{"id":"45575088-b987-42b5-95b3-1ea7a21b076e","arxiv_id":"2508.05979","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"The manuscript's abstract (a 'learning by teaching' AI study) does not match its full text (an energy-based test-time adaptation paper for depth completion).","lead":"The abstract describes a computer science education experiment where students teach an AI, but the full text is a computer vision paper about adapting depth-completion models to new environments. These are different studies with different authors, so the claimed result cannot be checked from this manuscript.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Abstract promises a statistically significant education outcome, but the full text contains no education experiment; the submission's stated central claim is unverifiable from the manuscript.","rationale":"I read the abstract as the claim under review: teaching an LLM improves CS student performance. For this to hold, the paper must report an experiment. The full text is a completely different paper (ETA) with different authors and topic. This is not a subtle flaw in a derivation; it is an absent object. The reader's UNVERDICTED verdict is appropriate because no evidence in the submission supports the abstract's stated result. I do not treat the FGSM/OOD concern as decisive: the body text itself acknowledges that limitation, and a reviewer could assess the ETA contribution with that caveat. The structural mismatch, however, prevents any verification of the stated education claim. Hence my read leaves the verdict unchanged.","tokens_in":18925,"tokens_out":3011,"duration_ms":38021,"concrete_test":"Audit the submission for the claimed experiment: search all text and any ancillary files for 'Socrates', 'course', 'student', 'historical cohort', 'pre/post', 't-test', and 'p-value'. If none of these appear with actual results, the abstract's 'statistically significant improvement' claim has no supporting study in this submission. Optionally check the arXiv version history to determine whether the education manuscript was replaced in error.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim — that the active-learning method produced statistically significant improvements over historical cohorts — requires a controlled study of students using Socrates. The full text supplied is instead a depth-completion test-time adaptation paper (ETA) by different authors; no section, table, or supplementary material describes the course, Socrates, the historical cohorts, or the statistical test. The load-bearing premise that the manuscript is the study described in the abstract therefore fails, and the headline result cannot be checked. For the body's own claim, the Limitations section already concedes that FGSM adversarial perturbations may not capture extreme or semantic target shifts; that is an acknowledged caveat rather than the primary blocker. The unaddressed, decisive issue is the structural absence of the stated experiment.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The submitted manuscript pairs a cs.CY title/abstract describing a CS education study ('Learning by Teaching: Engaging Students as Instructors of Large Language Models') with a full text that is a computer-vision paper titled 'ETA: Energy-based Test-time Adaptation for Depth Completion' by a different author group. The abstract claims that the proposed Socrates active-learning method 'led to statistically significant improvements in student performance compared to historical cohorts,' but the body contains no course description, no Socrates system, no cohort comparison, and no statistical test. The full-text paper instead proposes an energy-based test-time adaptation method for depth completion: it trains an energy model on source predictions and FGSM-perturbed versions of them (Eqs. 5–7) and then updates a lightweight adaptation module at test time by minimizing an energy-based loss (Eqs. 8–11). The body reports consistent MAE/RMSE improvements over baselines across three indoor and three outdoor datasets and four depth-completion architectures (Tables 1–4).","tokens_in":1441,"tokens_out":1467,"duration_ms":66999,"significance":"If the abstract's education claim were substantiated, it would be a practically valuable contribution to CS education. However, that study is entirely absent from the submitted manuscript, so its significance cannot be assessed. The ETA method, if correct, is a plausible and potentially useful contribution to test-time adaptation for depth completion: the benchmark coverage is broad, the comparisons include relevant baselines, and the region-based energy formulation is an interesting idea. The paper does not, however, ship code, and the central training objective as written (Eq. 7) is mathematically inconsistent with the stated goal. Because the stated central claim of the submission is missing and a load-bearing equation in the body is erroneous, the manuscript in its current form is not publishable.","major_comments":[{"comment":"The abstract promises a statistically significant improvement in student performance from the Socrates active-learning method. The full text is a depth-completion test-time adaptation paper by different authors; there is no course, no Socrates system, no historical cohort, and no statistical test anywhere in the manuscript. The headline education claim is therefore completely unsupported. This is not a presentation issue that revision can fix within the current manuscript's scope; the described study is absent.","section":"Abstract vs. Full Text"},{"comment":"The energy-model training loss is written as Lenergy = -1/|Ωp| Σ y(x) log(y(x)/e(x)). This is the negative KL divergence between the target energy y and the predicted energy e. Minimizing it maximizes KL(y||e), driving the predicted energy away from the target, which is the opposite of the stated objective. The standard cross-entropy objective would be -Σ y(x) log e(x). As written, the central training objective contradicts the method's purpose and the results in Tables 1–4 cannot be attributed to the described loss.","section":"Eq. (7), Section 3.1"},{"comment":"The abstract reports an average improvement over the previous state of the art of 6.94% outdoors and 10.23% indoors. Section 4.2 reports average improvements over ProxyTTA of 5.36% (MAE) and 1.97% (RMSE) outdoors and 10.13% (MAE) and 10.33% (RMSE) indoors. The indoor number matches the mean of the two metric improvements, but the outdoor 6.94% does not match the mean (3.67%) of the reported MAE/RMSE figures. The paper must state exactly which comparison, which averaging, and which metric combination produces the abstract's numbers.","section":"Section 4.2 and Abstract"},{"comment":"The main results are reported as single-run MAE/RMSE values with no error bars, no number of seeds, and no significance tests. For a benchmark paper this is often acceptable, but the paper uses strong language such as 'consistently' and 'state of the art.' If the abstract's 'statistically significant' language is meant to apply to the education study, it belongs to a study that is not present; if it is meant to apply to the ETA results, variance estimates or paired comparisons are needed to support it.","section":"Tables 1–4"}],"minor_comments":[{"comment":"Typos: 'intractible' (Sec. 3.1), 'frquency' (Supp. A), and awkward phrasing such as 'we again improve by and 58.2% and 27.13% over TEA' (Sec. 4.2).","section":"Throughout"},{"comment":"The statement that the energy model trained on one depth-completion model is not applicable to other models is important. Since Tables 1–4 report results for four architectures, please clarify whether each architecture requires its own energy model and whether this is factored into the comparison with baselines.","section":"Section 3.1"},{"comment":"The caption text '8 ×38' appears to contain a typographical spacing error; the intended region sizes should be stated precisely.","section":"Figure 4"}],"recommendation":"reject","confidential_remarks":"The submitted PDF appears to be a different paper from the one described in the title and abstract. This is not a normal revision issue: the education study promised in the abstract is entirely absent, and the body is a separate computer-vision manuscript. I recommend returning the submission to the authors so they can upload the correct manuscript; the ETA paper, if properly formatted and with Eq. (7) corrected, might be suitable for a computer-vision venue."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The submission has a major identity problem: the abstract and title describe a CS-education intervention ('Learning by Teaching' with a system called Socrates) and claim a statistically significant improvement over historical cohorts. The full text is a depth-completion test-time adaptation paper, ETA, with a different author list. There is no course, no Socrates, no cohort, no test statistic anywhere in the manuscript. So the paper's stated headline result cannot be checked. That is a structural failure, not a subtle one.\n\nThe actual technical content is a legitimate piece of work. ETA takes energy-based test-time adaptation, which has been applied to classification and segmentation, and extends it to dense depth regression. The contributions over TEA and ProxyTTA are real: a patch-level energy model that localizes errors, and the use of FGSM adversarial perturbations to synthesize OOD examples for training the energy model. The experiments cover six datasets and four architectures, including an outdoor-to-indoor shift, and ETA consistently beats the prior state of the art by a non-trivial margin. The method section is clear and the Limitations paragraph honestly flags that adversarial perturbations may not capture extreme or semantic target shifts.\n\nThe soft spots, in proportion. First, the code isn't shipped and key hyperparameters are deferred to 'released code', so Tables 1–4 cannot be reproduced now. Second, the self-referential nature of the energy model is a real concern: since the OOD examples are generated by perturbing the source model's own inputs, the critic is trained on that model's failure modes. The authors acknowledge it, but it remains the load-bearing assumption; if real deployment shifts are different in kind, the guidance could be misleading. Third, the benchmark tables are single runs with no variance, so we don't know the stability of the gains. These are all standard TTA-paper weaknesses; none of them makes the ETA method obviously wrong.\n\nWho gets value: a reader working on test-time adaptation or depth completion will find ETA worth reading once the code is available. The education claim, as submitted, has no evidentiary support at all. My recommendation to you: desk reject this submission because the abstract and the body are different papers, but the ETA part should be allowed to resubmit under its own title, with code and hyperparameters, and then it merits a serious referee.","headline":"The abstract promises an education study that isn't in the paper; the full text is a competent depth-completion TTA paper, and only the latter deserves a look.","tokens_in":19604,"tokens_out":4833,"would_cite":false,"duration_ms":49612,"reading_group":"maybe","serious_thinker":"unclear","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proposes ETA, a test-time adaptation method that minimizes a learned energy score to align depth predictions with the source distribution.","keywords":["depth completion","test-time adaptation","energy-based model","domain shift","adversarial perturbations","sparse depth","autonomous driving","dense regression"],"falsifier":"Take a pretrained depth completion model and a target shift that is not adversarial in character — for example, adding defocus blur to RGB inputs or changing the LiDAR beam pattern — then run ETA and check whether the energy-minimized predictions reduce error relative to the unadapted model. If the energy model assigns low energy to severely wrong predictions on such shifts, the adversarial-proxy premise fails.","tokens_in":18721,"feed_emoji":"🎯","tokens_out":6125,"duration_ms":63315,"temperature":0.7,"pith_summary":"This paper presents ETA, a method for adapting pretrained depth completion models to new environments at test time, without ground truth or multiple passes over target data. The central claim is that a patch-based energy model, trained on source data plus adversarially perturbed examples, can score how source-like each region of a predicted depth map is; minimizing that energy while updating only a small adaptation layer brings test-time predictions back in line with the training distribution. Across three indoor and three outdoor datasets, ETA consistently lowers MAE and RMSE relative to prior test-time adaptation baselines. The abstract of this submission describes a separate study in which students teach LLMs; the full text is the depth-completion work described here.","feed_headline":"One energy score lets depth models adapt to new scenes at test time","feed_subtitle":"A patch-level score tells the network which depth predictions look out-of-distribution, so it corrects only those regions.","key_machinery":"The central object is the energy-based model $E_\\phi$: a convolutional network that outputs a patch-wise energy map $e$ for a predicted dense depth map, conditioned on the sparse depth map. Energy is linked to prediction error through the target $y = 1 - \\exp(-\\Delta/\\tau)$, where $\\Delta$ is patch MSE; low energy corresponds to source-like, high energy to out-of-distribution. Out-of-distribution training examples are synthesized with FGSM adversarial perturbations of the RGB image and sparse depth. Test-time adaptation updates only the inserted adaptation layer $m_\\psi$ (and BatchNorm statistics) to minimize $L_{\\mathrm{adapt}} = w_e \\ell_{\\mathrm{energy}} + w_z \\ell_{\\mathrm{sparse}} + w_s","core_discovery":"The paper claims that depth completion models, when transferred to a target domain with covariate shift, can be adapted on the fly by treating the source distribution as an energy landscape. An energy model $E_\\phi$ maps a predicted depth map, conditioned on the sparse depth input, to a low-resolution map of per-patch energy, where low energy means 'likely under the source distribution'. The energy model is trained by assigning high target energy to predictions from adversarially perturbed inputs (FGSM), whose errors stand in for out-of-distribution shifts. At test time, only a lightweight adaptation layer and BatchNorm statistics are updated to minimize a weighted combination of energy, spa","pith_inferences":["The abstract and full text describe different papers; any claim about the education study (statistically significant improvements in student performance) is not supported by the content of this submission.","The weakest conceptual link — flagged by the paper's own Limitations section — is the premise that FGSM perturbations produce out-of-distribution examples representative of real deployment shifts; if real shifts are non-adversarial, the energy model may encode the source model's failure modes rather than the target environment.","A testable extension: evaluate ETA on non-adversarial corruptions (defocus blur, sensor noise, LiDAR density changes) and check whether energy maps still predict per-patch error; this would tell whether the adversarial proxy is load-bearing.","Since the energy model is trained conditioned on a specific depth network's parameters, a practical follow-up would be to train an architecture-agnostic energy model to enable cross-model deployment."],"forward_implications":["Depth completion models can be specialized to a new environment in a single forward-backward pass per batch, without retraining or target ground truth.","Region-level energy localizes corrections, so adaptation can fix specific areas of the depth map instead of applying uniform global updates.","A single energy model generalizes across multiple target domains because adversarial perturbations cover a broad range of out-of-distribution variations.","The same energy-minimization principle extends to other dense regression tasks such as optical flow, surface normal prediction, and medical image segmentation.","Because only a small adaptation layer is updated, the risk of catastrophic forgetting during test-time optimization is reduced."],"supporting_citations":[{"why":"Provides the pretrained depth completion models and the ProxyTTA baseline that ETA is built on and compared against.","marker":"[52]"},{"why":"Closest existing energy-based test-time adaptation method, originally for classification; ETA extends it to dense regression and localizes corrections.","marker":"[100]"},{"why":"Introduces FGSM, the adversarial perturbation method ETA uses to synthesize out-of-distribution training examples.","marker":"[21]"},{"why":"Baseline TENT-style batch-normalization adaptation, adapted here to depth completion.","marker":"[80]"},{"why":"CoTTA baseline for continual test-time adaptation used in comparisons.","marker":"[81]"},{"why":"VOID source dataset for indoor pretraining and adaptation evaluation.","marker":"[86]"},{"why":"KITTI source dataset for outdoor pretraining and adaptation evaluation.","marker":"[19]"},{"why":"MSG-CHN architecture used as one of the evaluated depth completion models.","marker":"[33]"},{"why":"NLSPN architecture used as another evaluated depth completion model.","marker":"[53]"},{"why":"BP-Net architecture evaluated in the large-shift outdoor-to-indoor experiments.","marker":"[74]"}],"fun_headline_variants":["Students who teach LLMs learn more, study finds","Inverted tutoring: students instruct AI to master CS","Teaching AI deepens student learning in CS courses","Student-led LLM instruction boosts exam scores","Active learning via teaching LLMs beats passive use"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"Adversarial perturbations of source inputs are a faithful stand-in for the out-of-distribution data a depth model will meet at deployment, so an energy model trained on those perturbations will learn a useful notion of 'source-like' predictions.","fun_headline_variants_meta":{"raw":{"variants":["Students who teach LLMs learn more, study finds","Inverted tutoring: students instruct AI to master CS","Teaching AI deepens student learning in CS courses","Student-led LLM instruction boosts exam scores","Active learning via teaching LLMs beats passive use"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000194,"raw_usage":{"total_tokens":1142,"prompt_tokens":646,"completion_tokens":496,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":390,"completion_tokens_details":{"reasoning_tokens":424}},"tokens_in":390,"tokens_out":496,"duration_ms":5435,"temperature":1.0,"reasoning_tokens":424,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T23:01:52.616011+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a pretrained depth completion model and a target shift that is not adversarial in character — for example, adding defocus blur to RGB inputs or changing the LiDAR beam pattern — then run ETA and check whether the energy-minimized predictions reduce error relative to the unadapted model. If the energy model assigns low energy to severely wrong predictions on such shifts, the adversarial-proxy premise fails.","supporting_citations":[{"cited_title":"Test- time adaptation for depth completion","cited_arxiv_id":null,"evidence_quote":"Provides the pretrained depth completion models and the ProxyTTA baseline that ETA is built on and compared against."},{"cited_title":"Tea: Test-time energy adaptation","cited_arxiv_id":null,"evidence_quote":"Closest existing energy-based test-time adaptation method, originally for classification; ETA extends it to dense regression and localizes corrections."},{"cited_title":"Tent: Fully test-time adapta- tion by entropy minimization","cited_arxiv_id":null,"evidence_quote":"Baseline TENT-style batch-normalization adaptation, adapted here to depth completion."},{"cited_title":"Continual test-time domain adaptation","cited_arxiv_id":null,"evidence_quote":"CoTTA baseline for continual test-time adaptation used in comparisons."},{"cited_title":"Unsupervised depth completion from visual iner- tial odometry","cited_arxiv_id":null,"evidence_quote":"VOID source dataset for indoor pretraining and adaptation evaluation."},{"cited_title":"A multi-scale guided cascade hourglass network for depth completion","cited_arxiv_id":null,"evidence_quote":"MSG-CHN architecture used as one of the evaluated depth completion models."},{"cited_title":"Non-local spatial propagation network for depth completion","cited_arxiv_id":null,"evidence_quote":"NLSPN architecture used as another evaluated depth completion model."},{"cited_title":"Bilateral propagation network for depth completion","cited_arxiv_id":null,"evidence_quote":"BP-Net architecture evaluated in the large-shift outdoor-to-indoor experiments."}],"review_version":1}