acceptodds
Withdrawn from ICLR 2027

Multi-Timescale Verification for Test-Time Adaptation of Social Agents

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Self-evolving agents have demonstrated the potential to improve through accumulated experience, particularly in software engineering, where executable feedback can evaluate proposed changes. For social agents in unknown environments, however, both perceived events and their behavioral explanations are uncertain. Repeated reflection can therefore reinforce errors rather than correct them. This raises a central question: how can social agents build on past interactions without compounding unreliable beliefs? We introduce the *Verified Social Evolution Loop* (VSEL), a framework for social agents to evolve through verification at multiple timescales in unseen continuous environments. Within episodes, a *Physical Consistency Verifier* checks perceived states against spatio-temporal constraints. The retained evidence is consolidated into a *decision map* and provided to the visual-language model (VLM) for event-triggered reasoning and evidence reassessment in a zero-shot manner. Across episodes, a *Hypothesis Verifier* evaluates behavioral hypotheses against accumulated support and counterevidence. Using only egocentric RGB-D observations and poses, VSEL sets new state-of-the-art (SOTA) results across benchmarks on the Habitat 3.0 platform for social agents. Particularly, in social navigation, VSEL achieves an episode success rate of 0.73, surpassing the strongest compared reinforcement-learning baseline by 70% in relative terms. Overall, our framework offers a practical route to social agents that accumulate and revise knowledge through verified experience.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Social Agents, Test-Time Adaption, embodied AI, Self-evolving, Zero-Shot Learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_DXmQOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

VSEL combines physical consistency checks, event-triggered VLM reasoning, and cross-episode hypothesis verification for embodied social agents. The framework is well motivated and achieves strong navigation results, but the reliability and distinct contribution of its verification mechanisms require further evidence.

Critical strengths

Navigation success improves from 0.43 to 0.73 while collisions decrease substantially. Component ablations and additional VLM backbones support the approach’s utility. The distinction between uncertain observations and behavioral hypotheses is conceptually valuable.

Critical weaknesses
  1. The unique contribution of the proposed method relative to existing feedback mechanisms needs to be more clearly established. Further comparisons and discussions are needed to clarify its relationship to relevant studies, including feedback-driven planning in Inner Monologue, execution monitoring in DoReMi, and experience-guided adaptation in AeroWeaver.

  2. The robustness and statistical uncertainty of the adaptation mechanism are insufficiently addressed. Figure 4 presents a continuous evaluation of 1,200 episodes involving only a single collaborator, whereas the main rearrangement experiments involve four collaborators, resulting in differences in experimental settings. Furthermore, in the rearrangement task, the proposed method achieves only a marginal improvement in success rate, from 72.13% for the strongest reported reinforcement learning baseline to 72.67%.

  3. Runtime and reproducibility details remain incomplete. Appendix G reports an average VLM inference time of approximately 0.406 seconds per executed step, but it remains unclear whether the collaborator continues moving during model inference.

Questions for authors

1.Do the gains persist against conventional filtering and matched reflection-memory baselines?

2.How reliably does HV identify false hypotheses and prevent contradicted beliefs from influencing subsequent strategies?

3.Are adaptation gains consistent across collaborators and randomized episode orders?

Reviewer_8qbXOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes VSEL, a framework that lets a social agent built on a frozen VLM improve over repeated interactions without turning perception and reasoning errors into persistent beliefs. Verification runs at three timescales: a physical consistency verifier, written offline as code by a VLM, filters perceptual claims at each step, VLM decisions are reviewed when a queried entity is detected in the current image, and hypotheses about the collaborator's behaviour are verified across episodes and turned into a strategy. On Habitat 3.0, VSEL raises Social Navigation episode success from 0.43 to 0.73 and reports state-of-the-art results on Social Rearrangement.

Critical strengths

• Without privileged human positions or an oracle detector, VSEL reaches 0.73 episode success versus 0.43 for the strongest RL baseline (Table 1), and the gain holds with two other VLM backbones (Table 2b). This is the main reason I see value in the paper: filtering perceptual claims before VLM reasoning lets a frozen VLM outperform task-trained RL policies, even though those policies use an oracle human detector.

• Treating occlusion or missing evidence as unresolved rather than as evidence of absence, for both perceptual claims and behavioural hypotheses, is a principled design choice that other VLM-agent pipelines could adopt.

Critical weaknesses

(major) The title and abstract centre on agents that evolve across episodes, but the effect of the cross-episode loop is isolated only in Figure 4: a single collaborator, with no repeated runs or variance reported. Its values (success rate of about 75–85%) differ from Figure 3 (72.67%) without explanation, and after roughly 400–500 iterations the success-rate and efficiency curves of VSEL stop improving and slightly decline, which does not support the statement that VSEL "progressively improves its strategy". The paper also does not report whether the retained hypotheses match the collaborator's actual behaviour. Without this, the claim that verification prevents errors from becoming persistent beliefs is not demonstrated.

• (major) Removing the physical consistency verifier drops episode success to 0.25 (Table 2), below the 0.32 of the ET-Baseline (Table 1), which lacks the verifier, the decision map, the review step and hypothesis verification. This suggests the other components hurt performance without the verifier, yet the verifier is never added to the baseline on its own, and its accuracy against simulator ground truth is not reported. As removing it costs more than VSEL's entire gain over ET-Baseline, it remains unclear what the verifier actually filters.

• (major) None of the reported results include standard deviations, confidence intervals, or the number of runs (Tables 1–2, Figures 3–4). This matters most where differences are small: on Social Rearrangement, the success-rate gain over Plan-Pop3 is 0.54 points (72.67 vs. 72.13, Figure 3), yet Section 4.2 states that VSEL "substantially improves task success", and removing the decision map or the review step changes episode success by only 0.02–0.04 (Table 2). The largest rearrangement difference is in collision rate (0.07 vs. at best 0.20 for the RL baselines).

• (minor) The "Zero-shot" label in Table 1 should be qualified: no privileged information is used at test time, but the verifier's rules and thresholds were refined against simulator ground truth on training episodes (Appendix C).

• (minor) The main VLM backbone (Qwen3.8-27B-FP8) is named only in Appendix F, although Appendices E.3 and F state that it is specified in Section 4.1.

Questions for authors

1 How do Figures 3 and 4 relate, and are the Figure 4 curves running averages? Per-episode results with variance over several collaborators and seeds would address my concerns.

2 How accurate are the retained hypotheses against the known behaviour of the planner-based collaborators?

3 How often are claims labelled Verified, Unverified and Falsified actually correct? How does ET-Baseline perform with the verifier alone, and why does VSEL without the verifier fall below ET-Baseline?

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

The paper proposes VSEL, a zero-shot agent for Habitat 3.0 social tasks built around a frozen VLM (Qwen3.8-27B-FP8) and SAM 3.1 grounding, with verification at three timescales. At every step, a Physical Consistency Verifier (PCV), an executable program synthesized offline by GPT from training episodes with simulator feedback, labels each perceptual claim VERIFIED, UNVERIFIED or FALSIFIED using depth support, geometric plausibility and a feasible-displacement bound (Eq. 9). Non-falsified claims are rendered into a decision map. The VLM is invoked on events, and a positive segmentation match for a queried entity triggers a second review call. Across episodes with the same collaborator, a VLM-based Hypothesis Verifier assigns support, contradiction or unresolved verdicts to behavioral hypotheses, new hypotheses are discovered from unexplained evidence every B=8 episodes, and hypotheses with at least one supporting verdict are compressed into a strategy prompt for the next episode. Using only egocentric RGB-D and poses, VSEL reports episode success 0.73 on Social Navigation against 0.43 for SDA-S2, and SR 72.67 on Social Rearrangement against 72.13 for Plan-Pop3, with ablations of within-episode components and VLM backbones.

Critical strengths
  1. A zero-shot agent without privileged test-time inputs outperforms trained RL policies that use an oracle human detector on Social Navigation. Table 1 lists hDet = GT for DDPPO, Proximity and SDA-S2, whereas VSEL uses no privileged information at test time, yet VSEL reaches ES 0.73, F 0.56 and CR 0.10 against SDA-S2's 0.43, 0.39 and 0.57. The collision rate is also below that of the map-privileged Heuristic Expert (0.52). The input asymmetry runs against the proposed method, so the main Social Navigation result is not explained by extra information.
  2. The evidence states are defined carefully, and verification is separated from semantic correctness. Section 3.2 and Appendix C.3 distinguish UNVERIFIED (insufficient support, for example occlusion) from FALSIFIED (conflict with evidence), Appendix A.3 specifies that a falsified match still triggers review without becoming admissible evidence, and the Limitations section states that physical consistency does not guarantee semantic correctness. This makes the design inspectable and the claimed scope clear.
Critical weaknesses

The claim that the PCV works by keeping erroneous perception out of the decision context is supported only by end-task deltas, and those deltas have a pattern the paper does not explain. In Table 2a, removing the PCV lowers ES from 0.73 to 0.25. This is below ET-Baseline (0.32, Table 1), which removes the PCV and also the decision map, the post-grounding review and hypothesis verification (Appendix E.3). Adding the decision map and review without the PCV therefore makes the agent worse than having none of these components. One reading is that the map renders unfiltered claims that mislead the VLM, so the PCV mainly repairs a failure that the decision map introduces; under that reading the 0.48 drop overstates the PCV's standalone value, and the comparison that measures verification is ET-Baseline with and without the PCV. The size of the gain also depends heavily on the backbone: with gemma-4-12B-it, ET-Baseline already reaches ES 0.54 and VSEL 0.68 (Table 2b), so the 128% relative improvement over ET-Baseline in Section 4.2 is measured with the backbone on which ET-Baseline is weakest (0.32). Finally, the paper never reports how often SAM 3.1 produces erroneous claims during test episodes or how accurate the PCV verdicts are. The main rule shown (Eq. 9, Figure 5) is a feasible-displacement gate, and without verdict-level numbers it is not possible to tell what the synthesized checks catch. Simulator reference positions are already used for scoring at test time (Appendix E.1), so a confusion matrix of PCV verdicts against reference humanoid positions, and the rate of perception errors with and without the PCV, can be computed from existing evaluation logs. Cross-episode test-time adaptation, which the title and framing center on, has the weakest evidence. (i) The Social Navigation headline (ES 0.73, "70% in relative terms" in the abstract) is ablated only over within-episode components (Table 2a), and the paper does not say whether hypothesis verification is active in Social Navigation or report a row without it. (ii) On Social Rearrangement, VSEL's SR of 72.67 exceeds Plan-Pop3 (72.13) and Plan-Pop4 (72.02) by 0.54 and 0.65 points, and its RC of 0.47 is below Plan-Pop4's 0.52 (Figure 3, right). Against the RL baselines, the clear gains are in RE (122.33 against 116.59 for Plan-Pop1) and CR (0.07 against 0.20 for Plan-Pop2), and ET-Baseline already reaches CR 0.11 without any verification. (iii) The only evidence for hypothesis verification is Figure 4, given as curves without numbers, and the "w/o hypothesis verification" variant is not defined: it could accept all discovered hypotheses, use none, or skip discovery. (iv) The setting is stated inconsistently: Section 4.3 evaluates "1,200 episodes with a fixed human collaborator", while the caption of Figure 3 reports results "with 4 unseen planner-based collaborators". Reporting SR, RE and CR for the first and last portions of each partner stream, for the full method, the w/o-HV variant and a variant whose cross-episode knowledge K_(<e) is reset every episode, would directly measure adaptation. These numbers are available from the runs already performed.

Questions for authors
  1. On the Social Navigation test episodes, what are the precision and recall of FALSIFIED and VERIFIED verdicts for humanoid-position claims against the simulator reference positions, and how often does SAM 3.1 produce a humanoid claim that is more than the following tolerance away from the reference? In the "− PCV" ablation, are all claims rendered on the decision map as admissible evidence?
  2. Is cross-episode hypothesis verification active in the Social Navigation results of Table 1? If so, what is VSEL's ES without it?
  3. For Social Rearrangement, how many collaborators and partner streams are evaluated, and how many episodes per stream? What exactly does "VSEL w/o hypothesis verification" do, what are its final SR, RE and CR, and how do the full method's SR, RE and CR compare between early and late episodes of a stream?

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

Related papers

Discussion (0)

Sign in to comment.