acceptodds
Under review as a conference paper at ICLR 2027

Learning When to Trust: Risk-Adaptive Supervision for Flow Map Models

Abstract

Flow map models (FMMs) achieve efficient one-step generation by directly learning finite-time transport, yet existing training strategies largely treat supervision reliability as globally uniform. We challenge this assumption and show that the relative reliability of different supervision signals can vary substantially across transport states. To formalize this observation, we develop a flow-map-aware risk framework that characterizes state-dependent supervision reliability and derives the optimal trade-off between complementary supervision sources. Building on this framework, we propose Risk-Adaptive Flow Map Learning (RAFL), which estimates state-dependent reliability from target-space disagreement and uses it to construct adaptive finite-time supervision. RAFL provides a lightweight post-training refinement for pretrained FMMs without modifying their inference procedure or introducing additional inference-time cost. Extensive experiments across diverse FMMs reveal substantial supervision heterogeneity and demonstrate consistent improvements over existing supervision strategies.

Primary area:
generative models
Keywords:
Flow Map Models, Risk-Adaptive Supervision, Representation Guidance, One-Step Generation, Reliability Estimation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_jGknOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper argues that, when training flow map models (FMMs), the coefficient that balances stochastic conditional supervision and pretrained guidance should not be set globally, but should instead adapt to the transport state. To this end, the authors propose Risk-Adaptive Flow Map Learning (RAFL), which approximates the closed-form optimal balance coefficient through kNN-based risk estimation combined with an amortized predictor. Experiments on CIFAR-10 and class-conditional ImageNet-256 show that RAFL outperforms the baselines and support the effectiveness of the proposed risk-based adaptation.

Critical strengths

The motivation is clear and insightful, and it is partially supported by empirical observations. The method appears to have the potential for broad applicability, and its benefits are well demonstrated in the experiments. Overall, these points lead me to lean toward acceptance.

Critical weaknesses

I am not an expert on flow-map models, but the claimed derivation of "the optimal balance between stochastic conditional supervision and potentially biased pretrained guidance" (i.e., Prop. 3.1) appears to be a standard bias–variance trade-off result. This leaves me uncertain about how much of the theoretical analysis goes beyond results that are already classical in this field or in statistics. It may be more appropriate to frame the first contribution around demonstrating why λ needs to be modeled as a risk-adaptive coefficient.

In addition, several details of the experimental settings and results are not sufficiently explained (see Questions).

Questions for authors
  1. Which teacher model provides the v_φ signal during training?
  2. Since the estimation of V and B is one of the core contributions, is there any experiment that directly evaluates the accuracy of these estimates? The current experiments use a kNN-based metric to show that the estimates are self-consistent, but do not characterize their bias or variance.
  3. In Fig. 2(a), all curves appear to start at an FID of approximately 3.7 at iteration 0, whereas the shared pretrained checkpoint has an FID of 3.47 (Tables 1 and 10), and Fig. 3(b) starts at approximately 3.47 for the same model. Could the authors clarify this discrepancy?
  4. Could the authors report the additional training cost that RAFL introduces relative to standard training?
4 comments

We thank the reviewer for the constructive feedback and the suggestion to emphasize why and when supervision should be risk-adaptive. We address each point below.

W. Theoretical contribution and the motivation for risk-adaptive weighting.

We agree that the scalar bias-variance minimization in Proposition 3.1 is standard. We have revised the contribution statement to emphasize why and when supervision should be risk-adaptive, using this proposition as a building block.

First, we formulate reliability in the finite-time target geometry, Gξ=Aξ⊤AξG_\xi=A_\xi^\top A_\xi. This accounts for how instantaneous uncertainty and teacher bias propagate into the supervision target actually optimized by the model.

Second, Corollary 3.3 quantifies the benefit available to fully state-dependent mixing over the best transport-condition-only coefficient:

\mathbb E\left[W_\xi(\lambda_\xi^\star-\lambda_\kappa^\star)^2\mid\kappa\right],$$ where $\kappa=(r,t)$ and $W_\xi=V_\xi+B_\xi$. Positive $H_{\mathrm{sup}}$ identifies when state adaptation can strictly reduce supervision-target risk. Third, Theorem 3.2 and Corollary 3.3 characterize the optimal coefficient under restricted information and decompose this opportunity into representation-resolved and unresolved components. Section 4 further quantifies the policy discrepancy introduced by unresolved teacher-bias variation in disagreement calibration. These results motivate state-conditioned risk prediction and calibration. The optimality statements concern supervision-target risk within the specified mixture family at a fixed target-network snapshot. ### Q1. Which teacher model provides the $v_φ$ signal during training? For each experiment, we use a frozen copy of the corresponding pretrained baseline as the teacher, and initialize the student from the same checkpoint. Thus, the teacher matches the backbone being refined: MF uses its pretrained MF baseline, iMF uses its pretrained iMF baseline, and similarly for the other evaluated FMMs. For example, the controlled MF-XL/2 experiments use the 300k-step MF-XL/2 checkpoint with one-step FID 3.47. For the MF formulation, instantaneous guidance is obtained through diagonal evaluation: $$v_\phi(z_t,t):=u_\phi(z_t,t,t).$$ The teacher parameters $\phi$ remain fixed throughout post-training, while the student parameters $\theta$ are updated by RAFL.

Q2. Accuracy of the estimated risk components.

To directly assess the accuracy, bias, and variability of the estimated VξV_\xi and BξB_\xi, we added an independent oracle-based evaluation in Appendix A.4 (Table 7).

On ImageNet, the teacher bias bt=vϕ(zt,t)−v⋆(zt,t)b_t=v_\phi(z_t,t)-v^\star(z_t,t) cannot be directly evaluated because the true marginal velocity v⋆v^\star is unavailable. Squared target-space disagreement between teacher and conditional supervision has conditional expectation Vξ+BξV_\xi+B_\xi, so it does not directly measure BξB_\xi. We therefore use an eight-component 2D Gaussian mixture whose analytic posterior allows direct computation of ground-truth VξV_\xi and BξB_\xi for a frozen MF snapshot.

An invertible representation gives population-level Cb=0C_b=0, isolating estimation and amortization errors. We evaluate four transport conditions using 20 independent evaluation banks per condition (16,384 samples each) and three predictor-training runs. Training data and calibration banks are separate from test queries and evaluation banks; oracle targets are used only for evaluation.

At K=32K=32:

MethodVV biasVV SDVV RMSE ↓BB biasBB SDBB RMSE ↓
Uncorrected kNN0.41391.66302.74500.37870.51661.1111
Corrected kNN0.03301.71662.60920.14420.43400.7774
Amortized QψQ_\psi0.01400.46701.15630.08810.00640.1575

Bias and SD are computed per query across independent banks for kNN or training runs for QψQ_\psi, then averaged over queries. All entries are arithmetic means of the four condition-level statistics; RMSEs are not pooled across conditions.

The finite-sample correction reduces BB RMSE by 30.0%, while the amortized predictor further reduces VV RMSE by 55.7% relative to corrected kNN. The predictor also achieves coefficient MAE of 0.02088 and normalized excess risk of 0.00326. These oracle-based measurements directly characterize the bias, variability, and accuracy of the estimated risk components. Ground-truth accuracy claims are restricted to this controlled setting; the ImageNet diagnostic evaluates policy alignment.

Q3. Consistency of the initial FID across figures.

We thank the reviewer for identifying this inconsistency. All strategies start from the same 300k-step MF-XL/2 checkpoint with 1-NFE FID 3.47, and iteration 0 denotes this checkpoint before any post-training updates. The approximately 3.7 starting value in the original Fig. 2(a) was incorrect. We have corrected the panel using the corresponding evaluation records; the revised Fig. 3(a) now shows the common initial FID of 3.47, consistent with Table 1.

Q4. Additional training cost.

For iMF-XL/2 on ImageNet-256, RAFL increases FLOPs per training step by approximately 6.64% relative to standard continued training. Both methods start from the same 1M-step checkpoint and use batch size 1024 for 40k additional steps.

MethodBatch sizeAdditional stepsCompute (PFLOPs/step)Total post-training compute (EFLOPs)
iMF-XL/2 + continued training102440k2.020180.804
iMF-XL/2 + RAFL102440k2.154386.172

RAFL adds 0.1342 PFLOPs per step, totaling 5.368 additional EFLOPs over 40k steps. Under the same per-step accounting, the entire RAFL post-training stage requires approximately 4.27% of the original pretraining compute.

Thank you for the detailed response. Though I'm not an expert in this field, I learned something from this paper. I will maintain the positive rating.

It has been our honor to have you carefully review our work, and we sincerely appreciate the time and effort you devoted to evaluating our manuscript. We are truly grateful for your many valuable comments and constructive suggestions, which have helped us improve our work substantially.

Reviewer_XKU1OpenReview ↗
Rating
3
Confidence
2(+1)
Informativeness
3(+1)
Presentation
3(+1)
Soundness
2
Read review
Paper summary

This paper addresses the problem of state-dependent variability in supervision reliability during the training of flow mapping models and proposes a risk-adaptive learning method, RAFL. Paper analyzes the stochasticity of conditional supervision and the bias of pretrained teachers, and derive optimal mixing weights for the two supervision sources. The method leverages supervision disagreement in the finite-time target space, combined with representation-neighborhood calibration and a lightweight risk predictor, to assign supervision weights to individual states.

Critical strengths
  1. A discrepancy indeed exists between the conditional velocity and the pretrained teacher velocity, and both require adjustment as noted in the paper.
  2. It derives a theoretically grounded mixing coefficient, balancing the stochastic uncertainty of conditional supervision against the bias of pretrained guidance.
  3. The auxiliary modules participate only during training and do not enter the actual generation or inference stages, thereby introducing no additional computational cost.
Critical weaknesses
  1. This relies on a key nearest-neighbor assumption. The method assumes that teacher bias is relatively consistent within representation neighborhoods, yet it neither proves nor directly measures that CbC_b is sufficiently small; feature similarity does not necessarily imply bias similarity.
  2. A gap exists between theory and empirical performance. The optimality result pertains to supervision risk under a fixed target network and does not appear to guarantee optimal final generation quality.
  3. The effect of RAFL appears insufficiently pronounced. The FID improvement of RAFL over direct gating is only 3.07→3.01.
  4. Implementation details are inadequate. The introduction of new modules brings new design choices and hyperparameters; for example, the definition of “compatible temporal conditions” and the update procedures for calibration and the buffer lack sufficiently specific implementation descriptions.
Questions for authors
  1. Could you track the predicted mixing coefficient λ and estimated risks for a few fixed held-out transport states across training checkpoints, with a concrete example illustrating how and why reliance on the teacher versus conditional supervision changes during training?
  2. Could you provide additional information or code to facilitate reproduction?
4 comments

We thank the reviewer for the positive assessment and constructive feedback. We address each concern below.

W1. Neighborhood consistency and unresolved teacher-bias variation.

Section 4.1 explicitly characterizes how teacher-bias variation within representation neighborhoods affects the calibrated supervision policy. We quantify this effect through the unresolved transformed teacher-bias variation:

=\operatorname{Tr}\!\left( \operatorname{Cov}[A_\xi b_t\mid h_t=h,\kappa] \right).$$ At a fixed target-network snapshot, Eq. (8) and Corollary A.1 give $$\lambda_{\mathrm{cal}}^\star-\lambda_h^\star =\frac{C_b(h,\kappa)}{\mathbb E[W_\xi\mid h_t=h,\kappa]},$$ and $$R_{h,\kappa}(\lambda_{\mathrm{cal}}^\star) -R_{h,\kappa}(\lambda_h^\star) =\frac{C_b(h,\kappa)^2}{\mathbb E[W_\xi\mid h_t=h,\kappa]},$$ where $W_\xi=V_\xi+B_\xi$, $R_{h,\kappa}(\lambda)=\mathbb E[R_\xi(\lambda)\mid h_t=h,\kappa]$, and the denominator is positive. These identities do not require $C_b=0$: they characterize the consequence of departures from teacher-bias consistency. In particular, positive unresolved variation increases teacher reliance relative to the optimal representation-conditioned policy. The coefficient gap depends on this variation relative to the total supervision risk. We have not established that this normalized variation is small on ImageNet, and explicitly acknowledge this limitation in Appendix A.5. Finite-neighborhood estimation, cross-snapshot drift in cached disagreements, and predictor approximation are separate sources of practical error. Alignment with a held-out calibrated reference is therefore used as a policy diagnostic, rather than direct evidence of teacher-bias consistency. ### W2. Scope of the optimality result. Our result characterizes the ideal mixing coefficient that minimizes conditional supervision-target risk within the specified mixture family at a fixed student snapshot. It provides a snapshot-conditioned principle for constructing supervision targets, without claiming global training convergence or optimal final FID. During training, $\bar\theta=\operatorname{sg}(\theta)$ is refreshed at each iteration, while the pretrained teacher remains frozen. RAFL uses neighborhood calibration and amortized risk prediction to estimate the mixing coefficient as the student evolves. We assess final generation quality separately through matched controls. Starting from the same MF-XL/2 checkpoint with FID 3.47, continued training reaches 3.39 and RAFL reaches 3.01 after the same additional 20k steps (Table 10). This provides empirical evidence of improvement beyond additional optimization alone. ### W3. Improvement over direct adaptive gating. RAFL derives its mixing coefficient from the relative magnitudes of two calibrated risk components in finite-time target space. The predictor is trained through risk regression with generative gradients blocked, providing an explicit reliability-estimation objective and interpretable component estimates. Our fixed-snapshot analysis further characterizes the risk-optimal weighting rule and the policy discrepancy introduced by disagreement calibration. To evaluate this design across backbones, we extend the MF-XL/2 comparison in Table 4 to three additional backbones under matched initialization, post-training budgets, and evaluation settings. All results use 1-NFE sampling. | Backbone | Pretrained FID ↓ | Direct gate FID ↓ | RAFL FID ↓ | |---|---:|---:|---:| | MF-B/2 | 6.04 | 5.98 | **5.81** | | MF-RAE | 2.03 | 2.00 | **1.91** | | iMF-XL/2 | 1.72 | 1.70 | **1.63** | Although the FID gap on MF-XL/2 is small, RAFL also achieves lower FID than direct gating on each additional backbone. The reductions are 0.17, 0.09, and 0.07 on MF-B/2, MF-RAE, and iMF-XL/2, respectively.

W4 & Q2. Implementation details and reproducibility.

We have expanded Appendix A.3 (pp. 17–18) and added Algorithm 1 (p. 25) to specify the retrieval protocol, calibration schedule, predictor optimization, and update order. We summarize the key implementation details below.

Temporal compatibility and retrieval. For a query with transport condition (ri,ti)(r_i,t_i), historical entry jj is eligible when max⁡{∣ri−rj∣, ∣ti−tj∣}≤0.05.\max\{|r_i-r_j|,\ |t_i-t_j|\}\le0.05. Among eligible entries, we retrieve the K=32K=32 most similar representations using cosine similarity and compute normalized softmax weights aija_{ij} with temperature τ=0.1\tau=0.1.

Calibration and valid queries. Calibration is scheduled every 10 student steps. Queries with fewer than KK eligible neighbors are excluded. We also exclude queries for which 1−si≤10−61-s_i\le10^{-6}, where si=∑jaij2s_i=\sum_j a_{ij}^2, to avoid unstable finite-sample corrections. Each scheduled calibration step with valid targets performs one predictor update; if no valid queries remain, only the predictor update is skipped.

Buffer management and update order. The FIFO buffer stores up to 16,384 tuples (hj,rj,tj,δj)(h_j,r_j,t_j,\delta_j). Current representations are extracted using the frozen encoder, while target-space disagreements are computed using the current student snapshot. Historical retrieval and scheduled predictor updates occur before the student update; current tuples are appended afterward, with the oldest entries evicted when capacity is exceeded. This prevents current queries from serving as their own neighbors. Historical disagreements are reused as cached values computed at their generation iterations, rather than recomputed under the current snapshot.

Predictor architecture and optimization. We normalize the representation as ht/max⁡(∥ht∥2,10−8)h_t/\max(\|h_t\|_2,10^{-8}) and concatenate the scalar inputs (r,t)(r,t). The predictor uses two hidden layers of width 256 with SiLU activations and two Softplus-transformed risk outputs.

The predictor is trained against detached calibration targets using log-space Huber loss, with Huber threshold 1 and log-stability constant 10−610^{-6}. The two component losses are summed and averaged over valid queries. We use AdamW with constant learning rate 3×10−43\times10^{-4}, betas (0.9,0.999)(0.9,0.999), weight decay 10−410^{-4}, optimizer epsilon 10−810^{-8}, gradient clipping norm 1, and reliability-loss weight β=1\beta=1.

Warm-up and gradient routing. During the first 1,000 steps, the student uses conditional supervision with λwarm=0\lambda_{\mathrm{warm}}=0, while buffer collection and scheduled predictor updates remain active. After warm-up, predicted coefficients are used if the predictor has received at least one valid calibration update; otherwise, λ=0\lambda=0 is retained. The teacher and encoder remain frozen. Both the mixing coefficient and finite-time student target are detached, so Lrel\mathcal L_{\mathrm{rel}} updates only QψQ_\psi, while LRAFL\mathcal L_{\mathrm{RAFL}} updates only the student.

Q1. Fixed-state evolution of predicted risks and teacher reliance.

Following the reviewer’s suggestion, we added a fixed-state checkpoint analysis in Appendix A.4, Figure 4 (p. 21). We track four held-out transport states at 2k, 5k, 10k, 15k, and 20k post-training steps. Within each state, the image, class label, noise realization, and transport times remain fixed. All four states use (r,t)=(0.2,0.8)(r,t)=(0.2,0.8).

At each checkpoint, we record V~ξ\widetilde V_\xi, B~ξ\widetilde B_\xi, and the resulting teacher coefficient λ^ξ\hat\lambda_\xi. Between 2k and 20k steps, State A's V~ξ\widetilde V_\xi decreases from approximately 0.80 to 0.20, while B~ξ\widetilde B_\xi increases from 0.18 to 0.28, reducing teacher reliance from 0.82 to 0.42. State B shows the opposite trend, with teacher reliance increasing from 0.32 to 0.66 as B~ξ\widetilde B_\xi decreases relative to V~ξ\widetilde V_\xi. State C exhibits non-monotonic changes, while State D remains near 0.67 as both risks decrease approximately proportionally.

Although the teacher and encoder remain frozen, the student-induced target-space geometry and risk predictor evolve. Teacher reliance therefore follows changes in relative predicted risks and need not share a common monotonic schedule across states.

The rebuttal addressed my concerns. So I raised Presentation and Informativeness scores, and increased my Confidence. However, I still feel the paper is somewhat short of a strong accept, so I maintain my rating at 3. (After all, 4 is the maximum score.)

We sincerely thank the reviewer for the thoughtful and constructive feedback. Your insightful and targeted suggestions have significantly improved the clarity, presentation, and overall quality of our paper.

Reviewer_HWm5OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2(+1)
Soundness
2
Read review
Paper summary

This paper proposes a new technique for finetuning Flow Map Model (FMM) based on the idea of risk adaptivity. The authors define risk as the expected MSE of the targets given the "transport state" ξ=(zt,r,t)\xi = (z_t, r, t), which is a function of the interpolation weight λ\lambda. The paper then analyzes this risk function for building a new training strategy for FMMs. Empirical results on CIFAR-10 and ImageNet-256 demonstrate empirical gains of the method as a plug and play over existing methods.

Critical strengths

For me, the main strong point of this paper is its empirical evaluation. It considers a relevant benchmark (ImageNet-256), and demonstrates that the proposed method works well as a plug and play over existing models (e.g., MF-B), with a consistent improvement of performance over the base FMMs.

Critical weaknesses

The overall problem is that the paper seems harder to read than it needs to be. For instance, the authors introduce, in the very beginning, the transport state as ξ=(zt,r,t)\xi = (z_t, r, t), however, I think that largely obscures notattion throughout the entire paper, making it harder to parse. For instance,

  • vξ⋆=v⋆(zt,t)v_{\xi}^{\star} = v^{\star}(z_t, t) does not depend on rr (section 2, line 86) and hξ=sg[Pool(Hηℓ(zt,t))]h_{\xi} = \text{sg}[\text{Pool}(H_{\eta}^{\ell}(z_t, t))] only depend on ztz_t and tt, so readers must keep in mind the gymnastics around the definition of ξ\xi when reading these identities

  • The objects JξJ_{\xi} and Tξ\mathcal{T}_{\xi} could have been better described at first use. As far as I understood it, JξJ_{\xi} is a Jacobian matrix, while Tξ\mathcal{T}_{\xi} seems to be an affine map of the supervision target ww. If that is correct, could the authors specify that at first use? Say, around line 100 in Section 2, and around equation (2)? On top of this, if I am not missing something, the authors never mention the space they work on, which I assume to be the standard d−d-dimensional Euclidean space. Mentioning that would help ground intuition, and the objects JξJ_{\xi} and Tξ\mathcal{T}_{\xi} would be better defined.

  • vϕ,ξv_{\phi,\xi} is not properly defined at 1st use (1st paragraph of page 3, around lines 108 - 110), and ϕ\phi is not properly defined either.

  • The risk functional Rξ(λ)R_{\xi}(\lambda) is a function of the coefficient λ\lambda. It should be at least highlighted/warned in prose that this is different from the usual notion of risk in machine learning, which is a functional defined over functions (e.g., the velocity field vθv_{\theta} vs. the ground-truth v⋆v^{\star}).

  • I think the insights of the paper are currently scattered along sections 3 and 4, without a centralized exposition of how the proposed Risk Adapted Flow Map Learning. In this sense, an algorithm box and an illustrative figure would make the paper much clearer.

Questions for authors

N/A


Here is some guidance for the authors rebuttal. While I am initially leaning on a reject, if the authors clarify the problems I raised in the weakness, and especially provide a centralized algorithm for their paper (either in the main paper, if the authors can find the space, or in the appendix) and a clear illustration figure of how their algorithm works, I would gladly consider raising my score.


Post-rebuttal: the authors have added a figure summarizing their algorithm, and an algorithm box detailing the steps. I now consider this submission to be easier to read.

6 comments

We thank the reviewer for the constructive suggestions. We have clarified the notation and added an overview figure and algorithm to explain the training procedure.

W1. Clarifying the dependence on the transport state.

We now write the marginal velocity explicitly as v⋆(zt,t)v^\star(z_t,t) and the representation as ht=h(zt,t)h_t=h(z_t,t), making their independence from rr explicit. Section 2 now explains that the marginal and sampled conditional velocities do not involve the interval endpoint rr, while the full state ξ=(zt,r,t)\xi=(z_t,r,t) is used for finite-time quantities that also depend on rr (lines 86-92). Section 4.1 similarly clarifies that hth_t does not depend on rr; the risk predictor receives (ht,r,t)(h_t,r,t), with the transport condition supplied separately (lines 266-269).

W2. First-use definitions and underlying spaces.

We now explicitly define Jξ∈Rd×dJ_\xi\in\mathbb R^{d\times d} as the spatial Jacobian of uθˉu_{\bar\theta} and Tξ:Rd→Rd\mathcal T_\xi:\mathbb R^d\to\mathbb R^d as the affine map from instantaneous supervision to a finite-time target. These definitions appear at first use in Section 2, immediately before Eq. (2), together with an explicit specification of the underlying Euclidean space Rd\mathbb R^d (lines 86 and 102-106).

W3. Defining the pretrained guidance and its parameters.

The Complementary Supervision paragraph in Section 2 now introduces ϕ\phi as the parameters of the pretrained teacher and states that they remain fixed throughout training (lines 111-115). For the MF instantiation, we explicitly define the instantaneous guidance through the frozen FMM teacher’s diagonal evaluation: vϕ(zt,t):=uϕ(zt,t,t).v_\phi(z_t,t):=u_\phi(z_t,t,t). This also distinguishes the fixed teacher parameters ϕ\phi from the current student snapshot θˉ\bar\theta used to construct the target-space operator.

W4. Clarifying the meaning of the risk Rξ(λ)R_\xi(\lambda).

Following the reviewer's suggestion, we have added an explicit clarification immediately after its definition in Section 3.1 (lines 137-143). At a fixed transport state and target-network snapshot, Rξ(λ)R_\xi(\lambda) is a scalar function of the mixing coefficient that measures the conditional error of the supervision target. It does not denote the population prediction risk of the learned generative model. This clarification specifies the scope of the subsequent optimal-coefficient analysis.

W5. Providing a centralized explanation of the training procedure.

To provide a unified account of RAFL, we have added Figure 1 in the main paper (p. 5) and Algorithm 1 in the appendix (p. 25).

The workflow is now explained explicitly. RAFL first computes target-space disagreements using conditional supervision, frozen teacher guidance, and the current student snapshot. Periodically, historical disagreements within temporally compatible representation neighborhoods provide calibrated risk labels for training QψQ_\psi. After warm-up, the predicted risks determine a detached teacher coefficient, which mixes conditional and teacher supervision before the finite-time target transformation. The student regresses the resulting detached target through LRAFL\mathcal L_{\mathrm{RAFL}}, while QψQ_\psi is trained only through Lrel\mathcal L_{\mathrm{rel}}. Current tuples are appended to the FIFO buffer after the student update, ensuring that retrieval uses historical samples.

Figure 1 illustrates the relationship between these components, while Algorithm 1 specifies warm-up, calibration scheduling, and update order. Appendix A.3 additionally provides the predictor architecture, regression loss, temporal compatibility rule, and optimization settings.

Reviewer_HWm5OpenReview ↗

Small response

Dear authors,

I acknowledge receiving your answers via openreview. However, note that your manuscript is not up to date, making the changes in W5 impossible to check.

Dear Reviewer,

We sincerely apologize for the oversight in not updating the PDF. The file has now been updated. Could you kindly check the figure on page 5 and the algorithm on page 25?

Best regards, On behalf of all the authors of Submission 37638

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.