whether it does.
Reviewer_yoRrOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper explores how the bridge entity (or key) in a multi-hop QA task is computed and handed off across the layers in hybrid language models. While dense transformers apply attention at every layer, hybrid models interleave attention and recurrent layers. Prior work attributes retrieval in dense transformers and hybrid models to the attention layers. This work attempts to explain the role of recurrent layers in hybrid models, specifically on multi-hop QA task. The authors find that recurrent layers are crucial for propagating the key to the attention layers which then perform retrieval of the answer based on the key. In some hybrid models, the recurrent layers that follow the final attention layer can also perform the key-based answer retrieval.

Critical strengths

S1. Good experimental design: The interventions and evaluations effectively characterize the role of recurrent layers in passing along information.from the patched activation S2. Comprehensive coverage: The paper covers a range of models and tasks and explores multiple competing explanations in a methodical manner.

Critical weaknesses

W1. Lack of clarity: Frequently used terminology crucial to describing the experiments (for example, donor, recipient) lacks precise definitions, leaving the reader to infer what was actually meant. W2. Contributions with respect to prior work: The paper does not clearly state how this work extends prior knowledge about recurrence in hybrid models. For instance, the authors mention that Afendulev3 et al (2026) explores the role of attention and recurrence in hybrid models. Beyond the mechanism used for ablation (substitution vs removal), it is unclear how the findings presented by the authors support/disprove/extend those in prior work.

Questions for authors

Q1. Please justify the selection of the reference candidate. In the running example, 'hall' is used as the reference entity because it is not one of the three candidates of interest (donor's answer, key-based, row-based). Is 'hall' the only entity that satisfies that requirement? It would be good to have a robustness check with the reference being an entity that does not appear in either contexts if that makes sense. Q2. Can the authors confirm that the donor's answer does not appear anywhere in the recipient's context? Q3. What happens on a "null donor" (one which does not contain the key)? Q4. How does activation patching of a recurrence layer change logit assignments from patching an attention layer?

5 comments

Thank you for the review and the four questions. We ran the null-donor control you ask for in Q3 and report it below. References are to the revised paper.

W1. Donor and recipient. The revision defines both at first use in the Introduction and formally in Section 2: the donor is the prompt a hidden state is taken from, the recipient the prompt it is written into. Section 2 now follows the running example of Figure 1b, and Appendix A has a notation table.

W2. Relation to Afendulev et al. (2026). The two papers ask different questions. They remove or swap a hybrid's whole attention cache or recurrent state on one-hop lookups and find that the recurrent state alone cannot return a stored value. We ask where across depth a key the model computes itself is carried and where it becomes an answer, which needs a two-hop task and interventions on one layer at a time. The results fit together:

  1. Both find that attention performs the retrieval. In all twelve of our models the key is converted at an attention layer, and in the three hybrids where we test heads, a few heads of that layer do the reading (Sections 3 and 4.2).
  2. We add what recurrence does while attention is in place: it delivers the key. Swapping the output of single recurrent layers before the converting attention layer changes which key is looked up (Section 4.1).
  3. We qualify "recurrence cannot retrieve". With attention left in place, writing a different fact into the recurrent memory after the last attention layer moves the answer toward that fact in four of seven hybrids (Section 5).

Their paper appeared on arXiv on 3 September 2026 (arXiv:2609.04434), shortly before our submission. The revision discusses it in Section 2 and in the Discussion.

Q1. Choice of the reference candidate. The reference is a role, not a fixed entity. On each pair the donor's key predicts one of the recipient's three places, the row of the donor's fact predicts a second, and the third, predicted by neither, is the reference. Across the pairs of one fact list every place takes every role, so a preference for one place cannot pass for key-based retrieval.

We also computed an alternative reference: the mean of the two other scored candidates that no account predicts. Key effect at the input of the resolving layer, in logits:

ModelPaper's referenceAlternative reference
Granite-H1.071.24
Nemotron-H1.131.29
Qwen2.5-7B0.901.17
Nano1.061.35

200 fact lists per model; interval half-widths 0.11 to 0.20. The key effect is larger under the alternative in every model, so the paper's choice is the conservative one and the contrast that separates key from position, κ − π, uses no reference at all. We did not use an entity absent from both prompts, because its score would also reflect the model's preference for tokens that appear in its context.

Q2. Does the donor's answer appear in the recipient's context? Yes, in every pair, by design: every candidate appears in both prompts, so none can win or lose merely by being present (Section 2). In the recipient it is bound to a key that no fact points to (checked on all 1,800 pairs). Figure 1b draws only the rows the predictions use; Appendix A shows the complete pair. A state holding the answer and one saying "find office and repeat it" give the same output, and we do not separate them (Appendix B). The key effect does not depend on this, since it compares two of the recipient's own bindings.

Q3. Null donor (new control). The null donor is the original donor prompt with its first hop redirected to a key that has no fact, so it never computes a key. On the same 200 fact lists, at the input of the resolving layer:

Modelκ, original → null donorπ, original → null donor
Granite-H1.07 → 0.030.25 → 0.00
Nemotron-H1.13 → 0.020.36 → 0.01
Qwen2.5-7B0.90 → −0.010.77 → −0.04
Nano1.06 → 0.060.16 → −0.04

Every null-donor interval contains zero. Both effects therefore require the donor to have computed its key, and neither is a by-product of patching a foreign state into the recipient.

Q4. Patching a recurrent layer versus an attention layer. The sweep applies the same patch at every layer's input, so the difference lies in what the later layers do with it. Patched before the converting attention layer, the state carries a key: at the input of Nemotron-H's L40 the key effect is 1.13 logits and the donor-answer effect −0.61. Patched one layer later, it carries the answer: −0.62 and 3.49. Attention layers account for 95% of the rise in the donor-answer effect, although they are 4 of Nemotron-H's 28 attention and recurrent layers.

Thank you for the clarifications. However, I'd prefer to keep my score as I found the presentation quite dense and the writing difficult to follow.

Dear Reviewer yoRr,

Thanks for reading our reply. We agree that presentation was the main problem. That's why the revision rewrites the abstract, the introduction and Section 2 rather than patching them.

Your two weaknesses are now handled in the main text. Donor and recipient are defined at first use with the task introduced before the method (Section 2). The comparison with Afendulev et al. is in Section 2 and the Discussion. The four questions are answered there too, with two new controls: the null donor and an alternative reference.

In case you saw the original version, we'd be grateful if you could take a look at the revised PDF before the discussion closes. We think it reads very differently now.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.