acceptodds
Under review as a conference paper at ICLR 2027

SUPERVISION BEFORE STRUCTURE: WHAT A WORLD MODEL NEEDS TO KNOW WHAT IT CANNOT ANSWER

Abstract

A world model asked a counterfactual question should first decide whether its evidence determines the answer at all. We study what such a model needs for that decision on a benchmark where the answer is known exactly: ten environments over seven systems (six physical, one structural causal model) whose hidden mechanism is drawn from a finite bank, so compatible laws, answer dispersion and the value of every probe are enumerated rather than estimated. We test the causal world-model design the field is converging on — a mechanism posterior with per-mechanism counterfactual heads — against a supervision-matched control with none of that machinery. Three findings follow. With an answerability label on every training query, learned structure is redundant: eight of nine ablations are null, the posterior's own dispersion scores answerability far below a learned head, and the one margin the architecture showed was an artifact of the baseline's training loss. With scarce labels — as few as forty — a factorised posterior improves label efficiency by up to AUROC, needs no mechanism labels, and transfers to a non-physics causal model; auxiliary supervision, generic factorised capacity, independent heads and width do not account for it, nor, on most environments, do the posterior's own auxiliary losses. Acting on the query is at oracle level for both models; stopping and answering after self-gathered evidence are weak for both. We release the benchmark, every control, and the failed hypotheses.

Primary area:
causal reasoning
Keywords:
world models, identifiability, counterfactual reasoning, causal representation learning, mechanism posterior, selective prediction, label efficiency, active evidence acquisition, benchmark

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_f6CMOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The study Supervision Before Structure demonstrates that explicit causal architecture in world models is redundant for determining question answerability under full supervision, but provides an effective inductive bias that improves label efficiency when answerability labels are scarce.

Critical strengths

This paper raises a seemingly interesting question.

Critical weaknesses
  1. This paper is very hard to read as there seems to be no editing at all. It thus lacks the most basic presentation requirements for scientific results and discovery. For example, lines 216- 258 lack clarity, as the entire paragraph is presented as one block. Moreover, the experimental setup and tables are impossible to understand.
  2. The primary limitation is that the claimed label-efficiency benefit of structured world models is demonstrated under an artificial supervision regime in which identifiability labels are selectively withheld while closely related counterfactual and evidence-acquisition targets remain fully available. When these supervision sources are jointly reduced, the reported advantage largely disappears. Furthermore, the fact that removing or randomizing mechanism supervision does not degrade performance raises questions about whether the improvement reflects meaningful causal structure rather than a generic architectural inductive bias. Overall I think these findings substantially weaken the paper's central claim and its practical significance.
Questions for authors

I encourage the authors to carefully revise their work before considering resubmission.

1 comment
AuthorsOpenReview ↗

Thank You reviewer

"When these supervision sources are jointly reduced, the advantage largely disappears." We agree, and the new experiments make the point stronger than the submitted table did. With the evidence-gain target confined to the labelled queries, the structured model has no forty-label lead on five of eight environments; with all three targets confined, the structured model is behind the control on five of eight (Table 4). You wrote that these findings weaken the paper's central claim. The central claim of the revised paper is those findings: the margin is not label efficiency, it requires an auxiliary target that contains the label, and it owes nothing to causal semantics (Section 4).

"Removing or randomising mechanism supervision does not degrade performance." Agreed, and the revision says it in those terms: what survives when both auxiliary targets are fully available is specific to the full structured model, with or without mechanism labels, so it owes nothing to causal semantics. A generic K-expert head given to the control does not reproduce it, and we have not isolated the ingredient further (Section 4, "What remains"; Appendix D).

Presentation. The paper is rewritten. Old lines 216–258 are now four short paragraphs with one claim each, and the numbers they carried are in Table 4 and Appendix Table 6. Every table caption now defines its columns, and the two models are drawn side by side in Figure 4.

Does the restructured Section 4 address your concern about the central claim?

Reviewer_94DZOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates whether explicit causal-mechanism representations are necessary for world models to predict the identifiability of counterfactual queries, and under what supervision regimes such representations are beneficial. The authors introduce ID-Phys++, a synthetic benchmark spanning ten environments across seven physical and causal systems. By enumerating a finite bank of candidate mechanisms, the benchmark provides exact, bank-relative labels for counterfactual identifiability, answer dispersion, and the expected value of evidence-acquisition probes.

Critical strengths
  • The experiments are generally carefully designed, with supervision-matched baselines, multiple seeds, several OOD evaluations, and extensive ablations. The authors also acknowledge important limitations.
  • The motivation and overall experimental questions are well articulated, and I appreciate the transparent presentation of negative results and revised conclusions.
Critical weaknesses
  • My main complaint is that the motivation of this research direction is unclear. The authors introduce ID-Phys++, a synthetic benchmark spanning ten environments; however, it is unclear how this benchmark connects with real world situations. I would recommend rejections in the current form.

  • The scope of the benchmark is pretty limited. All environments are synthetic and the identifiability ground truth is relative to a known finite hypothesis bank. Several fully supervised tasks are near saturation.

  • The identifiability labels are computed from parameterized mechanisms in the full compatible bank, whereas the learned categorical posterior operates over a smaller collection of family-level components. Can the authors quantify how often non-identifiability arises from disagreement among parameters within the same family, and whether that helps explain the gap between posterior dispersion and the learned identifiability head?

Questions for authors

See my comments above.

1 comment
AuthorsOpenReview ↗

Thank You Reviewer

Motivation and connection to real situations. The introduction now opens with the problem and a concrete case. A model that has seen one patient's response at one dose is asked what a doubled dose would do. If a linear and a saturating law both fit what it saw, the useful output is not a number but a request for a measurement at a second dose. Deciding which case holds is answerability prediction, and it gates both abstaining and evidence gathering. Explicit mechanism posteriors are the design the literature proposes for this decision; our question is whether they are what makes it work. The benchmark is synthetic because that is the only way to know the right answer for every query, and we argue in the introduction and Discussion why a finite known bank is the most favourable setting for the proposal under test.

Within-family versus across-family non-identifiability. This was computable exactly and we had not done it; thank you. For every non-identifiable test query we attributed the disagreement (Appendix Table 21):

envacross families onlywithin a family onlybothAUROC, exact dispersion over familiesover hypotheses
ramp0.180.140.680.9740.993
spring0.290.090.610.9770.991
pendulum0.110.370.520.8350.974
two-body0.060.380.570.9440.989
spring (cont.)0.210.170.620.9650.995
SCM dose0.120.180.700.9660.996

Within-family disagreement is the only source for 9–18 % of non-identifiable queries on four systems and 37–38 % on the pendulum and two-body systems. An exact family-level posterior would still score 0.97–0.98 AUROC on the ramp and spring systems, where the learned posterior's dispersion scores 0.66–0.69. Family granularity therefore explains at most 0.02–0.03 of that 0.30 gap; the rest is the quality of the learned posterior. On the pendulum granularity explains more (family ceiling 0.84).

Scope and saturation. See the headroom result in the general response (Table 2) and the harder-regime result (Table 3): at 0.88–0.94 AUROC the two models are still equivalent.

Reviewer_UzWKOpenReview ↗
Rating
2(−1)
Confidence
3
Informativeness
2
Presentation
1(−1)
Soundness
2(−1)
Read review
Paper summary

This paper studies whether a world model can tell when the available evidence is sufficient to answer a counterfactual question.

The authors introduce ID-Phys++, a benchmark covering ten environments across seven physical and causal systems, where a finite bank of candidate mechanisms allows answerability and probe value to be computed exactly. They compare a model with reference architecture predictions against a simpler model trained with the same answerability supervision. With abundant labels, the explicit structure generally offers little benefit, and an apparent advantage on the ramp tasks disappears after scaling the prediction loss by the answer tolerance. When answerability labels are scarce, however, the structured model learns more efficiently, even without mechanism labels, although the precise source of this benefit remains unresolved. Both models select useful probes nearly as well as the oracle, but struggle to decide when to stop and to answer correctly after collecting new evidence. The findings highlight how strongly the benefits of structure depend on supervision and training choices.

Critical strengths
  • The authors can check exactly whether the evidence determines an answer, rather than treating every wrong prediction as the same kind of failure.

  • The comparison is fairer than a standard baseline comparison. They also teach the simpler model which questions are answerable, helping separate the value of the architecture from the value of extra training labels.

  • The paper finds a concrete setting where structure helps. When answerability labels are scarce, the structured model usually performs better, even without labels identifying the underlying mechanism.

  • Both models choose good experiments, but still struggle to use the new evidence and answer correctly. This shows why good experiment selection alone is not enough.

Critical weaknesses
  • The setting is still quite restricted. All experiments use synthetic environments with a predefined mechanism bank. The ground truth is exact within that bank, but this does not establish whether the models can recognize insufficient evidence when the possible mechanisms are unknown.

  • The main positive result is not fully explained. Structure helps when answerability labels are scarce, but removing mechanism labels or scrambling their meaning does not remove the benefit. The controls rule out several explanations, yet leave unclear what specifically makes the proposed representation work better.

  • Most experiments reduce only answerability labels while retaining counterfactual-answer and probe-value targets. The results therefore show reduced dependence on one type of label, rather than learning from very little supervision overall. The scarce-label setting still provides substantial supervision.

  • -Some environments offer limited room for a meaningful comparison. The collision tasks are nearly saturated, and the causal-model task contains strong query-type shortcuts. These cases provide weaker evidence about whether explicit structure matters for difficult reasoning.

  • How sensitive are the findings to the particular structured architecture used here? The paper tests several ablations and capacity settings, but alternative designs, such as continuous latent-parameter models or different ways of combining mechanism predictions, could change the comparison. Could the authors clarify which conclusions are specific to QID-AWM and provide evidence or discussion on whether alternative structured designs might improve or worsen performance ?

Questions for authors

Could the authors provide an architecture diagram or pseudocode showing how the observation history and query are processed, how the mechanism posterior is computed, and how it interacts with the counterfactual and answerability heads? It would also help to state explicitly how this computation differs from B1+ID and the KHead controls.

2 comments

Thank you reviewer

Architecture diagram and pseudocode. Added: Figure 4 draws QID-AWM and the control side by side with the components that differ highlighted, and Appendix Figure 7 gives pseudocode for both and for the KHead controls.

"The main positive result is not fully explained." It is now explained further, and it is smaller than we claimed. The margin rests on the evidence-gain target, which contains the label (general response, experiments 1–3). The revised paper presents it as a qualified effect with its conditions, not as a contribution. One candidate we can rule out by construction: the encoder, including its early conditioning on the query, is identical in the structured model and in every control (Figure 4; Appendix Figure 7), so the margin is not an encoder difference. What we could not do is isolate which part of the posterior-plus-heads combination carries it, and the paper says so.

"Reduced dependence on one type of label, not learning from little supervision." Agreed. Table 4 now shows all three regimes side by side, and with all three targets scarce the structured model is behind the control at 0.1 %.

Which conclusions are specific to QID-AWM? All of them are about one family: a categorical posterior over a finite bank read by per-mechanism heads. Within that family we varied the number of components (2–24), width, depth, training length, the posterior losses, mechanism supervision and head sharing (Appendix B and D). Continuous latent-parameter designs are not tested and we claim nothing about them; the introduction and the Limitations paragraph say so.

Restricted setting, saturated collision tasks, SCM shortcut. Agreed on all three. The collision environments are saturated and carry no weight in our conclusions; on the other eight the models are equivalent on seven at full supervision. On the SCM you are right: it has no query type with a balanced base rate, and the query-type prior alone reaches 0.949 AUROC there, so at full supervision it is weak evidence; its informative results are the scarce-label ones. On the other environments the shortcut does not explain the result: on balanced query types the prior falls to 0.53–0.68 while both models keep 0.91–1.00 (Appendix Table 13). The off-bank experiment is our only step outside the bank (Appendix Table 19).

Thank you for updating the manuscript. While it explores an interesting question, I cannot recommend it in its current form, as several critical issues remain unresolved and the writing is not clear enough.

Reviewer_7siTOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2(+1)
Soundness
1
Read review
Paper summary

The paper studies whether a world model needs explicit causal structure in order to decide whether a counterfactual query can be answered from evidence. The authors introduce a benchmark, ID-Phys++ with 10 simulated environments. They compare a structured model against a baseline that receives the same information without a posterior. Since the hidden mechanism in each environment is drawn from a small fixed set, whether a query is answerable can be computed exactly. With full supervision, both models perform the same, and an initial advantage of the structured model turns out to be due to the training loss. With very few labels, the structured model shows a small improvement.

Critical strengths
  1. Exact labels + fair baseline: Because the hidden mechanism is drawn from a small fixed set, the authors can compute exactly whether each query is answerable. This makes for a strong testbed, since it lets them control exactly what information each model receives during training and evaluate against the true answer. They also compare against a baseline that receives the same labels without the causal machinery, which is the right baseline for studying this question.
Critical weaknesses
  1. Presentation: The manuscript is extremely difficult to follow. The text defers its main results to an appendix table (Table 17), along with an over-reliance on the appendix in the primary text. Furthermore, sections 3–4 are essentially numbers, but written in prose. Instead of presenting a coherent argument, the results are told chronologically, making it very hard to determine what the core argument presented in the paper actually is, and what the main claims are. For example, a claim is initially made, only to be retracted later in the paper. There are also several internal inconsistencies (e.g. Table 6 vs Tables 1/20/21; "nine of ten environments" in Sec. 5 vs the eight in Table 31). The AI-use statement notes that AI was used to polish the writing; whatever the cause, even if the conclusions were strong, the manuscript in its current form does not meet the presentation standard expected at ICLR.

  2. The positive result is unsupported: The main positive claim of the paper is that structure helps when identifiability labels are scarce, but I am not convinced this is supported. In the scarce-label experiments, only the identifiability label is removed. Both models still receive the evidence-gain target for every query, and this target is derived from the same enumeration, so it likely leaks most of the same information. When the authors remove it as well (Table 16), the advantage mostly disappears.

  3. Benchmark and environments: The environments presented in this paper are largely toy examples and the results thus do not directly translate to more complex environments that are actually relevant to world models. This is also supported by their presented result numbers. e.g., under the full supervision regime, both models score >0.96 AUROC on every environment (Table 17). The central claim that learned structure is redundant is therefore measured on a saturated benchmark, where no architecture can actually show a reasonable advantage. Harder environments (more complex mechanisms/interactions, high dimensional observations, real video rather than toy images) are required before the results can be read as evidence about world model design.

Questions for authors
  1. What is the original ID-Phys benchmark? And what are the improvements made to ID-Phys++? It is also not made clear what each state observation actually contains (e.g., the rendered environments are described as 6×32×32, but it is not explained what the 6 refers to, whether frames or channels).

  2. Both models presented in this paper are trained to score how much each extra experiment would help answer a query (i.e., the evidence-gain target). If a query is already answerable, no experiment can help, so this score seems to reveal the answerability label indirectly. How well can answerability be predicted from this score alone? And do the scarce-label results in Table 1 still hold when this target is also removed?

2 comments
AuthorsOpenReview ↗

Response to Reviewer 7siT

Thank you reviewer for the specific questions; both changed the paper.

Q: How well can answerability be predicted from the evidence-gain score alone? At the data level, almost perfectly: the negated maximum oracle gain predicts the label at 0.97–0.996 AUROC on the eight non-saturated environments. From the trained evidence heads: 0.87–0.95 for the structured model, 0.67–0.86 for the control, at every label fraction (Appendix Table 22). You were right that this target leaks the label.

Q: Do the scarce-label results hold when this target is also removed? No, not at the scarcest setting. With the evidence-gain target confined to the labelled queries (counterfactual target kept), QID-AWM minus control at 0.1 % labels is:

rampspringvis. rampvis. springpendulumtwo-bodyspring (cont.)SCM
+0.011−0.039−0.004−0.021−0.007+0.006−0.003+0.028

against +0.010 / −0.004 / +0.018 / −0.005 / +0.036 / +0.013 / +0.009 / +0.024 with the target on every query. At 1 % the margin shrinks on five of eight environments and stays positive on all eight (+0.001…+0.013). With all three targets confined, the structured model is behind the control on five of eight environments at 0.1 % (Table 4). The revised paper no longer claims that structure helps when labels are scarce. It reports this margin as the product of a label-revealing auxiliary target, which the structured model exploits better than the control, and not as label efficiency.

Presentation. We agree with the assessment and rewrote the paper; the general response lists what changed. The argument is no longer chronological, the retracted claim is out of the main text, and the main comparison is a main-text table.

Inconsistencies. Both were real.

  • Old Table 6 quoted a QID-AWM reference at 1 % from validation-selected runs (0.961 on the ramp) while old Tables 1/20/21 used final-checkpoint runs (0.955). It now uses the same runs (Appendix Table 10).
  • "Nine of ten environments" in old Section 5 referred to a table with eight. The sentence now reads "seven of the eight environments evaluated" (Appendix Table 38).

Saturation and toy environments. Two answers. First, the null is not a ceiling effect: with an exact posterior in place of the learned one the structured model scores up to +0.016 AUROC above the control, and +0.014 above its own learned version, on the ramp environments (Table 2), so there was room for the machinery to help and it did not. Nor does a gap open when the task is made harder: under observation noise that lowers both models to 0.88–0.94 AUROC they remain equivalent on all four environments tested (Table 3). Second, we scope the claim. A finite, known bank is the setting most favourable to a categorical mechanism posterior, because its support matches the truth exactly. If the machinery adds nothing there, a misspecified setting gives it less reason to help. We do not claim the result transfers to video world models, and the Discussion says so.

ID-Phys++ and 6×32×32. There is no published "ID-Phys"; the suffix marks the second generation of our own pilot, now stated at first use. Rendered observations are six greyscale 32×32 frames per episode.

Does the revised claim resolve your soundness concern?

Thank you for the detailed response and the updated manuscript. The revised claim does resolve my concern about the label-efficiency result, as the paper no longer actually makes this claim. I appreciate that the authors ran the evidence-gain experiment and report its outcome openly. The presentation is also much better, and the inconsistencies and questions on ID-Phys and the 6×32×32 inputs have been addressed.

However, the revision essentially changes the main claim of the paper. The one positive result is withdrawn, most of the text is rewritten, and substantial new experiments have been added. Changes of this scale are difficult to evaluate fully within the discussion period. Furthermore, what remains is largely a negative result on a near-saturated benchmark. This is also supported by the new results, e.g., the headroom in Table 2 is at most ~0.016 AUROC with all models still above 0.98, and the noise experiment only makes the task harder by removing information. I am not convinced this is sufficient for a conference such as ICLR in its current form, and thus I will keep my current scores while improving the score for the presentation.

Comments on the paper

3 comments
AuthorsOpenReview ↗

We thank all four reviewers.

We thank all four reviewers. The reviews agree on two things: the paper was hard to read, and its one positive claim (structure helps when labels are scarce) was not supported. Both are right. We have rewritten the paper and changed its claim.

The claim now. On a benchmark with exact answerability labels, none of the structured model's apparent advantages reflects the causal machinery doing what it was designed to do, and none depends on causal semantics. The revised paper shows this three times, each with a controlled test (new Figure 2):

apparent advantage of the structured modelsizecause, and the test that shows it
lead at full supervision on the two ramp environments+0.009 AUROCscale of the training loss: under a tolerance-scaled loss the margin is ≤ 0.001 (Table 1)
label efficiency at 0.1 %–1 % of the labelsup to +0.036an auxiliary target that contains the label, which the structured encoder exploits better (Appendix Table 7): confining the evidence-gain target to the labelled queries leaves the structured model with no lead on five of eight environments (Table 4)
larger scarce-label margin under best-checkpoint selection+0.01 to +0.03 on topselection on labelled validation data (Appendix Table 11)

The second row is the objection raised by Reviewers 7siT and f6CM. We ran the experiment they asked for, it confirmed their reading, and it is now a headline result of the paper, not a caveat.

New experiments since submission (all five seeds, final checkpoints, both models):

  1. Evidence-gain target withheld with the label (7siT): eight environments, 0.1 % and 1 %. The structured model then has no forty-label lead on five of eight environments and is significantly behind on both spring systems (−0.039, −0.021; they had no margin to begin with, −0.004 and −0.005). At 1 % it shrinks on five of eight and stays positive on all eight (+0.001…+0.013).
  2. All three targets withheld together, extended from three environments to eight and to 0.1 %: the structured model is behind the control on five of eight environments at 0.1 % (by 0.03–0.09 AUROC) and within −0.003…+0.007 at 1 %.
  3. How much the evidence target reveals (7siT): the negated maximum oracle gain predicts the label at 0.97–0.996 AUROC. A linear readout of the trained evidence head's utilities predicts it at 0.87–0.95 for the structured model and 0.67–0.86 for the control (Appendix Table 22).
  4. Where non-identifiability comes from (94DZ): exact attribution of every non-identifiable test query to across-family and within-family disagreement (Appendix Table 21).
  5. Headroom (7siT, UzWK, 94DZ on saturation): a structured model given the exact compatible family set scores +0.014 and +0.016 AUROC above the control on the two ramp environments and +0.003 to +0.009 on the three spring environments (Table 2; on the rendered spring the learned model already has +0.005 of the +0.009, and the two rendered rows use three oracle seeds). The full-supervision null is therefore not a ceiling effect on five of eight environments.

(Experiments 6–7 and the summary of the rewrite continue in part 2.)

AuthorsOpenReview ↗

**New experiments since submission, continued** (five seeds, final checkpoints, both models):

  1. Harder regime (7siT, UzWK, 94DZ on saturation): both models retrained on four state environments under observation noise σ = 2 and 4, two and four times the heaviest level in the submission (Table 3). AUROC falls to 0.93–0.96 and then 0.88–0.94. The two models are equivalent in all eight cells (largest gap 0.0008, every 95 % interval within ±0.002). The +0.009 spring margin we had reported at σ ≤ 1 is gone at σ = 2 and 4. We state the limit of this test in the paper: noise makes the task harder by removing information, and at σ = 4 both models are 0.04–0.07 above the query-type prior, so it shows that structure does not degrade more gracefully, not that it would fail on a task made harder in another way.
  2. Forty-label probe. To test whether the scarce-label margin is only a head that is easier to fit, we fitted a logistic readout on each frozen belief vector using exactly the labelled training queries (Appendix Table 7). The readout on the control's belief stays below the structured model's head on seven of eight environments at 0.1 % (by 0.006–0.076 AUROC) and on all eight at 1 %, and the same readout scores higher on the structured belief in every cell (+0.003 to +0.046). The margin is in the representation: the structured encoder, trained on the same leaking target, lays answerability out more linearly. This is an advantage of the structured model, and we report it as one: it exists only while the leaking target is on every query, and it does not depend on mechanism labels or their meaning.

A caution on the 0.1 % column. Training is not bitwise reproducible: re-training the same configuration and seeds at 0.1 % labels moves a five-seed mean by up to 0.02 AUROC (Limitations). Tables 7 and 22 use a second training run, so their head columns differ from Table 4 by up to that amount. At 0.1 % we therefore rest conclusions only on effects larger than 0.02 or consistent in sign across cells: the reversals of 0.03–0.09 when all three targets are confined, and the probe comparison, which has the same sign in all 16 cells. The 1 % results are not affected at this scale.

AuthorsOpenReview ↗

**The rewrite.**

The paper now states the problem, the question, the method and the answer in that order. Concretely:

  • The introduction has labelled paragraphs (the problem, why it matters, the proposed solution and our question, how we test it, what we find) and a one-figure summary.
  • Each results section opens with its question, method and answer; each paragraph makes one claim with one or two numbers.
  • Final claims come first. The tolerance-scaled loss is the default from Section 2; the raw-unit loss is an ablation. The chronology, including the claim we retracted, moved to an appendix audit note.
  • The headline comparison is Table 1 in the main text (it was Appendix Table 17). Main-text pointers to the appendix went from over 30 to 10.
  • An architecture diagram of both models is Figure 4; pseudocode is Appendix Figure 7.
  • Figures are redrawn: a one-figure summary of the three margins (Figure 2) and a diagram of how the exact labels are computed (Figure 3) are new; the noise plot (Figure 5) now extends to the harder regime.
  • Action selection, stopping and calibration are one paragraph; the detail is in the appendix.
  • Both inconsistencies Reviewer 7siT found were real and are fixed (see that reply).

Old to new numbering. Reviews cite the submitted numbering; the revised PDF differs:

submittedrevisedcontent
Table 1Table 6scarce-label matrix
Table 2Table 1loss control / headline comparison
Table 3Table 5noise sweep
Table 4Table 8sequential agent
Table 5Table 9posterior quality
Table 6Table 10K sweep
Table 7Table 11checkpoint selection
Table 8Table 12posterior-shaping losses
Table 9Table 13query-type shortcut
Table 10Table 14oracle posterior
Table 11Table 15KHead gate
Table 12Table 16width
Table 13Table 17no mechanism labels
Table 14Table 18absolute AUROC under noise
Table 15Table 19off-bank law (compact)
Table 16Table 20joint reduction
Table 17Table 24full IID results
Table 18Table 25OOD splits
Table 19Table 26visual OOD
Table 20Table 27full label-fraction sweep
Table 21Table 28label-fraction variants
Table 22Table 29four loss recipes
Table 23Table 30learning-rate check
Table 24Table 31off-bank law
Table 25Table 32ablations
Table 26Table 33stopping
Table 27Table 34capacity
Table 28Table 35sequential (all policies)
Table 29Table 36dispersion threshold
Table 30Table 37calibration
Table 31Table 38conformal coverage
(new)Table 2headroom at full supervision
(new)Table 4scarce-label margin under three withholding regimes
(new)Table 3harder regime (heavy observation noise)
(new)Table 21within- versus across-family attribution
(new)Table 22evidence-head readout
(new)Table 23evidence-only reduction
(new)Table 7forty-label probe

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.