acceptodds
Under review as a conference paper at ICLR 2027

Identity Supervision for Readable Intermediate States

Abstract

How do looped transformers organize intermediate states that new downstream computations can use across source views? We investigate this question through controlled studies of modular addition and two-hop relational composition, comparing weight-tied and untied transformers at matched effective depth. Our analysis combines temporary supervision interventions, state diagnostics across checkpoints, and tests of new readers and retained downstream computation. These tests distinguish identity information readable from a state from its usefulness to a specified attribute reader, and examine the roles of supervision location, gradient exposure and source variation. The tested identities are known, and the later reader generalizes their observed labels across source views. In relational composition, direct identity supervision yields gains of 5.2–7.6 percentage points across seven settings under a common bounded gradient rule, while several comparisons with fixed calibration remain unresolved. In modular addition, near-perfect affine identity reading coexists with substantially weaker random-attribute reading. Depth-by-strength comparisons qualify a pure supervision-distance explanation, and matching identity grouping reveals different costs for source-factor reading. Together, these results characterize concrete dependencies and limits of intermediate-state use rather than a universal benefit of weight sharing. The study connects source-view generalization to analysis of internal states while distinguishing reader-specific accessibility from stronger claims about computational mechanisms.

Primary area:
interpretability and explainable AI
Keywords:
looped transformers, intermediate representations, representation probe

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

This paper studies when intermediate states of looped transformers form useful interfaces for downstream computation. The key setting is source-view generalization: multiple sources express the same known intermediate identity, and a newly trained reader must transfer an observed identity-dependent label from a few source views to other views of the same identity.

The main intervention, Direct-ID, temporarily attaches a fresh affine identity classifier directly to the exposed intermediate state during source-model training. The classifier is later discarded, the source model is frozen, and new downstream readers are trained from scratch. The authors compare Direct-ID with native training, additional atomic cross-entropy, representation contraction, supervised contrastive learning, and several control objectives.

Experiments cover two-hop relational composition and decomposed modular addition, using both weight-tied recurrent transformers and untied twelve-block models. The study additionally varies supervision depth, auxiliary-gradient strength, intervention timing, source-view coverage, and reader type. Direct-ID substantially improves linear identity readability and source-view generalization. Under the bounded gradient rule, it improves the graph fresh-reader endpoint by 5.2--7.6 percentage points across seven tested settings. The experiments also show that near-perfect identity decoding does not imply equally strong linear decoding of arbitrary identity attributes, and that representations with similar identity grouping can retain different amounts of source-specific information. Overall, the paper argues that the quality of an intermediate interface depends jointly on identity accessibility, training exposure, and the distinctions required by the downstream reader.

Critical strengths
  1. The experimental controls are unusually careful. The paper explicitly separates supervision location, readout location, gradient magnitude, parameter sharing, and intervention timing. In particular, the supervision-depth scan over s = {0, 2, 4, 8}, together with shared and untied architectures and fixed versus bounded gradient rules, provides a substantially stronger test of the apparent supervision-distance effect than a simple auxiliary-loss comparison. I also appreciate that the paper limits its mechanistic claims and does not attribute the remaining depth effect directly to Jacobian attenuation.

  2. The paper makes a useful distinction between several notions of representation quality. Identity probes, random identity-attribute readers, retained computation, nearest-view retrieval, and source-factor readers measure meaningfully different properties. The geometry-matched comparison is particularly informative: Direct-ID and Contraction can have similar W/B while preserving substantially different amounts of source-specific information. This provides concrete evidence that a single class-separation statistic is insufficient to characterize the usefulness of an intermediate state.

Critical weaknesses
  1. The external validity is limited. Both task families are synthetic and highly controlled. Every evaluated identity is already known, every downstream identity-attribute label has been observed on training views, and the processing interface is a single-token state. The experiments therefore study source-view transfer for known identities, which remains substantially removed from pretrained multi-token interfaces, unseen source distributions, or unseen identities. This limits the strength of the broader representation-reuse claim.

  2. The gradient-budget control does not actually equalize the realized auxiliary pressure across supervision depths. In the shared graph model, the bounded controller attains mean auxiliary/native norm ratios of 0.088 at s = 0 and 0.059 at s = 8, while the coefficient cap is active for 32.7% and 70.3% of updates, respectively. Consequently, the remaining depth gap still mixes supervision location with cumulative gradient magnitude, gradient direction, and optimizer history. The paper acknowledges this limitation, but it makes the mechanistic interpretation of the supervision-depth result suggestive rather than conclusive.

p.s.: The discussion section seems to be too long. More weights are suggested to be put in the explanation and analysis of the results. (This reflects my personal preferences only but does not form the basis for any scoring results.)

Questions for authors

Overall, the current illustrated experiments are excellent in general. However, there are two questions below. My confidence score will be raised if the questions could be fully answered.

Q1: Could you add an experiment that more closely matches the realized cumulative gradient exposure across supervision depths, for example by increasing c_max or using a controller that matches the time-integrated auxiliary/native gradient-norm ratio?

Because the s = 8 condition hits the coefficient cap very frequently, it is currently difficult to quantify how much of the residual s = 0 versus s = 8 gap in Table 9 is attributable specifically to supervision location.

Q2: Most transfer evaluations train a new reader. In the Discussion section, you describe a stronger functional-transfer test in which the native continuation is held fixed and receives a state taken from another source context. Could you include this experiment in the current synthetic setting?

It would substantially strengthen the claim that the intermediate state is directly reusable by downstream computation and would separate easy readout relearning from compatibility with an already learned continuation.

3 comments
AuthorsOpenReview ↗

Response to Reviewer CMYn

Thank you for the careful reading and the two concrete requests. Both are now experiments (E1, E2), summarized in Sections 4.2 and 4.4 of the revised PDF, with details in Appendix E.

Q1 and W2, matching realized exposure (E1). We add three arms on the control cohort of Table 9.

  • Cap raise. With c_max = 30 the s = 8 realized ratio in the looped graph model rises from 0.059 to 0.096 (cap active on 28% of updates). With c_max = 100 the cap never binds, and both depths attain the nominal 0.100. The requested coefficient then peaks near 60×λ₀ late in the pulse, so we keep c_max = 10.
  • Matching from the other side. Lower γ at s = 0 until its pulse-mean realized ratio equals the same cell's s = 8 value (γ′ ≈ 0.061–0.067). γ′ starts from the linear prediction 0.1·r̄₈/r̄₀ and is refined with pulse-only reruns until the ratios agree within 0.001; no 200k outcome is consulted. A time-integrated rule is unnecessary, because with c_max = 100 the realized ratio equals γ at every update.
  • Shadow heads. We record the cosine between the s = 0 and s = 8 processing-route gradients on identical minibatches.

Results in pp, s = 0 minus s = 8 (Table 13, Figure 3c):

Layoutfixedbounded (c = 10)c = 100matched
graph/looped5.351.911.64 [0.05, 3.23]1.55 [0.13, 2.97]
graph/untied2.991.061.11 [−0.11, 2.32]0.81 [0.06, 1.55]
addition/looped5.412.011.91 [0.72, 3.09]1.50 [−0.01, 3.01]
addition/untied3.733.161.47 [−0.71, 3.65]2.53 [0.58, 4.48]

Direct answer to Q1 (graph, looped): of the 1.91-point gap under the bounded rule, equalizing realized exposure removes 0.36 [−0.07, 0.79] pp, which is not resolved, and 1.55 pp remains at matched exposure. Across layouts this residual is 0.8–2.5 pp; under matching its interval excludes zero in three of four layouts (the fourth just includes zero), and with the raised cap in two. The larger fixed-rule gap is mostly dose: the bounded rule alone removes 64 / 65 / 63 / 15% of it while exposure is still unequal (Table 19). The gradient cosine falls from 0.37 to 0.13 over the pulse, so equal magnitude leaves direction and optimizer history unmatched. Section 4.2 now reads: most of the depth effect is dose, and the residual is associated with location but does not identify Jacobian attenuation.

Q2, functional transfer under the frozen continuation (E2). We freeze each 200k model's decoder, the native continuation of your question, and refit nothing (Section 4.4, Figure 5, Tables 14–15).

  • Recipients and prototypes. For graphs, the reserved source of each selected identity with all its composition-training queries (about 1,700 per task); for addition, all A sources with 32 queries. The prototype is the mean state of the other views of that identity (four training views on graphs, the other three A sources on addition), so it never contains the recipient.
  • Conditions. Prototype substitution C(μ_m, r); foreign-residual transplant; prototype shift toward a donor identity, scored as donor- and recipient-following; norm-matched random shift.

Graph results (Native / Atomic-CE / Direct-ID / Contraction / SupCon):

  • Prototype accepted: 51.4 / 68.0 / 90.0 / 95.0 / 93.7%.
  • Shift, donor-following: 44.7 / 59.6 / 83.8 / 92.1 / 89.9%.
  • Shift, recipient-following: 26.1 / 18.5 / 6.6 / 2.3 / 2.5%.
  • Random shift leaves 78–98% of answers unchanged.

Direct-ID minus Atomic-CE is +22.0 [13.9, 30.1] pp, and +21.5 [18.0, 25.1] in bounded-rule models; addition is similar (94.3 vs 75.3%). Because nothing is refitted, the gain cannot come from easier readout relearning, the distinction you asked us to draw. The native decoder partly re-derives identity from the residual, whereas after identity supervision it relies on the prototype. Contraction states are the most interchangeable, consistent with tighter grouping and larger source-factor loss. This shows interchangeability under one fixed decoder, not algorithm identity.

W1, external validity. We agree. The abstract and Section 1 now state the scope and list what we do not claim. We did not add a pretrained setting in this cycle; Section 5 names structured identities, unseen sources and pretrained multi-token interfaces as the next tests.

Moreover, following your suggestion, the interpretation of each result now sits next to it in Section 4 (e.g., "Why Atomic-CE trails" and "What changing depth changes" in Section 4.2), and the Discussion is organized around three lessons drawn from the results, each related to prior work.

Thank you again for your two concrete requests; both are now addressed in the revised version.

E1 equalizes realized exposure across depths in two ways. Of the 1.91-point bounded-rule gap, equalization removes 0.36 points, leaving 1.55 points. This quantifies how much of the observed gap is attributable to exposure differences rather than location, although the underlying mechanism is not fully resolved. E2 tests the frozen continuation without any refitting (Section 4.4).

We would welcome your any further questions or concerns.

Thank you for your response. E1 addresses my concern about a fair comparison across supervision depths by showing how much of the gap remains when relative gradient strength is matched. E2 answers my second question in the synthetic setting. These results show extra evidence in the findings, especially the evidence for reuse by a fixed decoder. My overall assessment remains consistent with my original score.

Reviewer_bJY6OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors study how different forms of intervention on intermediate layers affect how well the intermediate states can be leveraged to solve the tasks with a new readout. They construct interventions that help learn intermediate results, such as moving similar facts together in latent space or adding a simple readout for the intermediate solution on intermediate layers.

Critical strengths

The paper is very thorough when analysing hypotheses for what helps models do multi-step compositional tasks. It analyses multiple different auxiliary losses, compares the effect sizes of the gradient, and checks for a multitude of potential confounders which strengthens the results.

The authors pose an interesting question: how do looped transformers represent intermediate states? This can help both understand and help improve current architectures.

Critical weaknesses

The abstract poses the question of how loop transformers organise intermediate states. This seems to be very sparsely explored in the main text. Most experiments don’t evaluate representation with a standard loop transformer but add an auxiliary target. The paper is in general, quite hard to follow, and it is not always fully clear to me which numbers are the most relevant. A lot of intervals are part of the main text, making it difficult to understand which ones are most relevant or how strong the effect size is. I might be misunderstanding something, but the results seem, in parts, a bit detached from the abstract. This makes judging the claims difficult.

My understanding is that if you train on the intermediate goal, this will improve the latent representation. And then later, a readout is trained to check if we can find that intermediate representation. This seems cyclical if we train to move similar groups together in representation space or linearly read out an intermediate result, then this should become easy to read out by construction. Secondly, this seems largely influenced by where the loss is applied. Table 10 seems to suggest that less signal arrives the later the additional loss is placed. Given that Atomic-CE performs worse than the others and that the others place their loss on the earlier layer, this seems somewhat confounded. The effect of which supervising signal to use becomes additionally small once the readout is placed at k=8. So it seems this mainly evaluates how early information becomes available. This then questions whether some methods are truly better. What is interesting is that the information stays available although the supervisor signal is only given for a short period of time.

Questions for authors

Did you run any experiments with Contraction and SupCon placed at later layers? Why are we reading at an earlier layer if a later layer seems to provide more information?

7 comments
AuthorsOpenReview ↗

Response to Reviewer bJY6

Thank you for the constructive review. Both questions led to new experiments (E3, E4), W4 is addressed by E1, and all three are integrated into Section 4 of the revised PDF.

W1, abstract versus body. We agree the abstract over-promised a study of how looped transformers organize states. The new abstract states the question (where identity supervision should act and what the state must keep) and four results with numbers. Two clarifications are now in the paper. (i) The standard looped transformer is the Native row of Tables 3–5, 14 and 17: without auxiliary loss it organizes identity only partly at w (identity probe 73.6 / 90.0%, Section 4.1; W/B 1.95 / 2.74, Table 5). (ii) Why a loop: supervision depth then means only the number of further applications of the same block between w and a temporary head. The auxiliary gradient reaches the block through the two processing calls at every depth, in looped and untied models alike; weight sharing adds coupling, which the twelve-block untied model removes (Section 3.1).

W2, readability. Section 4 now opens with a six-row reading guide giving the headline number for each question. Bracketed intervals in the main-text prose drop from about 17 to 2 (the primary estimates); the rest are in tables, with new Table 19 collecting those cited in the text.

W3, "cyclical". The test labels f_r are random partitions never shown to the model, and identity readability does not make them readable: identity probe 99.4% versus attribute probe 44.1% on the same states (Table 3). Atomic-CE supervises identity with the same labels and pairs and trails Direct-ID by 5.3 / 7.6 pp. E5 measures the by-construction part: a head on one coarse partition g makes g readable (+19.7 pp) but yields only 40% / 36% of Direct-ID's gain on unrelated partitions. E2 uses no fitted reader: with the decoder frozen, an identity prototype is accepted in 90.0% (Direct-ID) versus 68.0% (Atomic-CE) of graph queries (Section 4.4).

W4, Table 10 and Atomic-CE. Your reading is right, and Section 4.2 now states it ("Why Atomic-CE trails"). Atomic-CE supervises at depth eight through the trained head and receives the least gradient of the identity heads (realized ratio 0.038, cap active on 81% of updates). Its 6.1-point deficit under the bounded rule splits into three paired steps (Figure 3d, Table 19): +4.18 for a fresh head at the same depth, a step that changes head and exposure together; +1.55 for moving that head to w at matched exposure; and +0.36 for the remaining exposure.

E1 asks how much of the depth effect is dose. Lowering γ at s = 0 to the realized s = 8 level leaves depth gaps of 1.55 / 0.81 / 1.50 / 2.53 pp (graph looped / untied, addition looped / untied); a non-binding cap (both depths at 0.100) leaves 1.64 / 1.11 / 1.91 / 1.47. Most of the fixed-rule gap disappears once both depths receive the target ratio (the bounded rule alone removes 64% of it on graphs), and exact equalization changes the rest little. Because the two depths' gradients point in increasingly different directions (cosine 0.37 → 0.13), we associate the 0.8–2.5 pp residual with location without attributing it to a mechanism.

On k = 8: we agree that the objectives converge there, and Section 4.3 says so. They differ in what the exposed state retains, which is what a new module reads, so we claim a trade-off rather than one objective being better in general (Section 1). On persistence, we agree and now state it as a result: the loss acts for 5k updates, its head is discarded, and the gain is similar when the pulse moves to 40k or 160k (6.7 / 5.9 pp; Table 1).

Q1, Contraction and SupCon at later depths (E3). Both losses are applied after s ∈ {0, 2, 4, 8} consumer calls under the bounded rule, with readers still on w. Moving from s = 0 to 8 costs 1.84 [0.67, 3.00] (Contraction) and 1.55 [0.59, 2.51] (SupCon) pp on graphs, and 2.61 [1.54, 3.69] and 3.41 [2.84, 3.99] on addition, similar in size to the identity head's 1.91 / 2.01 (Figure 3b, Table 16). The bounded-rule depth gap is not specific to the cross-entropy head.

Q2, why read an earlier layer (E4). The question is which state to expose to a new computation; when the original consumer suffices, one should use it. Along the retained consumer, identity-attribute accuracy of Atomic-CE graph states rises from 31.5% to 96.1%, while source-factor accuracy falls from 74.4% to 6.4% (chance 6.25%), and every objective reaches the same values by k = 8 (Figure 4c, Table 17). Source-factor readability is highest at w, so a system that must also answer source-dependent requests has to retain those distinctions somewhere.

Thank you again for your constructive review. We wanted to check whether the new experiments address your questions:

E3 places Contraction and SupCon at later depths and finds depth gaps similar to those observed with the identity head. E4 shows that source factors fall to chance at later exits while identity attributes become more prominent. E1 measures how much of the depth effect can be attributed to gradient dose. Section 4.2 now explicitly addresses your interpretation of Table 10.

If any concerns remain, we would be happy to address them before the discussion period closes.

Thank you for the rewrite of the abstract as well as the main text with the figures. The results and implications are easier to find and more understandable. The additional experiments are also helpful and clear up several of my concerns. E4 is especially interesting as it justifies why we can’t just read late. This trade-off now allows a more informed choice of where to place a supervisory loss. 
I have two remaining concerns.
1) There might be a probe that is able to read out more than any individual layer by concatenating an early and a late layer, even from the Native model, as Native performs well at different layers. For example, large-scale diffusion models such as FLUX.2 read out multiple layers at different depths of the Qwen or Mistral text encoder. They fit quite well as a chained compositional example, e.g., a dog chasing a brown cat, where composing brown with cat would be a subgoal. In comparison, supervised methods would be useful if they outperform concatenation. 
2) The usability heavily depends on knowing the useful information for downstream tasks. This is given in the paper’s setting but is substantially harder in real-world applications where it is not clear which identities are relevant. For example, in the diffusion example, supervising on being able to read out brown cat is non-trivial, nor is it clear whether it is useful or will hurt generalisation on other tasks.

Reviewer_syz8OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The authors mainly study how to make an intermediate hidden state easy to use for new downstream modules. It trains small transformers on two synthetic tasks: two-hop relational lookup and modular addition. They train the model to encode some internal representations which are later used to decode the prediction. This objective is called Direct IDs. After training, the model is frozen, and new readers use new inputs and the authors find that these intermediate representation-based approach improves over the baseline decoder approach. Further, they conduct more analysis on the effect of the location of the supervision, etc.

Critical strengths

The experiments are well controlled such that the comparisons are carefully matched in data exposure, initialization and optimizer state. One of the finding was interesting that two objectives that cluster identities equally well can still differ in how much other information they can keep.

Critical weaknesses

The key findings are largely expected. The authors train a linear identity head on an intermediate state which makes the state linearly "identity" readable. Then, every test label is just a function of that identity, and every test identity was actually seen in training, so the readers mostly have to recognize the identity. This is expected. And the effect of the work is unclear beyond the hardest setting. With four examples per identity, the gain drops to about 2 points. On the original task it is about 0–1 point, and the contrastive objective actually beats Direct-ID on the main metric.

The paper presentation in the current form needs substantial revisions. It is substantially hard to parse through the text to understand what the premise is, what the approach is, etc. And honestly, the overall experiments and goals are fundamentally not that hard to understand. It's just the way they are presented right now. I'd request the authors to consider adding some more descriptive figures to explain the premise and make the draft accessible to new readers.

Questions for authors

NA

2 comments
AuthorsOpenReview ↗

Response to Reviewer syz8

Thank you for the review. The revision changes the presentation where your summary shows it was unclear, and adds two experiments (E2, E5) aimed at the by-construction concern. Numbers refer to the revised PDF.

What we changed in response to the summary. (1) New readers do not see new identities: every tested identity is known; what is new is the source view and the label family f_r, which the model never saw. Section 1 now opens with a worked example (Figure 1a), and Section 3.3 adds a reader table. (2) We do not compare against the original decoder: the decoder followed by a lookup of the observed label is an upper reference (99.8–99.9%, Table 19), now stated in Section 3.4. The comparison is among training objectives for the same new reader on w. (3) Direct-ID is an affine identity head on w for 5k of 200k updates, then discarded (Eq. 2). Atomic-CE is indeed supervision through the decoder; Section 3.2 now says so and states how the two differ. (4) Your "about 2 points" for four views is right for the cohort of Figure 2 (+1.9 / +2.2 pp). The control cohort gives +2.9 / +2.6 and the full panel +4.8 / +1.8, where only the graph interval excludes zero; Section 4.1 reports all three.

W1, expected / by construction. We agree that part of the result holds by construction, and Section 1 now says which part ("What is expected, and what is not"). Three results do not follow from it. (a) Atomic-CE uses the same identity labels, pairs, window and gradient rule, yet trails Direct-ID by 5.3 / 7.6 pp (Tables 4, 19), so training on identity is not sufficient: how the labels reach w matters. Section 4.2 decomposes the gap (Figure 3d): under the bounded rule a fresh head at the same depth recovers 4.18 of the 6.1 points (a step that changes head and exposure together), moving it to w at matched exposure 1.55, and remaining exposure 0.36. (b) Identity access does not give attribute readability: on the same Direct-ID states an affine identity probe reaches 99.4%, an affine probe for random attributes 44.1% (Table 3). Random partitions of identity centroids are almost always affinely separable at these sizes, but their margins are small relative to within-identity scatter (Section 4.1, Appendix C). (c) The result you list as a strength is not implied by identity training: at matched grouping Direct-ID keeps +10.9 / +5.4 pp more source-factor readability than Contraction (Table 12, Figure 4).

New E5 separates the by-construction part. A temporary head on one coarse random 16-way partition g makes g readable (+19.7 pp for new readers; logistic 85.0% vs 32.1%) but adds only +2.44 [0.85, 4.04] / +2.25 [1.01, 3.49] pp for unrelated partitions, 40% / 36% of the Direct-ID gain (Section 4.1, Figure 2c). Supervising a label makes that label readable; the gains we report are for labels no objective supervised. New E2 uses no fitted reader. With the model's decoder frozen, replacing a view's state by its identity prototype, computed without that view, yields the correct answer in 90.0% of graph queries after Direct-ID, 68.0% after Atomic-CE and 51.4% natively (+22.0 [13.9, 30.1] pp; Section 4.4, Figure 5). Its ordering largely follows grouping (Table 5), so E2 shows use by the model's own computation rather than ruling out a grouping account.

W2, effect size beyond the hardest setting. Gains shrink with coverage, as the four-view numbers above show. We focus on one view because a reader that must carry a label from a single view depends most on how the state groups views. Under the bounded rule the one-view gain excludes zero in all fourteen settings of Table 1 (graph 5.2–7.6, addition 6.3–9.6 pp). On the original task we find no detectable difference (+0.90 [−0.62, 2.41] / +0.02 [−1.23, 1.26] pp, Table 19), so the transfer gain comes without a measurable cost there; Section 1 lists an original-task gain among the claims we do not make.

W3, Contraction beats Direct-ID. Correct: +2.72 [0.04, 5.40] pp on graphs and +1.22 [−1.30, 3.75] on addition, now reported in Table 19. Section 1 states that we do not claim Direct-ID is the best objective. Our claim is the trade-off: identity access at a smaller source-factor cost. E2 adds that Contraction states are the most interchangeable (+5.0 pp over Direct-ID), consistent with their tighter grouping and larger factor loss.

W4, presentation. We agree and have revised accordingly: a worked-example Figure 1 for the premise, a reader table (Section 3.3), a reading guide opening Section 4, intervals moved from prose to tables, a new abstract and a shorter Discussion.

We hope these changes address the soundness concern and would be grateful if the assessment could be revisited in that light.

Thank you again for your review. We would like to check whether our response and the revised version has addressed your concerns or not.

Currently, Section 1 now separates what follows from the design from what does not, and two new experiments test this directly: a head trained on one coarse partition makes that partition readable but transfers little to unrelated labels (E5), and the frozen decoder accepts identity prototypes without any refitting (E2). On presentation, Figure 1 now explains the premise with a worked example, and Section 4 opens with a reading guide.

We would be glad to clarify anything that remains unclear.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.