acceptodds
Withdrawn from ICLR 2027

EvoBelief: Evolving Behavioral Beliefs via Conformal Multimodal Diffusion

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

In driving, the ego car continually negotiates with other road users under imperfect perception, and the same observations often admit more than one reasonable maneuver—at a merge, a yield, or an unprotected turn the scene does not yet reveal whether to proceed or hold. Such interactive, ambiguous moments call not for a single confident trajectory, but for an explicit account of the competing behaviors the scene still admits. Diffusion-based trajectory planners can represent multiple plausible future behaviors, yet existing methods largely exploit this multimodality for trajectory generation rather than behavioral uncertainty estimation. We present EvoBelief, which treats the multimodal distribution of generated trajectories as explicit evidence of alternative driving behaviors, such as waiting or committing, and converts this evidence into safety-aware decisions. Since sample frequencies are not calibrated probabilities, EvoBelief uses split conformal prediction to calibrate behavioral evidence into statistically controlled safety decisions, with an online adaptive mechanism for deployment shift. On Bench2Drive, EvoBelief improves success rate from 89.25% to 91.36% over state-of-the-art. More broadly, EvoBelief offers a new paradigm that reads generative multimodality as an evolving behavioral belief and turns it into safe interactive driving decisions, rather than treating diffusion merely as a trajectory generator.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Diffusion policy, closed-loop planning, end-to-end autonomous driving

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_BUDbOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes a decision strategy for a diffusion-based driving planner. It uses Monte Carlo sampling to generate trajectories and speed profiles, then applies a speed-based rule to estimate the proportion of samples that suggest waiting. A threshold calibrated offline and adjusted online determines whether to select a slower trajectory or proceed. The authors show that this strategy improves success rate over the state of the art on the Bench2Drive Carla benchmark.

Critical strengths

The proposed strategy combines offline threshold calibration with online adaptation and empirically improves success rate on Bench2Drive. Ablations show performance gain over keeping the wait score threshold fixed and a relevant baseline PlanCP while the latter is not well explained.

Critical weaknesses
  1. The writing of the paper needs significantly improved to help readers understand the method and evaluate the novelty compared to the existing works. Many terms such as "behavioural belief" or "trajectory-level return band" are not explained and need clear math definitions if possible.

  2. Comfort score drops nearly half potentially caused by frequent stalling action makes the method sound impractical.

  3. Assuming constant velocity for surrounding agents can make the risk estimates unreliable. Reactive agents can behavior in many ways that is not captured by this assumption and I suggest the authors to look into the large existing research on modelling multi-agent behaviours for traffic agents.

Questions for authors
  • Could the author explain the baseline "PlanCP's trajectory-level return band"? Could you make it clear for the reviewers on the method contribution compared with the existing works?

  • Could the author find out the causes of the performance drop in the comfort score ?

  • This work builds on the assumption that diffusion based planners can produce safe trajectory samples, what happens if all the samples are concentrated in the risky regime?

Reviewer_pYQ9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

EvoBelief is a decision layer for diffusion-based driving planners in interactive scenes where the ego vehicle must choose to wait or commit.

The planner jointly predicts a path and a speed profile. At each control cycle it draws M=30M = 30 trajectories and labels each as wait or commit with a hand-crafted dynamics criterion DD. The fraction labeled wait, the sampled wait mass p^k(wait)\hat p_k(\text{wait}), serves as an uncalibrated score.

Split conformal prediction on expert wait frames sets a threshold τ(α)\tau(\alpha), so that the unsafe-commit rate USR=P(π=commit∣Z=wait)\mathrm{USR} = P(\pi = \text{commit} \mid Z = \text{wait}) is at most α\alpha under exchangeability.

At deployment, Proxy-ACI adapts αt\alpha_t with a dense risk-map proxy and caps it at α1=α\alpha_1 = \alpha, so the adaptive gate is never less cautious than the static one (Lemma 2).

On Bench2Drive, EvoBelief reports SR 91.36 ± 0.79 and DS 96.52 ± 0.62, against 89.25 ± 0.50 and 96.34 ± 0.55 for BridgeDrive. The paper also reports a 2 × 2 ablation over representation and belief source crossed with four decision layers, a PlanCP re-implementation, offline coverage curves, and an analysis of the 8 collisions in the deployment.

Critical strengths

Treating the sampled wait mass as an uncalibrated score and calibrating it class-conditionally is a clear and checkable design.

Section 2.3 states that p^k(wait)\hat p_k(\text{wait}) is a Monte Carlo estimate under qθq_\theta, not the true posterior pk∗p^*_k.

Section 2.4 calibrates only the safety-critical class, with the score si=−log⁡p^i(wait)s_i = -\log \hat p_i(\text{wait}).

Figure 3 (left) shows the empirical USR following α\alpha on an exchangeable split.

Section 2.5 states that trajectory selection lies outside the certificate.

Appendix I is candid about failures. 5 of the 8 collisions occur with zero sampled wait mass, where a gate that requires at least one sampled wait trajectory cannot force a wait (Proposition 3).

The gate and the αt\alpha_t update cost under a millisecond, while sampling M=30M = 30 trajectories roughly doubles the per-cycle latency (447.4 ms vs. 211.9 ms for M=1M = 1, Table 6).

This interface between a generative planner and a binary safety decision is the most useful part of the paper.

The closed-loop gain is of a useful size and is reported over three seeds. On Bench2Drive, EvoBelief reaches SR 91.36 ± 0.79, against 89.25 ± 0.50 for BridgeDrive and 86.8 ± 0.7 for LEAD (Table 1). Section 3.1 states that all variants share the same TransFuser-v6 backbone. The 2.11-point gain corresponds to about 4.6 of the 220 routes and is larger than the seed spread of either method. This supports the claim that the full system improves closed-loop success on this benchmark. The DS gain (+0.18) is within the seed spread.

Critical weaknesses

The paper attributes the SR gain to calibration and to the combination of its three components, but the reported controls do not isolate these effects.

Most cells are single runs. Section 3.1 states that only the full configuration was run with three seeds. Apart from the two no-gate cells that repeat the Table 1 means, every other cell in Table 2 is a single run. One route on Bench2Drive is 0.45 SR points, and the seed standard deviation of the full method (0.79) is close to two routes.

Table 2 contradicts its own caption. BridgeDrive with Proxy-ACI and EvoBelief with fixed-α\alpha both reach 89.54, above the 89.25 that only the bottom-right cell is said to clear (caption and Section 3.2).

The step that carries the gain is not explained by calibration. Most of the gain comes from the step from fixed-α\alpha to Proxy-ACI (89.54 to 91.36). Section 3.4(c) explains this step as restored coverage. However, Figure 3 (right) shows that offline, under the route-disjoint split, the static threshold keeps USR at about 0.10 at the deployed α=0.10\alpha = 0.10; under-coverage appears only at α=0.15\alpha = 0.15 and 0.200.20. Online USR is not reported.

The same step helps rows without a faithful wait signal. The step from fixed-α\alpha to Proxy-ACI also adds 1.36 points in the BridgeDrive row, whose scalar head the paper says cannot express the wait event (Proposition 4). It adds 1.36 points in the LEAD-variant row as well, whose MC-dropout belief "does not reliably populate" the modes (Section 3.2(b)). Fixed-α\alpha is below no-gate in three of the four rows.

This pattern fits a risk-triggered brake. Each errtproxy\mathrm{err}^{\text{proxy}}_t event on a commit lowers αt\alpha_t by γ(1−α)=0.045\gamma(1-\alpha) = 0.045. Two such events bring αt\alpha_t below the level at which τt=0\tau_t = 0, after which a single wait sample among 30 forces a wait (Figure 4 shows several stretches with τt=0\tau_t = 0). The paper's contribution is the claim that calibrating a behavioral belief produces the gain, so this gap weighs most in my assessment.

The claims overstate Lemma 2 and Proposition 2. Eq. 6 is a split-conformal bound for the static gate on frames exchangeable with the expert wait frames in ScalS_{\text{cal}}. For the deployed gate, Lemma 2 shows only that it waits whenever the static gate waits, and Section 2.4 itself says this gives no per-step bound at deployment. Still, Section 3.4(c) states that online adaptation "restores coverage". Section 2.4 and Remark 3 state that the proxy-trigger rate cannot persistently exceed α\alpha, although Proposition 2 contains the term (1−α)L/N(1-\alpha)L/N, which is neither bounded nor reported.

The gate is defined in two ways. Eq. 8 and Algorithm 1 wait when p^k(wait)≥τt\hat p_k(\text{wait}) \ge \tau_t. Lemma 2 and Proposition 3 add the condition Mp^k(wait)≥1M\hat p_k(\text{wait}) \ge 1. The two coincide whenever τt>0\tau_t > 0, and differ when τt=0\tau_t = 0, which happens once αt\alpha_t falls below about p0=0.024p_0 = 0.024 (and whenever αt≤0\alpha_t \le 0).

Under Algorithm 1, that state is absorbing: every cycle waits, and αt\alpha_t is updated only on commits.

Under the gate of Lemma 2, αt\alpha_t can recover, but then lower-saturation events LL enter Proposition 2. Remark 3 itself says that saturation "empirically occurs". Figure 4 shows αt\alpha_t rising while τt=0\tau_t = 0 and the wait mass is 0, so this second gate appears to be the one that was run.

USR does not measure collisions. USR counts commits on frames where DD labels the expert's realized speed profile as wait, so it measures disagreement with the expert's braking. The bridge P(collision)≤α+ϵdeployP(\text{collision}) \le \alpha + \epsilon_{\text{deploy}} in Appendix I is not derived and rests on 8 collisions.

Appendix I's conclusion does not fit Figure 5b. Appendix I concludes that the remaining collisions are "exactly the p0p_0" and that the calibration layer is sound. But in Figure 5b one collision occurs with a wait mass of about 0.57, above τ=0.23\tau = 0.23, where the gate would already wait. That failure lies in trajectory selection, outside the certificate. This matters because the abstract presents the decisions as statistically controlled for safety.

Questions for authors

(Clarification) Which gate was used in the experiments: Eq. 8 and Algorithm 1, or the support-constrained gate of Lemma 2 and Proposition 3? Over the 220 routes, please report: the fraction of cycles with τt=0\tau_t = 0; the ratio L/NL/N from Proposition 2; the online USR; the denominator behind "6.85 vs. 2.47 expected fires" in Appendix I. If the window is the 20 cycles used for recall, a suite-wide rate of 2.47 fires per window is above α=0.10\alpha = 0.10 per cycle.

A single algorithm with matching statements is needed before I can accept the online safety claims.

Does Proxy-ACI help through calibration or through the risk map? Two controls would separate them, both keeping the fixed-α\alpha gate and leaving αt\alpha_t unchanged:

for a fixed number of cycles after each errtproxy\mathrm{err}^{\text{proxy}}_t event, always wait (the slowest feasible sample); for the same number of cycles, wait whenever Mp^k(wait)≥1M\hat p_k(\text{wait}) \ge 1.

Together with three seeds for the fixed-α\alpha and Proxy-ACI cells of all four rows, this would show whether the step from 89.54 to 91.36 comes from calibration, and why the same step adds 1.36 points in the BridgeDrive and LEAD-variant rows. If the calibrated gate clearly beats such rules, my main concern would be resolved. Relatedly, is the drop in Comfortness from 23.24 (BridgeDrive) to 12.68 (Table 4) caused by forced waits, and how do the 19 failed routes split between collisions and other failure types?

(Clarification on the protocol) The risk map is a conditioning input to the diffusion planner (Appendix G). Do the LEAD and BridgeDrive heads in Table 2 also receive it? Where do ScalS_{\text{cal}} (nn = 1,398) and PlanCP's 38,826 calibration frames come from relative to the 220 evaluation routes? How were γ\gamma, τrisk\tau_{\text{risk}} and ρmax⁡\rho_{\max} chosen?

How is p^k(wait)\hat p_k(\text{wait}) computed for the scalar-speed heads, given that DD (Algorithm 2) takes a speed profile?

Since the gain is about 4.6 routes, these choices matter. In Appendix C, PlanCP uses a fixed threshold qlo=−0.761q_{lo} = -0.761, a deterministic deceleration-to-stop as its wait action and the medoid plan as its commit action, and Variant 2 is retrained. Was Proxy-ACI applied to either variant? The claim in Section 3.3 that the gap is "attributable entirely to the calibrated object" needs a comparison that also shares the wait action and the online adaptation.

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

This paper presents EvoBelief for wait-or-commit decision-making in interactive driving. It uses a diffusion planner to generate diverse path and speed candidates, then estimates a belief signal from the proportion of candidates classified as waiting. A class-conditional split-conformal gate uses this signal to control decisions, while a risk-map proxy adjusts the threshold online according to scene risk. Experiments on Bench2Drive show improved driving success and more adaptive behavior in uncertain interactions.

Critical strengths
  1. Waiting candidates can provide useful decision evidence even when most samples favor proceeding. Calibrating the sampled wait fraction is a concrete way to use this evidence, and the paper appropriately treats sample frequency as an untrusted score rather than a calibrated posterior.
  2. The full configuration achieves the highest reported success rate. The shared-backbone ablations, risk-map-only baseline, and analysis of collisions with zero wait mass provide useful evidence about the system’s performance and failure cases.
Critical weaknesses
  1. The main algorithm and the theoretical analysis use different gates Equation (8) and Algorithm 1 trigger waiting whenever the sampled wait fraction reaches the threshold. Lemma 2 and Proposition 3 additionally require at least one sampled wait candidate. When both the threshold and wait fraction are zero, the main algorithm waits while the appendix gate commits. This difference changes the policy to which the blind-spot lower bound and Figure 5’s explanation apply.

With a fixed calibration set, if the threshold reaches zero, Algorithm 1 repeatedly enters the wait branch. Because that branch leaves the adaptive level unchanged, the threshold stays zero. Remark 3 discusses saturation toward zero but does not establish whether an exact zero occurs in evaluation. The paper should state which gate is deployed and how it handles this case.

  1. The deployment safety claims exceed the established guarantees The offline bound controls how often the gate commits when the expert’s realized speed profile is labeled wait. This is a behavioral-label error, not a collision rate. Section 2.4 and Appendix D.3 explicitly state that the online gate is no more permissive than the static gate on the same input; they do not renew absolute coverage under deployment shift. Section 3.4(c) nevertheless says adaptation restores coverage, and the conclusion says calibration is preserved online. The paper should specify which error rate is restored and provide evidence for that claim.

Proposition 2 bounds risk-proxy triggers over commit cycles and includes the term (1 − α)L/N, where L counts lower-saturation events and N is the number of commit cycles. The bound need not approach α unless this term is controlled. Appendix I reports that all eight observed collisions had a warning within the preceding 20 cycles. This warning recall and the per-commit-cycle trigger frequency use different events and denominators; the reported recall alone does not establish the claimed collision bound of α plus ε.

  1. The experiments do not isolate the claimed source of improvement. The profile-diffusion configuration improves from 89.54 with a fixed gate to 91.36 with online adaptation. These results support the full system, but most ablations are single runs, limiting conclusions about small differences and component necessity. Sections 3.2(c) and 3.3 further attribute the gains to wait evidence providing earlier warning than trajectory risk. The results do not isolate this explanation from changes in execution or how readily the controller waits.

In Appendix C, adapted PlanCP waits by deterministic deceleration-to-stop, whereas EvoBelief executes a sampled trajectory. The stronger PlanCP variant also uses an auxiliary training loss. This comparison does not establish that the calibrated object alone explains the performance gap. A comparison with matched candidates and execution rules, together with signal timing and efficiency, would test the proposed explanation more directly.

  1. Trajectory selection and scalar-speed ablations lack executable definitions Algorithm 1 does not define when spatial modes agree or how CONSENSUSMODEREPRESENTATIVE is selected. Appendix C defines a medoid for PlanCP, but does not establish that EvoBelief uses the same rule. TARGETSPEED also needs a definition for a speed-profile output. Section 2.5 says selection is within the chosen mode, yet the wait branch explicitly filters candidates only by risk, without requiring D(y)=wait. It then selects the slowest feasible sample, while Appendix H describes selecting the lowest-risk sample. These rules can produce different actions. The scalar-speed ablations also do not explain how they obtain wait scores when D requires a speed profile. The omissions affect both reproduction and interpretation of the online results, since the chosen trajectory determines the risk feedback used to update the threshold.
Questions for authors
  1. Table 2 contains two other configurations at 89.54, contradicting the statement that only the full configuration exceeds 89.25. Section 3.2(a) compares LEAD no-gate at 86.8 with LEAD-variant Proxy-ACI at 87.72, changing the decision layer as well as the representation. Please correct both statements.

  2. Which gate is deployed when the threshold and wait fraction are both zero? Does an exact zero occur in evaluation, and how does the system recover? Please report the number of infinite calibration scores and specify the quantile convention for nonpositive adaptive levels, so Algorithm 1 can be reconciled with Proposition 3 and Figure 5.

  3. What exactly is restored in Section 3.4(c), given the limits stated in Section 2.4? Please account for the saturation term in Proposition 2 and explain how Appendix I’s collision-window recall supports any claimed bound on collisions.

  4. Can the behavioral score be compared with trajectory-risk gating using matched candidates and execution rules? Please report when each signal triggers, together with success, waiting behavior, and efficiency, or narrow the claim that the calibrated object explains the entire gap.

  5. Please define mode agreement, consensus representative selection, and TARGETSPEED. Is the wait candidate set restricted to D(y)=wait? Please reconcile the slowest-feasible and lowest-risk descriptions and explain the fallback when no candidate passes the risk filter. How are wait labels and scores obtained for the scalar-speed ablations?

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

Related papers

Discussion (0)

Sign in to comment.