{"id":"a8978a9d-a91a-4a2d-a9c3-c762d9a6500f","arxiv_id":"2412.06160","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"GP-ND adds a log-KL divergence penalty between a GP's predictive distribution and Gaussian blobs placed on negative data pairs, aiming to fit positive points while avoiding obstacles.","lead":"This paper proposes GP-ND, a Gaussian process regression method that treats certain input-target pairs as obstacles, modeled as Gaussian blobs, and adds a KL-divergence penalty to push the GP's predictions away from those points. The authors report improved test log-likelihood and faster convergence on synthetic and six real datasets, but the derivation of the core penalty is ambiguous and the real-data experiments use randomly shuffled labels as negatives.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The KL objective in Eq. (4) is written for joint distributions, but Eq. (6) is the univariate scalar KL; the paper never specifies whether it sums over negative points or uses the full covariance, so the objective actually optimized in Algorithm 1 is undefined.","rationale":"The paper's core proposal is the KL-divergence penalty in Eq. (5), so for the central claim to hold, the objective being optimized must be well-defined and must match the derivation. This is precisely where the paper fails: Eq. (4) writes a joint KL, Eq. (6) implements an univariate scalar KL, and Algorithm 1 does not say which one is used or how the per-point blobs are combined. This is not a minor notation issue, because the gradient of a sum of marginal KLs differs from the gradient of the joint KL when the predictive covariance couples the negative points. Without knowing which version was optimized, the experiments cannot be interpreted as validating the proposed mechanism. The random-shuffle construction of negative datapairs in §5.2 further limits the obstacle-avoidance interpretation, but the undefined objective is the more fundamental barrier: even the mechanism is unclear. I therefore agree with the reader's weakest assumption and with the REJECT verdict. The paper could be improved by specifying a sum over marginals (or using the full multivariate KL) and updating Eq. (4) and Algorithm 1 accordingly; the analytical check above would settle the inconsistency.","tokens_in":14534,"tokens_out":8311,"duration_ms":82160,"concrete_test":"Re-derive the gradient of the DKL term in Eq. (5) two ways: (i) using the multivariate KL between the joint predictive N(µ1, Σ1) at all negative inputs and the product of blob Gaussians N(µ2, Σ2), and (ii) using the sum of m univariate KLs of the form Eq. (6). Compare the resulting gradients; if they differ—specifically because the multivariate form includes tr(Σ2^{-1}Σ1) and log(|Σ2|/|Σ1|) terms absent from the univariate sum—then Eq. (4) and Eq. (6) are inconsistent, and the objective actually optimized by Algorithm 1 is not the one derived in §3.3.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.3 defines the penalty in Eq. (4) as DKL(p(y|θ, ¯X) ∥ q(¯y| ¯X)), where p(y|θ, ¯X) is the GP predictive distribution over the negative inputs. From Eq. (1), that predictive is a multivariate Gaussian with a full covariance matrix, not a scalar. The blob distribution q is defined per datapoint in §3.2. However, Eq. (6) is the univariate KL formula: log(σ2/σ1) + (σ1² + (µ1−µ2)²)/(2σ2²) − 1/2, which contains no trace or log-determinant terms and therefore cannot be the KL between the joint multivariate predictive and the product of blobs. The text then states p ≡ N(µ1, σ1) ≡ p(y|θ, ¯X), collapsing the multivariate predictive to a single variance. Algorithm 1 computes KLdiv = β·log DKL(p( ˆy|θ, ¯X) || N(¯y, σ²_neg)) without indicating whether this is a sum over m negative points, a single marginal, or a joint KL. Thus the training objective is ambiguous: if it is a sum of univariate KLs, Eq. (4) is mis-specified; if it is a joint KL, Eq. (6) is wrong. The empirical results cannot be attributed to the stated Eq. (5) unless this ambiguity is resolved, because the gradient of a sum of marginals differs from the gradient of the joint KL when the predictive covariance couples the negative points.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GP-ND, an extension of Gaussian process regression in which, in addition to ordinary positive training pairs, the user supplies negative pairs (inputs with target values that the fitted curve must avoid). The negative pairs are represented as small Gaussian blobs, and the training objective combines the standard GP log marginal likelihood with a log-KL penalty that pushes the GP predictive distribution away from the blobs. The method is presented as a generic wrapper that can be applied to exact GP, SVGP, and PPGPR. Experiments include a synthetic 2D trajectory prediction task and six real-world regression datasets, reporting improved NLL and RMSE, faster convergence, and modest runtime overhead relative to classical GP regression.","tokens_in":14844,"tokens_out":8686,"duration_ms":83851,"significance":"If made mathematically coherent, the core idea is simple and potentially useful: a plug-in penalty that lets a GP user encode 'avoid this output value at this input' without enlarging the covariance matrix. The paper also deserves credit for being explicit about the alternating update scheme and for framing the extension as applicable to several scalable GP models. However, the central derivation currently has a gap between the stated multivariate KL objective and the implemented univariate formula, and the real-data experiments rely on randomly shuffled labels rather than genuine obstacles, so the significance of the claims cannot yet be assessed from the manuscript.","major_comments":[{"comment":"The objective actually optimized is not well defined. Equation (4) denotes D_KL(p(y|θ, bar X) || q(bar y| bar X)), where p(y|θ, bar X) is the multivariate GP predictive distribution over the m negative inputs and therefore has a full covariance matrix. Equation (6), however, is the KL divergence between two univariate Gaussians, and the text immediately identifies p ≡ N(μ1, σ1) ≡ p(y|θ, bar X), collapsing the joint distribution to a single variance. The paper never states whether Algorithm 1 sums m univariate marginal KL divergences, uses a single marginal, or computes the joint KL with trace and log-determinant terms. These are different objectives with different gradients once the predictive covariance couples the negative points, so the experimental results cannot be attributed to the derivation in Eqs. (4)–(5). Please restate the objective consistently, for example as a sum of per-point KL terms, and specify exactly what Algorithm 1 computes.","section":"§3.3–§3.5 and Algorithm 1"},{"comment":"The real-data experiments do not evaluate obstacle avoidance. Negative pairs are produced by randomly shuffling labels and pairing each input with a mismatched target; these are not obstacles in any task-defined sense, and the condition y(x)_True ≠ y(x)_Shuffled merely restates that the labels were shuffled. Separating the predictive distribution from random wrong targets is a form of regularization, not avoidance of a specified forbidden region, so the claim that GP-ND 'outperforms' classical GP on these datasets does not support the paper's central obstacle-avoidance setting. Either use datasets with genuine negative constraints, such as forbidden output regions or known unsafe targets, or explicitly reframe GP-ND as a regularizer and evaluate it on that basis.","section":"§5.2, Random shuffling technique"},{"comment":"There is a mismatch between the stated objective and the training procedure. Equation (5) defines a joint minimization over the sum of the negative log likelihood and the β-weighted negative log-KL term, while Algorithm 1 alternates between minimizing the NLL and maximizing the KL divergence. The text states that alternating updates were found empirically more stable, but no comparison with joint optimization is reported. Since all reported results are produced by Algorithm 1, the paper must clarify whether the behavior reflects the joint objective or the alternating scheme, and should justify the alternation either theoretically or with an ablation.","section":"Algorithm 1 and Eq. (5)"},{"comment":"The headline empirical claims are not quantitatively supported. The trajectory comparison against NS GP in §5.1 is visual only, with no NLL or RMSE values reported for the three models. The real-world results in Fig. 4 are averages over 10 random splits without error bars or significance tests, and the 'faster convergence' claim in Fig. 5 is based on qualitative inspection of convergence curves. A table with means, standard deviations, and a significance test for the key comparisons, together with a quantitative convergence-rate statistic such as epochs to a fixed RMSE threshold, is needed to support the 'outperforms' and 'converges faster' claims.","section":"§5.1 and Fig. 5"}],"minor_comments":[{"comment":"Equation (3) appears to be missing the factor −1/2 on the quadratic term; as written it is not the standard log marginal likelihood, and the sign convention is inconsistent with the '− log p(y|θ, X)' used in Eq. (5).","section":"§2, Eq. (3)"},{"comment":"In Algorithm 1, 'NLL = - p(y|θ, X)' should read '-log p(y|θ, X)', and the KLdiv line should state explicitly whether β·log D_KL is summed over negative points or computed jointly; otherwise the pseudocode inherits the ambiguity raised in the first major comment.","section":"Algorithm 1"},{"comment":"The stated motivation for applying the logarithm to D_KL is not convincing as written: log(D_KL) is also unbounded above and tends to −∞ as D_KL → 0, so it does not 'fix' unboundedness. If the purpose is to change the gradient scale, the intended effect should be stated explicitly.","section":"§3.4–§3.5"},{"comment":"Section 3.6.2 defers the PPGPR likelihood to the original paper and provides only the predictive equations, but PPGPR-ND is then evaluated in §5; the manuscript should at least give the exact loss being optimized for PPGPR-ND.","section":"§3.6.2"},{"comment":"References [18] and [19] are duplicate entries for GPyTorch, and several references in Section 6 are not discussed in the body; the list should be trimmed to works actually used.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The main unresolved risk is the Eq. (4)–(6) objective ambiguity; if a revised version cannot state one coherent objective and show that Algorithm 1 implements it, I would not be able to recommend acceptance even after new experiments. I also note that the reference list contains a large number of self-citations that are disproportionate to the technical content and should be trimmed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere's my take on arXiv:2412.06160. The paper has a simple and attractive idea—define negative datapairs as Gaussian blobs and add a log-KL penalty that pushes the GP predictive away from them while fitting positive pairs. That's a reasonable drop-in for exact and sparse GPs, and the toy trajectory demo is visually effective. The author also tries an NS GP baseline and notes that the covariance matrix stays the size of the positive set, which is a fair point.\n\nThe soft spot is the one the stress-test note flags, and it's load-bearing. Eq (4) writes the penalty as a KL between joint predictive distributions (multivariate Gaussians), but Eq (6) is the univariate KL formula, with no trace or log-determinant terms. The text sets σ1 = cov(f∗), which is a matrix, into a scalar formula. Algorithm 1 never says whether it sums over the m negative points or evaluates a joint KL. So the objective actually optimized is undefined, and the gradient of a sum of marginals is not the same as the gradient of the joint KL when covariance couples negative points. This is not a minor typo; it sits at the center of the method.\n\nThe real-data evaluation is also weak. Negative pairs are made by randomly shuffling labels—that's not obstacle modeling, it's label noise. The reported NLL gains are plausibly a regularization effect. There are no error bars, no code, and no comparison to constrained-GP literature beyond NS GP. The hyperparameter discussion is heuristic.\n\nSo the paper as submitted does not establish its claims. But the core idea—a KL-avoidance penalty with fixed blobs—is worth a serious look. A corrected derivation, a real obstacle task, and seed variance could make it a solid workshop-level paper. I'd send it to peer review rather than desk reject, because the flaws are diagnosable and the idea deserves air time. I wouldn't cite it in this form.","headline":"A promising KL-avoidance idea for GP regression, but the objective is undefined as written and the real-data evaluation uses shuffled labels, so the claims don't hold.","tokens_in":15414,"tokens_out":4199,"would_cite":false,"duration_ms":38185,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Obstacle-aware Gaussian process regression claims that marking forbidden targets as negative datapairs and maximizing their KL divergence from the GP lets a model fit positive data while avoiding obstacles, with better NLL and RMSE than…","keywords":["Gaussian process regression","negative datapairs","KL divergence","obstacle avoidance","trajectory prediction","scalable Gaussian processes","negative log-likelihood","regression constraints"],"falsifier":"At the converged parameters from Algorithm 1, compute both the full multivariate KL divergence between $\\mathcal{N}(E[f(\\bar{X})], \\mathrm{cov}(f(\\bar{X})))$ and the product of the negative-blob Gaussians, and the sum of the univariate KL terms that Eq. (6) implements. If the two values are not equal, or are not monotonically related, then the objective the code maximizes is not the objective the paper derives, and the reported gains are for a different loss.","tokens_in":14228,"feed_emoji":"🚫","tokens_out":10856,"duration_ms":97137,"temperature":0.7,"pith_summary":"Obstacle-aware Gaussian Process Regression proposes a drop-in modification, called GP-ND, that lets a Gaussian process honor two kinds of training data: positive datapairs it must fit and negative datapairs it must avoid. Negative datapairs are modeled as small Gaussian blobs centered at forbidden target values, and the GP is trained to maximize the KL divergence between its predictive distribution at those inputs and the blobs, jointly with the usual marginal likelihood. The paper argues this turns any exact or scalable GP regressor into an obstacle-aware one. On a synthetic trajectory task and six real-world datasets, GP-ND versions of exact GP, scalable variational GP, and parametric GP regression report lower negative log-likelihood and RMSE than their classical counterparts, converge faster, and add modest runtime overhead. If correct, the framework encodes 'stay away from this value here' constraints into GP learning without enlarging the covariance matrix.","feed_headline":"Negative datapairs teach Gaussian processes to avoid obstacles","feed_subtitle":"A KL-divergence penalty pushes GP predictions away from forbidden targets, improving NLL and RMSE on real data.","key_machinery":"The central object is the negative datapair as a Gaussian blob $q(\\bar{y}\\mid\\bar{x}) = \\mathcal{N}(\\bar{y}, \\sigma_{\\mathrm{neg}}^2)$ together with the log-KL penalty $\\beta \\log D_{\\mathrm{KL}}(p(y\\mid\\theta,\\bar{X}) \\lVert q(\\bar{y}\\mid\\bar{X}))$ appended to the GP marginal likelihood. The KL term is the avoidance mechanism: it grows when the GP's predictive distribution overlaps a blob, so optimizing it pushes the predictive mean away from $\\bar{y}$ and inflates predictive variance near $\\bar{x}$. The implementation uses the closed-form univariate Gaussian KL formula from Eq. (6), evaluated with the predictive mean and variance at the negative inputs, and updates kernel parameters by alternating between the negative-log-likelihood step and the KL step in Algorithm 1.","core_discovery":"The central claim is that optimizing the combined objective $L(\\theta) = \\arg\\min_\\theta [-\\log p(y\\mid\\theta,X) - \\beta \\log D_{\\mathrm{KL}}(p(y\\mid\\theta,\\bar{X})\\,\\lVert\\, q(\\bar{y}\\mid\\bar{X}))]$ makes a GP fit positive datapairs while avoiding negative ones. In this formulation a negative datapair $(\\bar{x}, \\bar{y})$ is not a training point but an obstacle: it is represented as a Gaussian blob $q(\\bar{y}\\mid\\bar{x}) = \\mathcal{N}(\\bar{y}, \\sigma_{\\mathrm{neg}}^2)$ whose radius is controlled by $\\sigma_{\\mathrm{neg}}$. The training objective minimizes the negative marginal likelihood on positive pairs and simultaneously maximizes a log-transformed KL divergence between the GP's predictive distribution at the negative inputs and the blobs, with strength $\\beta$. The paper reports that this objective, updated alternately in Algorithm 1, outperforms classical GP, scalable variational GP, and parametric GP regression on negative log-likelihood and RMSE across datasets up to roughly 435,000 points, and that lowering $\\beta$ or raising $\\sigma_{\\mathrm{neg}}$ weakens the avoidance effect.","pith_inferences":["Editorial extension: Eq. (4) is written for joint multivariate distributions while the implemented Eq. (6) is the univariate scalar-Gaussian KL, so the paper leaves unresolved whether the loss sums over the $m$ negative points or uses the full predictive covariance; the empirical results stand or fall with whichever the code actually computes.","Editorial extension: on the six benchmark datasets the negative datapairs are random shuffled labels rather than spatial obstacles, so the reported gains may reflect regularization from the KL term more than genuine obstacle avoidance; a $\\beta \\to 0$ ablation would separate the two.","Editorial extension: Algorithm 1 alternates the NLL and KL updates instead of jointly optimizing Eq. (5), so the faster convergence could come from the optimization schedule rather than the penalty itself; comparing joint and alternating optimization on the same objective would clarify this.","Editorial extension: because the KL penalty uses only predictive mean and variance at a negative input, the mechanism is a local repulsive potential; one could test it against other forbidden-region shapes, such as intervals or uniform shells, to see whether the Gaussian-blob form is essential."],"forward_implications":["For navigation tasks, obstacles can be encoded directly as negative datapairs, so the GP trajectory bends around them without requiring a dense cloud of positive points near each obstacle.","Because only the likelihood term changes, the same KL-avoidance penalty can be attached to exact GP, scalable variational GP, and parametric GP regression without changing the size of the covariance matrix.","The two hyperparameters $\\beta$ and $\\sigma_{\\mathrm{neg}}$ control the avoidance radius: smaller $\\beta$ or larger $\\sigma_{\\mathrm{neg}}$ reduces the penalty's influence, giving practitioners a calibrated safety margin.","On real-world regression benchmarks, adding pseudo-negative datapairs by shuffling labels improves NLL and RMSE and steepens the convergence curve as data size grows, suggesting the penalty also acts as a regularizer.","The runtime overhead of the extra KL term stays modest relative to the underlying GP method, so the obstacle-aware variant inherits the scalability of the base regressor."],"supporting_citations":[{"why":"Establishes the classical GP regression framework and predictive equations that GP-ND extends with negative datapairs.","marker":"[39]"},{"why":"Supplies the scalable variational GP model whose likelihood is plugged into the GP-ND objective as a baseline and extension target.","marker":"[25]"},{"why":"Supplies the parametric GP regressor used as another sparse-GP baseline and plug-in for the negative-constraint penalty.","marker":"[27]"},{"why":"Defines the leveraged non-stationary GP baseline that incorporates negative constraints through the kernel, the direct comparison for the trajectory experiments.","marker":"[8]"},{"why":"Provides the pedestrian-trajectory modeling setup that inspires the 2-D obstacle trajectory experiment.","marker":"[11]"},{"why":"Provides the GPU-accelerated GP implementation used to run all models and baselines.","marker":"[18]"},{"why":"Supplies five of the six real-world benchmark datasets used in the empirical comparisons.","marker":"[10]"}],"fun_headline_variants":["GP regression dodges obstacles via negative datapairs","Negative datapairs steer Gaussian processes around obstacles","Obstacle-aware GP: KL penalty pushes predictions away","GP-ND: Gaussian processes learn to avoid forbidden targets","Teaching GPs to avoid obstacles with negative datapairs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the KL divergence actually maximized in the implementation, computed with the univariate Gaussian formula at each negative point, is the same objective as the joint-distribution KL divergence written in Eq. (4), because the paper never specifies how the per-point penalties combine.","fun_headline_variants_meta":{"raw":{"variants":["GP regression dodges obstacles via negative datapairs","Negative datapairs steer Gaussian processes around obstacles","Obstacle-aware GP: KL penalty pushes predictions away","GP-ND: Gaussian processes learn to avoid forbidden targets","Teaching GPs to avoid obstacles with negative datapairs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000173,"raw_usage":{"total_tokens":1303,"prompt_tokens":996,"completion_tokens":307,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":612,"completion_tokens_details":{"reasoning_tokens":229}},"tokens_in":612,"tokens_out":307,"duration_ms":3218,"temperature":1.0,"reasoning_tokens":229,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T19:57:13.513704+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"At the converged parameters from Algorithm 1, compute both the full multivariate KL divergence between $\\mathcal{N}(E[f(\\bar{X})], \\mathrm{cov}(f(\\bar{X})))$ and the product of the negative-blob Gaussians, and the sum of the univariate KL terms that Eq. (6) implements. If the two values are not equal, or are not monotonically related, then the objective the code maximizes is not the objective the paper derives, and the reported gains are for a different loss.","supporting_citations":[{"cited_title":"Lawrence","cited_arxiv_id":null,"evidence_quote":"Supplies the scalable variational GP model whose likelihood is plugged into the GP-ND objective as a baseline and extension target."},{"cited_title":"Parametric gaussian process regressors","cited_arxiv_id":null,"evidence_quote":"Supplies the parametric GP regressor used as another sparse-GP baseline and plug-in for the negative-constraint penalty."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the leveraged non-stationary GP baseline that incorporates negative constraints through the kernel, the direct comparison for the trajectory experiments."},{"cited_title":"Mod- elling pedestrian trajectory patterns with gaussian processes","cited_arxiv_id":null,"evidence_quote":"Provides the pedestrian-trajectory modeling setup that inspires the 2-D obstacle trajectory experiment."},{"cited_title":"Gaurav Shrivastava Gpytorch: Blackbox matrix-matrix gaussian pro- cess inference with gpu acceleration","cited_arxiv_id":null,"evidence_quote":"Provides the GPU-accelerated GP implementation used to run all models and baselines."},{"cited_title":"UCI machine learn- ing repository, 2017","cited_arxiv_id":null,"evidence_quote":"Supplies five of the six real-world benchmark datasets used in the empirical comparisons."}],"review_version":1}