nielsr HF Staff commited on
Commit
4def648
·
verified ·
1 Parent(s): 7149984

Update papers and claims (part 2)

Browse files
site/p/2110.01950.json CHANGED
@@ -1 +1 @@
1
- {"id":"2110.01950","title":"Classification of high-dimensional data with spiked covariance matrix structure","authors":"Yin-Jen Chen; Minh Tang","openreview_id":"","presentation":"poster","track":"journal:TMLR","verdict":"borderline","background":"Fisher's linear discriminant rule is Bayes-optimal for two Gaussian classes with a shared covariance, but its plug-in version can do no better than random guessing when the number of features p is comparable to or larger than the sample size n. Most high-dimensional fixes assume the discriminant direction β = Σ⁻¹(µ2 − µ1) is sparse and that Σ has bounded eigenvalues. Real high-dimensional data such as gene expression often have a few large (spiked) eigenvalues instead. The paper studies the common practice of reducing dimension with PCA before running LDA, and gives conditions under which this is asymptotically Bayes-optimal.","method":["lda∘pca (Algorithm 1) estimates the whitening matrix W = Σ^{-1/2} under a spiked model Σ = UΛUᵀ + σ²I. The top d eigenpairs (Û, Λ̂) of the pooled sample covariance are kept, σ̂² is set to (tr Σ̂ − tr Λ̂)/(p − d), and Ŵ = Û(Λ̂ + σ̂²I)^{-1/2}Ûᵀ + σ̂⁻¹(I − ÛÛᵀ). The method whitens the class means, keeps the s largest coordinates in absolute value of ζ̂ = Ŵ(X̄2 − X̄1), and applies the Fisher rule with threshold ln(n1/n2) on those whitened coordinates. d is the smallest k for which the top-k eigenvalues explain at least 90% of tr(Σ̂). s is chosen from {1,…,30} by five-fold CV, taking the smallest s on ties.","In simulation Model 1 (equal correlation), p = 800, Σ = ρ11ᵀ + (1 − ρ)I, µ1 = 0 and µ2 = (1₁₀, 0₇₉₀). Each class has 100 training and 100 test points. Every result is averaged over 200 Monte Carlo replicates. Baselines are CAT-LDA (R sda), DSDA (R TULIP), AdaLDA (ADAM Matlab code), DSDA and AdaLDA after the rotation preprocessing of Hao et al. (γ = 0.25), SLDA and NSC. All baseline tuning parameters are chosen by five-fold CV. The oracle is the Fisher rule with the true parameters, i.e. the Bayes error."],"claim":"In the equal-correlation simulation (p = 800, 100 training points per class, ρ = 0.5), lda∘pca has a test mis-classification rate of 1.74% (sd 1.00) against 3.27% (sd 1.47) for DSDA, the best non-oracle baseline. The Bayes error is 1.37%. lda∘pca uses about 12 whitened features, while DSDA uses about 96 raw features.","source":"Table 1, page 13 (feature counts from Table 2, page 14)","metric":"Test mis-classification rate (%), lower is better. Mean over 200 Monte Carlo replicates; the value in parentheses is the standard deviation across replicates, not a standard error.","comparison":"lda∘pca vs. DSDA (strongest baseline at ρ = 0.5), AdaLDA, CAT-LDA, DSDA(rot), AdaLDA(rot), SLDA, NSC, and the oracle Bayes rule","falsification":"Over 200 replicates of Model 1 with ρ = 0.5, the mean error of lda∘pca is not below DSDA's (or AdaLDA's), or the gap is smaller than one per-replicate standard deviation (about 1.0–1.5 points).","caption":"Table 1: Mis-classification rate (%) with standard deviations (%) in parentheses for equal correlation setting (model 1), based on 200 independent Monte Carlo replicates.","image":"img/2110.01950.png","kind":"table","image_w":1322,"image_h":466,"pwc":false}
 
1
+ {"id":"2110.01950","title":"Classification of high-dimensional data with spiked covariance matrix structure","authors":"Yin-Jen Chen; Minh Tang","openreview_id":"","presentation":"poster","track":"journal:TMLR","verdict":"borderline","background":"Fisher's linear discriminant rule is Bayes-optimal for two Gaussian classes with a shared covariance, but its plug-in version can do no better than random guessing when the number of features p is comparable to or larger than the sample size n. Most high-dimensional fixes assume the discriminant direction β = Σ⁻¹(µ2 − µ1) is sparse and that Σ has bounded eigenvalues. Real high-dimensional data such as gene expression often have a few large (spiked) eigenvalues instead. The paper studies the common practice of reducing dimension with PCA before running LDA, and gives conditions under which this is asymptotically Bayes-optimal.","method":["lda∘pca (Algorithm 1) estimates the whitening matrix W = Σ^{-1/2} under a spiked model Σ = UΛUᵀ + σ²I. The top d eigenpairs (Û, Λ̂) of the pooled sample covariance are kept, σ̂² is set to (tr Σ̂ − tr Λ̂)/(p − d), and Ŵ = Û(Λ̂ + σ̂²I)^{-1/2}Ûᵀ + σ̂⁻¹(I − ÛÛᵀ). The method whitens the class means, keeps the s largest coordinates in absolute value of ζ̂ = Ŵ(X̄2 − X̄1), and applies the Fisher rule with threshold ln(n1/n2) on those whitened coordinates. d is the smallest k for which the top-k eigenvalues explain at least 90% of tr(Σ̂). s is chosen from {1,…,30} by five-fold CV, taking the smallest s on ties.","In simulation Model 1 (equal correlation), p = 800, Σ = ρ11ᵀ + (1 − ρ)I, µ1 = 0 and µ2 = (1₁₀, 0₇₉₀). Each class has 100 training and 100 test points. Every result is averaged over 200 Monte Carlo replicates. Baselines are CAT-LDA (R sda), DSDA (R TULIP), AdaLDA (ADAM Matlab code), DSDA and AdaLDA after the rotation preprocessing of Hao et al. (γ = 0.25), SLDA and NSC. All baseline tuning parameters are chosen by five-fold CV. The oracle is the Fisher rule with the true parameters, i.e. the Bayes error."],"claim":"In the equal-correlation simulation (p = 800, 100 training points per class, ρ = 0.5), lda∘pca has a test mis-classification rate of 1.74% (sd 1.00) against 3.27% (sd 1.47) for DSDA, the best non-oracle baseline. The Bayes error is 1.37%. lda∘pca uses about 12 whitened features, while DSDA uses about 96 raw features.","source":"Table 1, page 13 (feature counts from Table 2, page 14)","metric":"Test mis-classification rate (%), lower is better. Mean over 200 Monte Carlo replicates; the value in parentheses is the standard deviation across replicates, not a standard error.","comparison":"lda∘pca vs. DSDA (strongest baseline at ρ = 0.5), AdaLDA, CAT-LDA, DSDA(rot), AdaLDA(rot), SLDA, NSC, and the oracle Bayes rule","falsification":"Over 200 replicates of Model 1 with ρ = 0.5, the mean error of lda∘pca is not below DSDA's (or AdaLDA's), or the gap is smaller than one per-replicate standard deviation (about 1.0–1.5 points).","caption":"Table 1: Mis-classification rate (%) with standard deviations (%) in parentheses for equal correlation setting (model 1), based on 200 independent Monte Carlo replicates.","image":"img/2110.01950.png","kind":"table","image_w":1322,"image_h":466,"pwc":true}
site/p/2412.09498.json CHANGED
@@ -1 +1 @@
1
- {"id":"2412.09498","title":"Gradient descent inference in empirical risk minimization","authors":"Qiyang Han; Xiaocong Xu","openreview_id":"","presentation":"poster","track":"journal:AoS","verdict":"borderline","background":"Gradient descent is the standard way to fit empirical risk minimization models, but when the sample size m and the dimension n are proportional (the mean-field regime), the gradient descent iterates are biased and depend on one another in ways classical theory does not describe. Valid confidence intervals for individual coordinates of the unknown signal are usually built from the converged estimator, and only for convex losses. A way to do valid inference at any iteration of gradient descent, with convex or non-convex losses and without waiting for convergence, would make early-stopped or non-convex fits usable for statistical inference.","method":["The authors derive a non-asymptotic state evolution for the gradient descent iterates mu^(t) in models Y_i = F(<A_i, mu*>, xi_i). The theory says that the debiased iterate mu_db^(t) = mu^(t-1) + eta * sum_{s in [1:t]} omega_{t,s} A^T d1L(A mu^(s-1), Y), with omega^[t] = (tau^[t])^{-1} the inverse of an 'Onsager correction' matrix (Eq. 1.9), is approximately b_db^(t) mu* + sigma_db^(t) Z coordinatewise, with Z standard Gaussian; in linear regression b_db^(t) = 1, so mu_db^(t) is approximately mu* + sigma_db^(t) Z. Algorithm 1 (the 'gradient descent inference algorithm') estimates the Onsager matrices tau^[t] and rho^[t] from the data alone, which gives mu_db^(t), a variance estimate sigma_db^(t) (Eq. 3.2), and the coordinatewise 95% confidence intervals mu_db;l^(t) +/- sigma_db^(t) z_{0.025} (Eq. 4.4 for linear regression).","Simulation setting (Section 6.1-6.2): m = 1200, n = 1000; mu* has i.i.d. |N(0,5)| entries; sqrt(n)A has i.i.d. N(0,1), standardized t(10), or standardized Bernoulli(1/2) entries; mu^(0) ~ N(0, I_n); 50 gradient descent iterations with step size eta = 0.3; B = 1000 Monte Carlo repetitions. Two linear regression settings: squared loss with N(0,1) noise, and pseudo-Huber loss (delta = 1) with t(2) noise. No regularization. Coverage is the fraction of the n coordinates whose CI contains mu*_l, shown as a box plot over repetitions at iterations 10, 20, 30, 40, 50."],"claim":"In linear regression with m = 1200, n = 1000 and no regularization, the 95% CIs built from the debiased gradient descent iterate via Algorithm 1 have median coordinatewise coverage of about 0.95 (repetition boxes roughly 0.94-0.96) at every iteration t = 10, 20, 30, 40, 50, for both squared loss and pseudo-Huber loss with t(2) noise, under Gaussian, Bernoulli and t(10) designs.","source":"Figure 1 (middle column), page 28","metric":"Empirical coverage of nominal 95% CIs averaged over the n = 1000 coordinates (target 0.95); box plots show the spread over B = 1000 Monte Carlo repetitions. No other error bars.","comparison":"Empirical coverage of the debiased-GD CIs vs. the nominal 95% level, across iterations, two losses and three design distributions. There is no competing inference method in the figure.","falsification":"Following the paper's setup, the median coverage at some iteration in {10, ..., 50} falls clearly outside about [0.93, 0.97] (e.g. below 0.92) for either loss or any of the three designs, or the standardized debiased coordinate (mu_db;1 - mu*_1)/sigma_db departs clearly from N(0,1) in a QQ plot.","caption":"Figure 1. Linear regression. Top row: Squared loss. Bottom row: Pseudo-Huber loss.","image":"img/2412.09498.png","kind":"figure","image_w":925,"image_h":616,"pwc":false}
 
1
+ {"id":"2412.09498","title":"Gradient descent inference in empirical risk minimization","authors":"Qiyang Han; Xiaocong Xu","openreview_id":"","presentation":"poster","track":"journal:AoS","verdict":"borderline","background":"Gradient descent is the standard way to fit empirical risk minimization models, but when the sample size m and the dimension n are proportional (the mean-field regime), the gradient descent iterates are biased and depend on one another in ways classical theory does not describe. Valid confidence intervals for individual coordinates of the unknown signal are usually built from the converged estimator, and only for convex losses. A way to do valid inference at any iteration of gradient descent, with convex or non-convex losses and without waiting for convergence, would make early-stopped or non-convex fits usable for statistical inference.","method":["The authors derive a non-asymptotic state evolution for the gradient descent iterates mu^(t) in models Y_i = F(<A_i, mu*>, xi_i). The theory says that the debiased iterate mu_db^(t) = mu^(t-1) + eta * sum_{s in [1:t]} omega_{t,s} A^T d1L(A mu^(s-1), Y), with omega^[t] = (tau^[t])^{-1} the inverse of an 'Onsager correction' matrix (Eq. 1.9), is approximately b_db^(t) mu* + sigma_db^(t) Z coordinatewise, with Z standard Gaussian; in linear regression b_db^(t) = 1, so mu_db^(t) is approximately mu* + sigma_db^(t) Z. Algorithm 1 (the 'gradient descent inference algorithm') estimates the Onsager matrices tau^[t] and rho^[t] from the data alone, which gives mu_db^(t), a variance estimate sigma_db^(t) (Eq. 3.2), and the coordinatewise 95% confidence intervals mu_db;l^(t) +/- sigma_db^(t) z_{0.025} (Eq. 4.4 for linear regression).","Simulation setting (Section 6.1-6.2): m = 1200, n = 1000; mu* has i.i.d. |N(0,5)| entries; sqrt(n)A has i.i.d. N(0,1), standardized t(10), or standardized Bernoulli(1/2) entries; mu^(0) ~ N(0, I_n); 50 gradient descent iterations with step size eta = 0.3; B = 1000 Monte Carlo repetitions. Two linear regression settings: squared loss with N(0,1) noise, and pseudo-Huber loss (delta = 1) with t(2) noise. No regularization. Coverage is the fraction of the n coordinates whose CI contains mu*_l, shown as a box plot over repetitions at iterations 10, 20, 30, 40, 50."],"claim":"In linear regression with m = 1200, n = 1000 and no regularization, the 95% CIs built from the debiased gradient descent iterate via Algorithm 1 have median coordinatewise coverage of about 0.95 (repetition boxes roughly 0.94-0.96) at every iteration t = 10, 20, 30, 40, 50, for both squared loss and pseudo-Huber loss with t(2) noise, under Gaussian, Bernoulli and t(10) designs.","source":"Figure 1 (middle column), page 28","metric":"Empirical coverage of nominal 95% CIs averaged over the n = 1000 coordinates (target 0.95); box plots show the spread over B = 1000 Monte Carlo repetitions. No other error bars.","comparison":"Empirical coverage of the debiased-GD CIs vs. the nominal 95% level, across iterations, two losses and three design distributions. There is no competing inference method in the figure.","falsification":"Following the paper's setup, the median coverage at some iteration in {10, ..., 50} falls clearly outside about [0.93, 0.97] (e.g. below 0.92) for either loss or any of the three designs, or the standardized debiased coordinate (mu_db;1 - mu*_1)/sigma_db departs clearly from N(0,1) in a QQ plot.","caption":"Figure 1. Linear regression. Top row: Squared loss. Bottom row: Pseudo-Huber loss.","image":"img/2412.09498.png","kind":"figure","image_w":925,"image_h":616,"pwc":true}
site/p/2501.09345.json CHANGED
@@ -1 +1 @@
1
- {"id":"2501.09345","title":"Rational Tuning of LLM Cascades via Probabilistic Modeling","authors":"Michael Zellinger; Matt Thomson","openreview_id":"","presentation":"poster","track":"journal:TMLR","verdict":"include","background":"An LLM cascade sends each query to a small model first and passes it on to a larger, more expensive model only when the current model's confidence is below a threshold. The k-1 thresholds set the trade-off between error rate and inference cost. Tuning them is usually done by grid search or black-box Bayesian optimization, which gets expensive as cascades get longer and needs enough labeled data to estimate how the models' errors interact.","method":["Each model's raw confidence (max answer-choice probability for multiple choice, or P('Y') from a self-verification prompt for generation) is calibrated by logistic regression on a hyperparameter-free log transform. The method then fits a joint model of the calibrated confidences of k models. Each marginal is a mixture of two scaled beta distributions plus point masses at the min and max confidence. Each adjacent pair is linked by a Gumbel copula with θ = 1/(1-τ), where τ is Kendall's tau, and the joint is built with a Markov factorization (model i depends only on model i-1). This gives closed-form P(Correct) and E[Cost] for any threshold vector. The thresholds are tuned by minimizing (1 - P(Correct)) + λ E[Cost] with L-BFGS-B over a geometric sweep of λ.","Setup: all 26 cascades of length k ≥ 2 built from Llama 3.2 1B/3B and Llama 3.1 8B/70B/405B, run zero-shot on MMLU, MedMCQA, TriviaQA, XSum, GSM8K and TruthfulQA. Open-ended answers are judged by Claude 3.5 Sonnet. Baselines: Bayesian optimization (HEBO, GP surrogate, same objective and λ sweep) and grid search over 2.5%-quantile grids followed by a Pareto filter. In the low-sample setting, calibrators and models are fitted on a balanced subsample of 20-30 training examples (pairs of one correct and one incorrect example per model) and evaluated on the ~1000-example test sets. The metric is the area under the test error vs. normalized-cost curve (AUC, lower is better)."],"claim":"With n ≤ 30 training examples, Rational Tuning's average test error-cost AUC over the 26 Llama cascades is 0.325, vs. 0.338 for Bayesian optimization (-4.29% on average). It beats BO on all 6 benchmarks, with paired Wilcoxon p < 0.01 on 5 of them (e.g. MMLU -7.74%, p = 1e-6); TruthfulQA (-0.73%, p = 0.12) is not significant.","source":"Table 8, page 18","metric":"Area under the test-set error-rate vs. normalized-cost curve (lower is better); %Δ is the mean per-cascade percentage change in AUC; p is from a Wilcoxon test paired by cascade (26 cascades); no seed std reported","comparison":"Rational Tuning (Markov-copula model + continuous optimization) vs. Bayesian optimization (HEBO) and high-resolution grid search, all tuning the same cascades on the same low-sample training subsets","falsification":"With the same Llama cascades and a balanced 20-30-example training subsample, Rational Tuning's mean per-cascade AUC reduction vs. Bayesian optimization is near zero or positive on most of MMLU, MedMCQA, XSum and GSM8K, or is not significant under a paired Wilcoxon test.","caption":"Table 8: Area under the error-cost curve (AUC) in the low-sample limit (n ≤ 30 training examples), showing that our Rational Tuning (\"RT\") framework for selecting confidence thresholds consistently outperforms both a Bayesian optimization baseline (\"BO\") and high-resolution grid search (\"GS\"). The mean percentage changes (%Δ) are statistically significant at the p < 0.05 on almost all benchmarks, as measured by Wilcoxon rank-sum tests paired by cascade (highlighted in bold).","image":"img/2501.09345.png","kind":"table","image_w":1333,"image_h":561,"pwc":false}
 
1
+ {"id":"2501.09345","title":"Rational Tuning of LLM Cascades via Probabilistic Modeling","authors":"Michael Zellinger; Matt Thomson","openreview_id":"","presentation":"poster","track":"journal:TMLR","verdict":"include","background":"An LLM cascade sends each query to a small model first and passes it on to a larger, more expensive model only when the current model's confidence is below a threshold. The k-1 thresholds set the trade-off between error rate and inference cost. Tuning them is usually done by grid search or black-box Bayesian optimization, which gets expensive as cascades get longer and needs enough labeled data to estimate how the models' errors interact.","method":["Each model's raw confidence (max answer-choice probability for multiple choice, or P('Y') from a self-verification prompt for generation) is calibrated by logistic regression on a hyperparameter-free log transform. The method then fits a joint model of the calibrated confidences of k models. Each marginal is a mixture of two scaled beta distributions plus point masses at the min and max confidence. Each adjacent pair is linked by a Gumbel copula with θ = 1/(1-τ), where τ is Kendall's tau, and the joint is built with a Markov factorization (model i depends only on model i-1). This gives closed-form P(Correct) and E[Cost] for any threshold vector. The thresholds are tuned by minimizing (1 - P(Correct)) + λ E[Cost] with L-BFGS-B over a geometric sweep of λ.","Setup: all 26 cascades of length k ≥ 2 built from Llama 3.2 1B/3B and Llama 3.1 8B/70B/405B, run zero-shot on MMLU, MedMCQA, TriviaQA, XSum, GSM8K and TruthfulQA. Open-ended answers are judged by Claude 3.5 Sonnet. Baselines: Bayesian optimization (HEBO, GP surrogate, same objective and λ sweep) and grid search over 2.5%-quantile grids followed by a Pareto filter. In the low-sample setting, calibrators and models are fitted on a balanced subsample of 20-30 training examples (pairs of one correct and one incorrect example per model) and evaluated on the ~1000-example test sets. The metric is the area under the test error vs. normalized-cost curve (AUC, lower is better)."],"claim":"With n ≤ 30 training examples, Rational Tuning's average test error-cost AUC over the 26 Llama cascades is 0.325, vs. 0.338 for Bayesian optimization (-4.29% on average). It beats BO on all 6 benchmarks, with paired Wilcoxon p < 0.01 on 5 of them (e.g. MMLU -7.74%, p = 1e-6); TruthfulQA (-0.73%, p = 0.12) is not significant.","source":"Table 8, page 18","metric":"Area under the test-set error-rate vs. normalized-cost curve (lower is better); %Δ is the mean per-cascade percentage change in AUC; p is from a Wilcoxon test paired by cascade (26 cascades); no seed std reported","comparison":"Rational Tuning (Markov-copula model + continuous optimization) vs. Bayesian optimization (HEBO) and high-resolution grid search, all tuning the same cascades on the same low-sample training subsets","falsification":"With the same Llama cascades and a balanced 20-30-example training subsample, Rational Tuning's mean per-cascade AUC reduction vs. Bayesian optimization is near zero or positive on most of MMLU, MedMCQA, XSum and GSM8K, or is not significant under a paired Wilcoxon test.","caption":"Table 8: Area under the error-cost curve (AUC) in the low-sample limit (n ≤ 30 training examples), showing that our Rational Tuning (\"RT\") framework for selecting confidence thresholds consistently outperforms both a Bayesian optimization baseline (\"BO\") and high-resolution grid search (\"GS\"). The mean percentage changes (%Δ) are statistically significant at the p < 0.05 on almost all benchmarks, as measured by Wilcoxon rank-sum tests paired by cascade (highlighted in bold).","image":"img/2501.09345.png","kind":"table","image_w":1333,"image_h":561,"pwc":true}
site/p/2502.03503.json CHANGED
@@ -1 +1 @@
1
- {"id":"2502.03503","title":"Why Transformers Struggle with Distribution-Independent In-Context Learning","authors":"Omar Naim; Jerome Bolte; Nicholas Asher","openreview_id":"7Ci5cBtBUO","presentation":"poster","track":"main","verdict":"include","background":"In-context learning (ICL) lets a transformer solve a new task from input-output examples in its prompt without weight updates. A common interpretation is that the transformer implements a learning algorithm such as gradient descent or least squares in its forward pass. If it did, its predictions should stay accurate whenever the context uniquely determines the target function, even when the test distribution differs from training. The paper tests this in noiseless polynomial regression, where classical regression extrapolates exactly, and asks whether strong in-distribution ICL implies distribution-independent ICL.","method":["A GPT-style decoder-only transformer (12 layers, 8 heads, hidden size 256, scalar tokens embedded by a learned linear map and decoded by a learned linear readout, final LayerNorm) is trained from scratch with the Garg et al. (2022) objective: squared error on next-label prediction across prompts (x1, g(x1), ..., xk, g(xk), x). For the chosen claim, g is a degree-1 polynomial with coefficients from U(-1,1) and inputs from U(-1,1). Training uses Adam (lr 1e-4, batch 64) for 500k steps with fresh prompts each step and a curriculum on prompt length from 11 to 41 pairs.","At test time inputs stay in U(-1,1) but coefficients are drawn from U(-sigma, sigma) for sigma from 1 to 10. Evaluation uses 5 seeds x 100 functions x 64 prompts of 41 points; MSE is averaged over predictions after the first n+1 positions. Baselines get the same in-context examples: Nadaraya-Watson with an RBF kernel, kernel ridge regression (KRR) with an RBF kernel and with a polynomial kernel (lambda = 1e-3, RBF lengthscale = std of the training input grid, no tuning on shifted data). Ordinary least squares, which is exact here, is shown separately as the dotted baseline in Figure 1, not in Table 1."],"claim":"On degree-1 polynomial regression with inputs from U(-1,1), a transformer trained on coefficients from U(-1,1) has MSE 9e-6 at sigma=1, but 4.29 at sigma=5 and 27.06 at sigma=10 when test coefficients come from U(-sigma, sigma). On the same in-context examples, KRR with a polynomial kernel stays at or below 2.3e-3 and KRR with an RBF kernel at or below 0.23.","source":"Table 1, page 4 (transformer row with ± std in Table 3, page 15)","metric":"Test MSE (lower is better). Table 1 gives means with no error bars. Table 3 reports the transformer as mean ± std over 5 evaluation seeds (one trained model, 100 test functions per seed): 4.3 ± 0.2 at sigma=5 and 27.1 ± 2.3 at sigma=10. No error bars are given for the kernel baselines.","comparison":"Transformer (trained on U(-1,1) coefficients) vs. Nadaraya-Watson (RBF), KRR (RBF) and KRR (polynomial kernel) on identical in-context examples, across coefficient scales sigma in {1,2,3,5,7,10}","falsification":"A re-trained transformer with the stated setup keeps MSE within the same order of magnitude as KRR (RBF) at sigma=5 and sigma=10 (e.g. below about 0.5 at sigma=10), or its in-distribution MSE at sigma=1 is not near zero (so the in-distribution vs. shifted contrast disappears).","caption":"Table 1: The evolution of MSE for degree-1 polynomial functions distribution shift, comparing the Transformer with Nadaraya–Watson regression (NW) and kernel ridge regression (KRR). Test inputs are sampled from U(−1, 1) and coefficients from U(−σ, σ). All methods receive the same in-context examples. Kernel hyperparameters are specified in Appendix C.","image":"img/2502.03503.png","kind":"table","image_w":1144,"image_h":333,"pwc":false}
 
1
+ {"id":"2502.03503","title":"Why Transformers Struggle with Distribution-Independent In-Context Learning","authors":"Omar Naim; Jerome Bolte; Nicholas Asher","openreview_id":"7Ci5cBtBUO","presentation":"poster","track":"main","verdict":"include","background":"In-context learning (ICL) lets a transformer solve a new task from input-output examples in its prompt without weight updates. A common interpretation is that the transformer implements a learning algorithm such as gradient descent or least squares in its forward pass. If it did, its predictions should stay accurate whenever the context uniquely determines the target function, even when the test distribution differs from training. The paper tests this in noiseless polynomial regression, where classical regression extrapolates exactly, and asks whether strong in-distribution ICL implies distribution-independent ICL.","method":["A GPT-style decoder-only transformer (12 layers, 8 heads, hidden size 256, scalar tokens embedded by a learned linear map and decoded by a learned linear readout, final LayerNorm) is trained from scratch with the Garg et al. (2022) objective: squared error on next-label prediction across prompts (x1, g(x1), ..., xk, g(xk), x). For the chosen claim, g is a degree-1 polynomial with coefficients from U(-1,1) and inputs from U(-1,1). Training uses Adam (lr 1e-4, batch 64) for 500k steps with fresh prompts each step and a curriculum on prompt length from 11 to 41 pairs.","At test time inputs stay in U(-1,1) but coefficients are drawn from U(-sigma, sigma) for sigma from 1 to 10. Evaluation uses 5 seeds x 100 functions x 64 prompts of 41 points; MSE is averaged over predictions after the first n+1 positions. Baselines get the same in-context examples: Nadaraya-Watson with an RBF kernel, kernel ridge regression (KRR) with an RBF kernel and with a polynomial kernel (lambda = 1e-3, RBF lengthscale = std of the training input grid, no tuning on shifted data). Ordinary least squares, which is exact here, is shown separately as the dotted baseline in Figure 1, not in Table 1."],"claim":"On degree-1 polynomial regression with inputs from U(-1,1), a transformer trained on coefficients from U(-1,1) has MSE 9e-6 at sigma=1, but 4.29 at sigma=5 and 27.06 at sigma=10 when test coefficients come from U(-sigma, sigma). On the same in-context examples, KRR with a polynomial kernel stays at or below 2.3e-3 and KRR with an RBF kernel at or below 0.23.","source":"Table 1, page 4 (transformer row with ± std in Table 3, page 15)","metric":"Test MSE (lower is better). Table 1 gives means with no error bars. Table 3 reports the transformer as mean ± std over 5 evaluation seeds (one trained model, 100 test functions per seed): 4.3 ± 0.2 at sigma=5 and 27.1 ± 2.3 at sigma=10. No error bars are given for the kernel baselines.","comparison":"Transformer (trained on U(-1,1) coefficients) vs. Nadaraya-Watson (RBF), KRR (RBF) and KRR (polynomial kernel) on identical in-context examples, across coefficient scales sigma in {1,2,3,5,7,10}","falsification":"A re-trained transformer with the stated setup keeps MSE within the same order of magnitude as KRR (RBF) at sigma=5 and sigma=10 (e.g. below about 0.5 at sigma=10), or its in-distribution MSE at sigma=1 is not near zero (so the in-distribution vs. shifted contrast disappears).","caption":"Table 1: The evolution of MSE for degree-1 polynomial functions distribution shift, comparing the Transformer with Nadaraya–Watson regression (NW) and kernel ridge regression (KRR). Test inputs are sampled from U(−1, 1) and coefficients from U(−σ, σ). All methods receive the same in-context examples. Kernel hyperparameters are specified in Appendix C.","image":"img/2502.03503.png","kind":"table","image_w":1144,"image_h":333,"pwc":true}
site/p/2502.10259.json CHANGED
@@ -1 +1 @@
1
- {"id":"2502.10259","title":"MITO: A Millimeter-Wave Dataset and Simulator for Non-Line-of-Sight Perception","authors":"Tara Boroushaki; Laura Dodds; Cusuh Ham; Fadel Adib","openreview_id":"hJlnaYhXuZ","presentation":"poster","track":"evaluations_datasets","verdict":"borderline","background":"Millimeter-wave (mmWave) radar signals pass through everyday occluders such as cardboard, fabric and plastic, so they can image objects that cameras cannot see, for example inside a closed box. mmWave images look very different from optical images: they are mostly specular, have edge-diffraction artifacts, no color, and are complex-valued. Learning-based perception on them is held back by a lack of data. MITO provides a real dataset (550 3D SAR images of 76 YCB objects at 24 and 77 GHz, in line-of-sight and under cardboard) and an open-source simulator that turns any 3D triangle mesh into a synthetic mmWave image, so models can be trained on synthetic data.","method":["The simulator takes a mesh and a set of radar positions. For each position it finds the visible mesh vertices, models each vertex reflection as exp(-j 4π|l−v|/λ_j), sums them into a raw signal, and passes the signals through the same SAR imaging pipeline (Eq. 1) used for real data. There are two variants: a specular simulation keeps only vertices whose surface normal points back toward the radar (angle below a threshold τ), and an edge simulation keeps only vertices on sharp edges (dihedral angle above τ_e). A combined image is the normalized weighted sum (α1·I_s + α2·I_e)/(α1+α2).","To measure accuracy, real and synthetic images are each turned into point clouds (one point per voxel whose magnitude exceeds τ_P), aligned with ICP (Open3D), and compared with the 3D F-score at distance threshold τ_F. For the combined simulation, the F-score is the maximum over a set W of weight pairs (α1, α2). Specular-only is α=(1,0) and edge-only is α=(0,1). The 25th, 50th and 75th percentiles of F-score over all objects in the dataset are reported."],"claim":"Across all objects in MITO, the combined (specular + edge, best weight) simulation matches real 77 GHz mmWave images with a 3D F-score of 85% / 94% / 98% at the 25th / 50th / 75th percentiles, versus 68% / 90% / 97% for specular-only and 49% / 71% / 88% for edge-only simulation.","source":"Table 3, page 6","metric":"3D F-score (higher is better) between thresholded, ICP-aligned real and simulated point clouds; percentiles over objects. No error bars or seeds reported.","comparison":"Combined simulation (best F-score over the weight grid W) vs. specular-only (α1=1, α2=0) vs. edge-only (α1=0, α2=1), all run by the authors on the same objects with the same pipeline","falsification":"Rerunning the released simulator and evaluation on the MITO real images, the combined simulation's 25th-percentile F-score is not clearly above specular-only (e.g. the 85% vs 68% gap shrinks to a few points), or the median of edge-only is not well below both others.","caption":"Table 3: Simulation 3D F-Score for combined (Comb), specular (Spec), & edge simulations.","image":"img/2502.10259.png","kind":"table","image_w":300,"image_h":222,"pwc":false}
 
1
+ {"id":"2502.10259","title":"MITO: A Millimeter-Wave Dataset and Simulator for Non-Line-of-Sight Perception","authors":"Tara Boroushaki; Laura Dodds; Cusuh Ham; Fadel Adib","openreview_id":"hJlnaYhXuZ","presentation":"poster","track":"evaluations_datasets","verdict":"borderline","background":"Millimeter-wave (mmWave) radar signals pass through everyday occluders such as cardboard, fabric and plastic, so they can image objects that cameras cannot see, for example inside a closed box. mmWave images look very different from optical images: they are mostly specular, have edge-diffraction artifacts, no color, and are complex-valued. Learning-based perception on them is held back by a lack of data. MITO provides a real dataset (550 3D SAR images of 76 YCB objects at 24 and 77 GHz, in line-of-sight and under cardboard) and an open-source simulator that turns any 3D triangle mesh into a synthetic mmWave image, so models can be trained on synthetic data.","method":["The simulator takes a mesh and a set of radar positions. For each position it finds the visible mesh vertices, models each vertex reflection as exp(-j 4π|l−v|/λ_j), sums them into a raw signal, and passes the signals through the same SAR imaging pipeline (Eq. 1) used for real data. There are two variants: a specular simulation keeps only vertices whose surface normal points back toward the radar (angle below a threshold τ), and an edge simulation keeps only vertices on sharp edges (dihedral angle above τ_e). A combined image is the normalized weighted sum (α1·I_s + α2·I_e)/(α1+α2).","To measure accuracy, real and synthetic images are each turned into point clouds (one point per voxel whose magnitude exceeds τ_P), aligned with ICP (Open3D), and compared with the 3D F-score at distance threshold τ_F. For the combined simulation, the F-score is the maximum over a set W of weight pairs (��1, α2). Specular-only is α=(1,0) and edge-only is α=(0,1). The 25th, 50th and 75th percentiles of F-score over all objects in the dataset are reported."],"claim":"Across all objects in MITO, the combined (specular + edge, best weight) simulation matches real 77 GHz mmWave images with a 3D F-score of 85% / 94% / 98% at the 25th / 50th / 75th percentiles, versus 68% / 90% / 97% for specular-only and 49% / 71% / 88% for edge-only simulation.","source":"Table 3, page 6","metric":"3D F-score (higher is better) between thresholded, ICP-aligned real and simulated point clouds; percentiles over objects. No error bars or seeds reported.","comparison":"Combined simulation (best F-score over the weight grid W) vs. specular-only (α1=1, α2=0) vs. edge-only (α1=0, α2=1), all run by the authors on the same objects with the same pipeline","falsification":"Rerunning the released simulator and evaluation on the MITO real images, the combined simulation's 25th-percentile F-score is not clearly above specular-only (e.g. the 85% vs 68% gap shrinks to a few points), or the median of edge-only is not well below both others.","caption":"Table 3: Simulation 3D F-Score for combined (Comb), specular (Spec), & edge simulations.","image":"img/2502.10259.png","kind":"table","image_w":300,"image_h":222,"pwc":true}
site/p/2503.07338.json CHANGED
@@ -1 +1 @@
1
- {"id":"2503.07338","title":"Aligned Delta-Triplane Transformers as Occupancy World Models","authors":"Haoran Xu; Peixi Peng; Shijun Guo; Guang Tan; Yiqian Chang; Yisen Zhao; Shuaixian Wang; Yonghong Tian","openreview_id":"TVtLbFPjtb","presentation":"poster","track":"main","verdict":"include","background":"Occupancy world models predict future 3D semantic voxel grids around an autonomous vehicle from past occupancy frames and the ego vehicle's past and planned motion. Most existing models compress each frame into a bird's-eye-view latent, use a large Transformer or diffusion model to align past frames implicitly, and predict the full future scene. That is costly and redundant, because once ego motion is removed most of the scene does not change over a few seconds.","method":["ADTT first pretrains a triplane autoencoder (xy, xz, yz feature planes, 16x100x100, 8 channels) to reconstruct 16x200x200 Occ3D-nuScenes occupancy grids. Historical triplanes are then explicitly warped into the target ego frame using the recorded ego poses (nearest-neighbour sampling, zeros where there is no historical correspondence). A lightweight conv tokenizer maps each aligned triplane to an 8x8x4 grid of 256 tokens (dim 96). A Transformer encoder builds a memory from the last 4 aligned frames. A Transformer decoder with 64 ego-motion queries (from an MLP on the future ego motion) and 192 learnable external queries predicts delta tokens. These are upsampled to a delta triplane, added to the latest aligned triplane and decoded to occupancy. Prediction is autoregressive. Training uses L1+L2 triplane losses plus cross-entropy/Lovasz occupancy losses, with the triplane encoder frozen and the decoder fine-tuned.","Evaluation follows the standard Occ3D-nus protocol: 4 past frames (2 s) are used to predict 6 future frames (3 s), reporting mIoU over 18 semantic classes and occupancy IoU averaged over 1 s, 2 s and 3 s. In the Table 4 ablation each variant removes one component from the full model M0. M4 removes motion compensation, so the historical triplanes are not warped into the target frame. M5 removes delta prediction, so the model predicts the full future triplane instead of a residual added to the aligned prior."],"claim":"On Occ3D-nus 3-second forecasting, full ADTT reaches 46.87 average mIoU and 52.63 IoU. Removing triplane motion compensation drops this to 31.87 mIoU and 37.48 IoU, and predicting the full future state instead of deltas drops it to 36.65 mIoU and 45.50 IoU.","source":"Table 4, page 8","metric":"Average mIoU (%) over 18 semantic classes and average occupancy IoU (%) across 1 s, 2 s and 3 s horizons, higher is better. Single runs with no seeds, std or confidence intervals reported.","comparison":"Full ADTT (M0) vs. the same model without motion compensation (M4) and without delta prediction (M5), under the same training settings","falsification":"Retrained variants without motion compensation, or with full-state instead of delta prediction, come within a few mIoU points of the full model (e.g. a gap under about 3 mIoU instead of 15 and 10), or match or beat it.","caption":"Table 4: Ablation on key designs in ADTT.","image":"img/2503.07338.png","kind":"table","image_w":544,"image_h":330,"pwc":false}
 
1
+ {"id":"2503.07338","title":"Aligned Delta-Triplane Transformers as Occupancy World Models","authors":"Haoran Xu; Peixi Peng; Shijun Guo; Guang Tan; Yiqian Chang; Yisen Zhao; Shuaixian Wang; Yonghong Tian","openreview_id":"TVtLbFPjtb","presentation":"poster","track":"main","verdict":"include","background":"Occupancy world models predict future 3D semantic voxel grids around an autonomous vehicle from past occupancy frames and the ego vehicle's past and planned motion. Most existing models compress each frame into a bird's-eye-view latent, use a large Transformer or diffusion model to align past frames implicitly, and predict the full future scene. That is costly and redundant, because once ego motion is removed most of the scene does not change over a few seconds.","method":["ADTT first pretrains a triplane autoencoder (xy, xz, yz feature planes, 16x100x100, 8 channels) to reconstruct 16x200x200 Occ3D-nuScenes occupancy grids. Historical triplanes are then explicitly warped into the target ego frame using the recorded ego poses (nearest-neighbour sampling, zeros where there is no historical correspondence). A lightweight conv tokenizer maps each aligned triplane to an 8x8x4 grid of 256 tokens (dim 96). A Transformer encoder builds a memory from the last 4 aligned frames. A Transformer decoder with 64 ego-motion queries (from an MLP on the future ego motion) and 192 learnable external queries predicts delta tokens. These are upsampled to a delta triplane, added to the latest aligned triplane and decoded to occupancy. Prediction is autoregressive. Training uses L1+L2 triplane losses plus cross-entropy/Lovasz occupancy losses, with the triplane encoder frozen and the decoder fine-tuned.","Evaluation follows the standard Occ3D-nus protocol: 4 past frames (2 s) are used to predict 6 future frames (3 s), reporting mIoU over 18 semantic classes and occupancy IoU averaged over 1 s, 2 s and 3 s. In the Table 4 ablation each variant removes one component from the full model M0. M4 removes motion compensation, so the historical triplanes are not warped into the target frame. M5 removes delta prediction, so the model predicts the full future triplane instead of a residual added to the aligned prior."],"claim":"On Occ3D-nus 3-second forecasting, full ADTT reaches 46.87 average mIoU and 52.63 IoU. Removing triplane motion compensation drops this to 31.87 mIoU and 37.48 IoU, and predicting the full future state instead of deltas drops it to 36.65 mIoU and 45.50 IoU.","source":"Table 4, page 8","metric":"Average mIoU (%) over 18 semantic classes and average occupancy IoU (%) across 1 s, 2 s and 3 s horizons, higher is better. Single runs with no seeds, std or confidence intervals reported.","comparison":"Full ADTT (M0) vs. the same model without motion compensation (M4) and without delta prediction (M5), under the same training settings","falsification":"Retrained variants without motion compensation, or with full-state instead of delta prediction, come within a few mIoU points of the full model (e.g. a gap under about 3 mIoU instead of 15 and 10), or match or beat it.","caption":"Table 4: Ablation on key designs in ADTT.","image":"img/2503.07338.png","kind":"table","image_w":544,"image_h":330,"pwc":true}
site/p/2505.13775.json CHANGED
@@ -1 +1 @@
1
- {"id":"2505.13775","title":"Beyond Semantics: The Unreasonable Effectiveness of Reasonless Intermediate Tokens","authors":"Karthik Valmeekam; Vardhan Palod; Kaya Stechly; Atharva Gundawar; Subbarao Kambhampati","openreview_id":"","presentation":"poster","track":"journal:TMLR","verdict":"include","background":"Reasoning models are trained to emit intermediate tokens ('chains of thought') before an answer, and these traces are often read as faithful records of the model's reasoning. It is unclear whether the semantic correctness of the traces is what makes them useful. In large pretrained models this cannot be checked, because traces are natural language with no ground truth. A small, fully controlled setting where traces are formally verifiable (A* search on grid mazes) lets one test directly whether training on semantically correct traces matters for final-answer accuracy.","method":["The authors train a Qwen2.5-0.5B architecture from scratch with a custom 944-token vocabulary (~380M parameters, 32k context) on 30x30 maze shortest-path problems. Each training example is a maze description, start and goal, a linearized A* execution trace (create/close x y cG cH operations, Manhattan heuristic), and the correct plan. Training sets have 500k problems; 250k steps, batch size 8, AdamW, peak LR 2.2758e-4, weight decay 0.1528, 100 warm-up steps, bf16. Mazes are generated either with Wilson's algorithm (acyclic) or with Searchformer-style random wall placement (SF-style, 30-50% walls, can contain loops).","The 'Swapped' dataset contains the same problems and correct plans, but each problem's A* trace is randomly permuted to a different problem, so traces have the right format but no relation to the maze they accompany. 'Normal' models train on the correct traces. Both are evaluated on 1000-problem test sets from five generators (Wilson, Kruskal, DFS, Drunkard's Walk, SF-style) by plan accuracy (a plan is valid if every move goes between adjacent free cells and ends at the goal; optimality is not required) and by trace validity among valid plans, checked by a formal A* trace verifier."],"claim":"For models trained from scratch on 500k SF-style maze problems, training on traces swapped between problems gives higher plan accuracy than training on correct A* traces, on both initialization seeds: 89.1% and 86.8% vs. 56.2% and 64.9% on in-distribution SF-style mazes, and 95.4% and 87.8% vs. 62.1% and 66.3% on out-of-distribution Drunkard's Walk mazes, while the swapped models have 0% trace validity.","source":"Table 6, page 19 (Appendix A.2); seed-1 values also in Table 3, page 10","metric":"Plan accuracy (%) on 1000 test problems per maze generator, higher is better; trace validity within valid plans (%) is also shown. No ± reported; two training seeds (original and Seed 2) are reported as separate rows.","comparison":"Swapped (SF-style) vs. Normal (SF-style), same architecture, same problems and plans, same training budget, matched per initialization seed","falsification":"A Normal (correct-trace) SF-style model trained under the same setup reaches plan accuracy within a few points of, or above, the Swapped model on the SF-style or Drunkard test sets, or the Swapped model's advantage disappears on a fresh seed.","caption":"Table 6: SF-style model performance across seeds. Original denotes the models whose performance were reported at Table 3. Seed 2 are the models trained on a different seed. Each cell shows Plan Accuracy (%) / Trace Validity within valid plans (%).","image":"img/2505.13775.png","kind":"table","image_w":1333,"image_h":350,"pwc":false}
 
1
+ {"id":"2505.13775","title":"Beyond Semantics: The Unreasonable Effectiveness of Reasonless Intermediate Tokens","authors":"Karthik Valmeekam; Vardhan Palod; Kaya Stechly; Atharva Gundawar; Subbarao Kambhampati","openreview_id":"","presentation":"poster","track":"journal:TMLR","verdict":"include","background":"Reasoning models are trained to emit intermediate tokens ('chains of thought') before an answer, and these traces are often read as faithful records of the model's reasoning. It is unclear whether the semantic correctness of the traces is what makes them useful. In large pretrained models this cannot be checked, because traces are natural language with no ground truth. A small, fully controlled setting where traces are formally verifiable (A* search on grid mazes) lets one test directly whether training on semantically correct traces matters for final-answer accuracy.","method":["The authors train a Qwen2.5-0.5B architecture from scratch with a custom 944-token vocabulary (~380M parameters, 32k context) on 30x30 maze shortest-path problems. Each training example is a maze description, start and goal, a linearized A* execution trace (create/close x y cG cH operations, Manhattan heuristic), and the correct plan. Training sets have 500k problems; 250k steps, batch size 8, AdamW, peak LR 2.2758e-4, weight decay 0.1528, 100 warm-up steps, bf16. Mazes are generated either with Wilson's algorithm (acyclic) or with Searchformer-style random wall placement (SF-style, 30-50% walls, can contain loops).","The 'Swapped' dataset contains the same problems and correct plans, but each problem's A* trace is randomly permuted to a different problem, so traces have the right format but no relation to the maze they accompany. 'Normal' models train on the correct traces. Both are evaluated on 1000-problem test sets from five generators (Wilson, Kruskal, DFS, Drunkard's Walk, SF-style) by plan accuracy (a plan is valid if every move goes between adjacent free cells and ends at the goal; optimality is not required) and by trace validity among valid plans, checked by a formal A* trace verifier."],"claim":"For models trained from scratch on 500k SF-style maze problems, training on traces swapped between problems gives higher plan accuracy than training on correct A* traces, on both initialization seeds: 89.1% and 86.8% vs. 56.2% and 64.9% on in-distribution SF-style mazes, and 95.4% and 87.8% vs. 62.1% and 66.3% on out-of-distribution Drunkard's Walk mazes, while the swapped models have 0% trace validity.","source":"Table 6, page 19 (Appendix A.2); seed-1 values also in Table 3, page 10","metric":"Plan accuracy (%) on 1000 test problems per maze generator, higher is better; trace validity within valid plans (%) is also shown. No ± reported; two training seeds (original and Seed 2) are reported as separate rows.","comparison":"Swapped (SF-style) vs. Normal (SF-style), same architecture, same problems and plans, same training budget, matched per initialization seed","falsification":"A Normal (correct-trace) SF-style model trained under the same setup reaches plan accuracy within a few points of, or above, the Swapped model on the SF-style or Drunkard test sets, or the Swapped model's advantage disappears on a fresh seed.","caption":"Table 6: SF-style model performance across seeds. Original denotes the models whose performance were reported at Table 3. Seed 2 are the models trained on a different seed. Each cell shows Plan Accuracy (%) / Trace Validity within valid plans (%).","image":"img/2505.13775.png","kind":"table","image_w":1333,"image_h":350,"pwc":true}
site/p/2505.22252.json CHANGED
@@ -1 +1 @@
1
- {"id":"2505.22252","title":"B-XAIC Dataset: Benchmarking Explainable AI for Graph Neural Networks Using Chemical Data","authors":"Magdalena Proszewska; Tomasz Danel; Antoni Antoszek; Dawid Damian Rymarczyk","openreview_id":"GHoBerj4sM","presentation":"poster","track":"evaluations_datasets","verdict":"include","background":"Graph neural networks are the standard model for predicting molecular properties, and many explainability (XAI) methods claim to show which atoms and bonds drive a prediction. Evaluating these explainers is hard because real molecular datasets rarely have known ground-truth rationales, and existing benchmarks are either synthetic or small single-task sets such as MUTAG. Common metrics also threshold importance maps or take top-k elements, which breaks down when no substructure is relevant at all. Without a benchmark with known rationales, it is unclear whether popular explainers recover the rules a model actually uses.","method":["B-XAIC samples 50K molecules from ChEMBL 35 (8/1/1 random split) and labels them for 7 rule-based binary tasks: presence of boron (B), phosphorus (P), halogens (X), indole, PAINS alerts, more than four rings (rings-count), and a ring with more than six atoms (rings-max). Each positive molecule has node- and edge-level ground-truth explanations (the atoms/bonds of the matched pattern). GCN, GAT, GIN and ProtGNN are trained on the tasks; GIN reaches weighted-F1 of 92.9-99.98% on all tasks (Table 3) and is the model the explainers are evaluated on in the main text.","Explanations are scored in two regimes. Null explanation (NE): for a molecule without the pattern, the explanation should contain no outliers; the score is 1 if Tukey's 1.5x IQR rule finds no outlier attributions and 0 otherwise, averaged over instances. Subgraph explanation (SE): for a molecule with the pattern, the AUROC of the attribution scores against the ground-truth nodes/edges. Explainers include gradient-based methods (Saliency, Deconvolution, InputXGradient, GuidedBackprop, IntegratedGradients), graph-specific methods (GNNExplainer, GraphMask, PGExplainer, PGMExplainer, FlowX, SubgraphX, GraphSHAP-IQ), perturbation-based (ShapleyValueSampling, EdgeSHAPer) and surrogate methods (DnX, FastDnX). Table 4 reports NE and SE averaged over the 7 tasks (nodes) or 4 ring/indole/PAINS tasks (edges)."],"claim":"On B-XAIC with a trained GIN, node-level gradient-based explainers other than Deconvolution score low on null explanations (Saliency 0.51 ± 0.12, InputXGradient 0.54 ± 0.14, GuidedBackprop 0.41 ± 0.12, IntegratedGradients 0.39 ± 0.26, averaged over 7 tasks) while PGMExplainer scores 0.98 ± 0.02 and SubgraphX 0.94 ± 0.05, even though these gradient methods rank the true pattern well (SE 0.83-0.87); Deconvolution is the gradient-based exception (NE 0.79 ± 0.10).","source":"Table 4, page 7","metric":"NE: fraction of pattern-free molecules whose node attributions contain no 1.5x IQR outliers (higher is better). SE: node-level AUROC against ground-truth pattern atoms (higher is better). Values are means over the 7 tasks; the ± is the standard deviation across the 7 tasks (the Table 4 caption says only 'uncertainty', but Table 10 reports the same node NE values as 'mean ± standard deviation across seven tasks'). Bold marks scores not significantly worse than the best under the paper's one-sided Wilcoxon test.","comparison":"Gradient-based explainers (Saliency, GuidedBackprop, IntegratedGradients, InputXGradient) vs. PGMExplainer and SubgraphX, all applied to the same trained GIN on the same B-XAIC test set.","falsification":"Using a GIN trained to similar accuracy on B-XAIC, Saliency (or IntegratedGradients/GuidedBackprop) reaches a task-averaged node NE within about 0.2 of PGMExplainer, or PGMExplainer's node NE falls below about 0.8.","caption":"Table 4: Evaluation of the node and edge explanations for the GIN model. Explainers are grouped as gradient-based (GB), graph-specific (GS), perturbation-based (PB), and surrogate (SU). Bold marks the best score and those not significantly worse (one-sided Wilcoxon test). Avg is the mean of NE and SE; uncertainty is reported for NE and SE.","image":"img/2505.22252.png","kind":"table","image_w":1144,"image_h":638,"pwc":false}
 
1
+ {"id":"2505.22252","title":"B-XAIC Dataset: Benchmarking Explainable AI for Graph Neural Networks Using Chemical Data","authors":"Magdalena Proszewska; Tomasz Danel; Antoni Antoszek; Dawid Damian Rymarczyk","openreview_id":"GHoBerj4sM","presentation":"poster","track":"evaluations_datasets","verdict":"include","background":"Graph neural networks are the standard model for predicting molecular properties, and many explainability (XAI) methods claim to show which atoms and bonds drive a prediction. Evaluating these explainers is hard because real molecular datasets rarely have known ground-truth rationales, and existing benchmarks are either synthetic or small single-task sets such as MUTAG. Common metrics also threshold importance maps or take top-k elements, which breaks down when no substructure is relevant at all. Without a benchmark with known rationales, it is unclear whether popular explainers recover the rules a model actually uses.","method":["B-XAIC samples 50K molecules from ChEMBL 35 (8/1/1 random split) and labels them for 7 rule-based binary tasks: presence of boron (B), phosphorus (P), halogens (X), indole, PAINS alerts, more than four rings (rings-count), and a ring with more than six atoms (rings-max). Each positive molecule has node- and edge-level ground-truth explanations (the atoms/bonds of the matched pattern). GCN, GAT, GIN and ProtGNN are trained on the tasks; GIN reaches weighted-F1 of 92.9-99.98% on all tasks (Table 3) and is the model the explainers are evaluated on in the main text.","Explanations are scored in two regimes. Null explanation (NE): for a molecule without the pattern, the explanation should contain no outliers; the score is 1 if Tukey's 1.5x IQR rule finds no outlier attributions and 0 otherwise, averaged over instances. Subgraph explanation (SE): for a molecule with the pattern, the AUROC of the attribution scores against the ground-truth nodes/edges. Explainers include gradient-based methods (Saliency, Deconvolution, InputXGradient, GuidedBackprop, IntegratedGradients), graph-specific methods (GNNExplainer, GraphMask, PGExplainer, PGMExplainer, FlowX, SubgraphX, GraphSHAP-IQ), perturbation-based (ShapleyValueSampling, EdgeSHAPer) and surrogate methods (DnX, FastDnX). Table 4 reports NE and SE averaged over the 7 tasks (nodes) or 4 ring/indole/PAINS tasks (edges)."],"claim":"On B-XAIC with a trained GIN, node-level gradient-based explainers other than Deconvolution score low on null explanations (Saliency 0.51 ± 0.12, InputXGradient 0.54 ± 0.14, GuidedBackprop 0.41 ± 0.12, IntegratedGradients 0.39 ± 0.26, averaged over 7 tasks) while PGMExplainer scores 0.98 ± 0.02 and SubgraphX 0.94 ± 0.05, even though these gradient methods rank the true pattern well (SE 0.83-0.87); Deconvolution is the gradient-based exception (NE 0.79 ± 0.10).","source":"Table 4, page 7","metric":"NE: fraction of pattern-free molecules whose node attributions contain no 1.5x IQR outliers (higher is better). SE: node-level AUROC against ground-truth pattern atoms (higher is better). Values are means over the 7 tasks; the ± is the standard deviation across the 7 tasks (the Table 4 caption says only 'uncertainty', but Table 10 reports the same node NE values as 'mean ± standard deviation across seven tasks'). Bold marks scores not significantly worse than the best under the paper's one-sided Wilcoxon test.","comparison":"Gradient-based explainers (Saliency, GuidedBackprop, IntegratedGradients, InputXGradient) vs. PGMExplainer and SubgraphX, all applied to the same trained GIN on the same B-XAIC test set.","falsification":"Using a GIN trained to similar accuracy on B-XAIC, Saliency (or IntegratedGradients/GuidedBackprop) reaches a task-averaged node NE within about 0.2 of PGMExplainer, or PGMExplainer's node NE falls below about 0.8.","caption":"Table 4: Evaluation of the node and edge explanations for the GIN model. Explainers are grouped as gradient-based (GB), graph-specific (GS), perturbation-based (PB), and surrogate (SU). Bold marks the best score and those not significantly worse (one-sided Wilcoxon test). Avg is the mean of NE and SE; uncertainty is reported for NE and SE.","image":"img/2505.22252.png","kind":"table","image_w":1144,"image_h":638,"pwc":true}
site/p/2505.22919.json CHANGED
@@ -1 +1 @@
1
- {"id":"2505.22919","title":"ER-Reason: A Benchmark Dataset for LLM Clinical Reasoning in the Emergency Room","authors":"Nikita Mehandru; Niloufar Golchini; Namrata Garg; Kathy T LeSaint; Christopher J Nash; Anu Ramachandran; Travis Zack; Liam Greg McCoy; Adam Rodman; David Bamman; Melanie Molina; Ahmed Alaa","openreview_id":"tI6dUEC7KR","presentation":"poster","track":"evaluations_datasets","verdict":"borderline","background":"LLMs score well on medical-licensing-style multiple-choice benchmarks, but prospective studies show much weaker clinical usefulness. A common critique is construct validity: most benchmarks measure knowledge recall or final diagnostic accuracy on stylized vignettes rather than clinical reasoning. The paper defines clinical reasoning as revising beliefs over a differential diagnosis as evidence accumulates (Ledley and Lusted's sequential testing loop). ER-Reason supplies 25,174 de-identified real emergency-room notes from 3,437 patients, with tasks across the ER workflow and Script Concordance Test (SCT)-style items, to measure whether models update diagnostic beliefs the way emergency physicians do.","method":["Emergency medicine attendings wrote 194 SCT-style cases from real ER provider notes. Each case gives a clinical scenario, 3-5 candidate diagnoses and a sequence of evidence steps (a lab, imaging, or exam finding), for 787 SCT items in total. At each step the respondent rates how the new evidence changes the likelihood of the diagnosis under consideration on a -2..+2 scale and re-ranks the differential. Two attending physicians annotated every item independently, a third adjudicated disagreements, and the consensus is the ground truth. The annotators also wrote a rationale for every step.","Eight LLMs (DeepSeek-R1, o4-mini, Gemini 2.5 Flash, GPT-5.2 Thinking, Claude Sonnet 4.5 Thinking, Phi-4, GPT-5.2, Claude Sonnet 4.5; temperature 0.1 where supported, via Azure OpenAI or OpenRouter) are prompted in three conditions. Zero-shot (ZS) gives the case, the differential and the new evidence. Single oracle (SO) also gives the physician rationale for the current step. Full oracle (FO) gives all physician rationales from steps 1..t. DxUpdate is the linearly weighted Cohen's kappa between model and physician ordinal scores, averaged over timesteps. DxTrajectory is the mean Spearman rho between model and physician differential rankings over all encounter-timestep pairs."],"claim":"On the 194-case ER-Reason SCT set, giving the model the physician's rationale for the current step (single oracle) raises DxUpdate for every one of 8 LLMs above the upper end of its zero-shot 95% CI, e.g. GPT-5.2 0.677 (0.646-0.701) -> 0.795 and o4-mini 0.691 (0.664-0.713) -> 0.809; DxTrajectory rises likewise (GPT-5.2 0.646 (0.618-0.680) -> 0.780 SO, 0.813 FO).","source":"Table 2, page 7","metric":"DxUpdate = linearly weighted Cohen's kappa vs. physician consensus (higher is better); DxTrajectory = mean Spearman rho of differential rankings (higher is better). Parentheses are 95% CIs from 1,000 bootstrap resamples, reported for zero-shot only; SO/FO are point estimates with no CI. One run per condition.","comparison":"Same model, same cases: zero-shot vs. single-oracle vs. full-oracle prompting, for each of 8 LLMs","falsification":"For the same (or comparable current) models, single-oracle DxUpdate or DxTrajectory falls inside the zero-shot bootstrap 95% CI (a gain of less than roughly 0.03-0.04) for most models.","caption":"Table 2: Performance in SCT reasoning tasks. (95% confidence intervals reported from 1,000 bootstraps.)","image":"img/2505.22919.png","kind":"table","image_w":1125,"image_h":388,"pwc":false}
 
1
+ {"id":"2505.22919","title":"ER-Reason: A Benchmark Dataset for LLM Clinical Reasoning in the Emergency Room","authors":"Nikita Mehandru; Niloufar Golchini; Namrata Garg; Kathy T LeSaint; Christopher J Nash; Anu Ramachandran; Travis Zack; Liam Greg McCoy; Adam Rodman; David Bamman; Melanie Molina; Ahmed Alaa","openreview_id":"tI6dUEC7KR","presentation":"poster","track":"evaluations_datasets","verdict":"borderline","background":"LLMs score well on medical-licensing-style multiple-choice benchmarks, but prospective studies show much weaker clinical usefulness. A common critique is construct validity: most benchmarks measure knowledge recall or final diagnostic accuracy on stylized vignettes rather than clinical reasoning. The paper defines clinical reasoning as revising beliefs over a differential diagnosis as evidence accumulates (Ledley and Lusted's sequential testing loop). ER-Reason supplies 25,174 de-identified real emergency-room notes from 3,437 patients, with tasks across the ER workflow and Script Concordance Test (SCT)-style items, to measure whether models update diagnostic beliefs the way emergency physicians do.","method":["Emergency medicine attendings wrote 194 SCT-style cases from real ER provider notes. Each case gives a clinical scenario, 3-5 candidate diagnoses and a sequence of evidence steps (a lab, imaging, or exam finding), for 787 SCT items in total. At each step the respondent rates how the new evidence changes the likelihood of the diagnosis under consideration on a -2..+2 scale and re-ranks the differential. Two attending physicians annotated every item independently, a third adjudicated disagreements, and the consensus is the ground truth. The annotators also wrote a rationale for every step.","Eight LLMs (DeepSeek-R1, o4-mini, Gemini 2.5 Flash, GPT-5.2 Thinking, Claude Sonnet 4.5 Thinking, Phi-4, GPT-5.2, Claude Sonnet 4.5; temperature 0.1 where supported, via Azure OpenAI or OpenRouter) are prompted in three conditions. Zero-shot (ZS) gives the case, the differential and the new evidence. Single oracle (SO) also gives the physician rationale for the current step. Full oracle (FO) gives all physician rationales from steps 1..t. DxUpdate is the linearly weighted Cohen's kappa between model and physician ordinal scores, averaged over timesteps. DxTrajectory is the mean Spearman rho between model and physician differential rankings over all encounter-timestep pairs."],"claim":"On the 194-case ER-Reason SCT set, giving the model the physician's rationale for the current step (single oracle) raises DxUpdate for every one of 8 LLMs above the upper end of its zero-shot 95% CI, e.g. GPT-5.2 0.677 (0.646-0.701) -> 0.795 and o4-mini 0.691 (0.664-0.713) -> 0.809; DxTrajectory rises likewise (GPT-5.2 0.646 (0.618-0.680) -> 0.780 SO, 0.813 FO).","source":"Table 2, page 7","metric":"DxUpdate = linearly weighted Cohen's kappa vs. physician consensus (higher is better); DxTrajectory = mean Spearman rho of differential rankings (higher is better). Parentheses are 95% CIs from 1,000 bootstrap resamples, reported for zero-shot only; SO/FO are point estimates with no CI. One run per condition.","comparison":"Same model, same cases: zero-shot vs. single-oracle vs. full-oracle prompting, for each of 8 LLMs","falsification":"For the same (or comparable current) models, single-oracle DxUpdate or DxTrajectory falls inside the zero-shot bootstrap 95% CI (a gain of less than roughly 0.03-0.04) for most models.","caption":"Table 2: Performance in SCT reasoning tasks. (95% confidence intervals reported from 1,000 bootstraps.)","image":"img/2505.22919.png","kind":"table","image_w":1125,"image_h":388,"pwc":true}
site/p/2505.24791.json CHANGED
@@ -1 +1 @@
1
- {"id":"2505.24791","title":"Accelerating Inference of Discrete Autoregressive Normalizing Flows by Selective Jacobi Decoding","authors":"Jiaru Zhang; Juanwu Lu; Xiaoyu Wu; Ziran Wang; Ruqi Zhang","openreview_id":"","presentation":"poster","track":"journal:TMLR","verdict":"include","background":"Autoregressive normalizing flows such as TarFlow reach state-of-the-art likelihoods and image quality by splitting an image into a sequence of patches and transforming each one conditioned on all previous ones with a causal transformer. Training is parallel, but sampling requires inverting every flow block, and the inverse is strictly sequential: each patch must wait for all earlier patches. For a 256x256 model with 1024 patches and 8 blocks, this makes generation slow. Jacobi decoding, which updates all positions in parallel from the previous iterate until a fixed point is reached, has been tried for language models and autoregressive image models with modest gains. This paper asks whether it works better for the deterministic, continuous inverse of a normalizing flow.","method":["Inverting one autoregressive flow block is treated as solving a triangular system of L nonlinear equations. Jacobi decoding (Algorithm 1) starts from z^0 = 0 and, in each iteration, recomputes every patch z_{k,l} in parallel from the previous iterate's prefix z^{t-1}_{k,<l} and the block input z_{k+1}, stopping when the max-norm change between iterates is below tau = 0.5. It converges exactly in at most L iterations, and in practice in about 4-7 iterations for most blocks (Table A3). No retraining or architecture change is needed.","Selective Jacobi Decoding (SJD, 'Ours') keeps standard sequential decoding with a KV cache for the first inverted block (the one that starts from Gaussian noise, where dependencies are strongest) and uses Jacobi iteration for all remaining blocks. Baselines are standard sequential (KV-cached) inference and uniform Jacobi decoding (UJD, Jacobi on all blocks). Time is the average seconds per batch of 256 over 10 runs on two L40S GPUs. FID is computed against the real dataset using as many generated samples as the dataset has images. CLIP-IQA and BRISQUE are also reported. AFHQ uses the released TarFlow 256x256 checkpoint (patch size 8, L = 1024, 8 blocks)."],"claim":"On AFHQ 256x256 with the released TarFlow checkpoint, selective Jacobi decoding generates a batch of 256 in 41.21 s vs. 186.28 s for sequential inference (4.5x faster) and 219.24 s for uniform Jacobi decoding, with FID 15.44 vs. 15.42 (sequential) and 15.44 (UJD).","source":"Table 1, page 8","metric":"Wall-clock time per batch in seconds (lower is better; mean over 10 batches, two L40S GPUs) and FID (lower is better). For AFHQ no error bars are reported (single run); for CIFAR-10/100 the ± is the maximum deviation over three runs.","comparison":"Selective Jacobi Decoding vs. standard sequential inference with KV cache and uniform Jacobi decoding (all blocks), same checkpoint, tau = 0.5, zero initialization","falsification":"On the released AFHQ checkpoint, SJD gives less than about a 2x wall-clock speedup over a properly KV-cached sequential sampler at the same batch size and hardware, or its FID rises by more than about 0.5 over sequential sampling, or UJD turns out to be as fast as SJD.","caption":"Table 1: Comparison of sequential inference, uniform Jacobi decoding, and our approach. The subscript indicates the maximum deviation in three runs.","image":"img/2505.24791.png","kind":"table","image_w":1333,"image_h":488,"pwc":false}
 
1
+ {"id":"2505.24791","title":"Accelerating Inference of Discrete Autoregressive Normalizing Flows by Selective Jacobi Decoding","authors":"Jiaru Zhang; Juanwu Lu; Xiaoyu Wu; Ziran Wang; Ruqi Zhang","openreview_id":"","presentation":"poster","track":"journal:TMLR","verdict":"include","background":"Autoregressive normalizing flows such as TarFlow reach state-of-the-art likelihoods and image quality by splitting an image into a sequence of patches and transforming each one conditioned on all previous ones with a causal transformer. Training is parallel, but sampling requires inverting every flow block, and the inverse is strictly sequential: each patch must wait for all earlier patches. For a 256x256 model with 1024 patches and 8 blocks, this makes generation slow. Jacobi decoding, which updates all positions in parallel from the previous iterate until a fixed point is reached, has been tried for language models and autoregressive image models with modest gains. This paper asks whether it works better for the deterministic, continuous inverse of a normalizing flow.","method":["Inverting one autoregressive flow block is treated as solving a triangular system of L nonlinear equations. Jacobi decoding (Algorithm 1) starts from z^0 = 0 and, in each iteration, recomputes every patch z_{k,l} in parallel from the previous iterate's prefix z^{t-1}_{k,<l} and the block input z_{k+1}, stopping when the max-norm change between iterates is below tau = 0.5. It converges exactly in at most L iterations, and in practice in about 4-7 iterations for most blocks (Table A3). No retraining or architecture change is needed.","Selective Jacobi Decoding (SJD, 'Ours') keeps standard sequential decoding with a KV cache for the first inverted block (the one that starts from Gaussian noise, where dependencies are strongest) and uses Jacobi iteration for all remaining blocks. Baselines are standard sequential (KV-cached) inference and uniform Jacobi decoding (UJD, Jacobi on all blocks). Time is the average seconds per batch of 256 over 10 runs on two L40S GPUs. FID is computed against the real dataset using as many generated samples as the dataset has images. CLIP-IQA and BRISQUE are also reported. AFHQ uses the released TarFlow 256x256 checkpoint (patch size 8, L = 1024, 8 blocks)."],"claim":"On AFHQ 256x256 with the released TarFlow checkpoint, selective Jacobi decoding generates a batch of 256 in 41.21 s vs. 186.28 s for sequential inference (4.5x faster) and 219.24 s for uniform Jacobi decoding, with FID 15.44 vs. 15.42 (sequential) and 15.44 (UJD).","source":"Table 1, page 8","metric":"Wall-clock time per batch in seconds (lower is better; mean over 10 batches, two L40S GPUs) and FID (lower is better). For AFHQ no error bars are reported (single run); for CIFAR-10/100 the ± is the maximum deviation over three runs.","comparison":"Selective Jacobi Decoding vs. standard sequential inference with KV cache and uniform Jacobi decoding (all blocks), same checkpoint, tau = 0.5, zero initialization","falsification":"On the released AFHQ checkpoint, SJD gives less than about a 2x wall-clock speedup over a properly KV-cached sequential sampler at the same batch size and hardware, or its FID rises by more than about 0.5 over sequential sampling, or UJD turns out to be as fast as SJD.","caption":"Table 1: Comparison of sequential inference, uniform Jacobi decoding, and our approach. The subscript indicates the maximum deviation in three runs.","image":"img/2505.24791.png","kind":"table","image_w":1333,"image_h":488,"pwc":true}
site/p/2506.06361.json CHANGED
@@ -1 +1 @@
1
- {"id":"2506.06361","title":"Tactile MNIST: Benchmarking Active Tactile Perception","authors":"Tim Schneider; Guillaume Duret; Cristiana de Farias; Roberto Calandra; Liming Chen; Jan Peters","openreview_id":"3yMkqerEOt","presentation":"spotlight","track":"evaluations_datasets","verdict":"borderline","background":"Tactile sensors such as the GelSight Mini give rich but very local information, so an agent has to choose where to touch and combine many touches over time to recognise an object (active tactile perception). The field has no standard benchmark. The paper introduces ap_gym (a Gymnasium-compatible active perception framework) and the Tactile MNIST Benchmark Suite (TMBS), which contains simulated tactile tasks, 13,580 3D-printable MNIST digit meshes, and a real dataset of 153,600 GelSight touches on 600 3D-printed digits. A CycleGAN trained on the real data renders realistic tactile images in simulation. Whether policies trained in simulation work on real tactile images is a key question for this kind of benchmark.","method":["The task is TactileMNIST classification with 16 touches per episode. Training uses TactileMNISTSnap, where each step the commanded sensor position snaps to the closest of 16 uniformly sampled positions. Real evaluation uses TactileMNISTRealSnap, which replays recorded real touches from the test split: the commanded position snaps to the closest of the next 16 prerecorded touches of the current digit. The agent is APPLE-SAC (a transformer actor-critic with a vision encoder), trained for 2.5M environment steps on 5 seeds per configuration.","The full sim-to-real recipe has four parts: (1) during training, simulated depth maps are rendered with the CycleGAN depth-to-RGB generator; (2) at test time, each real tactile image is passed through a CycleGAN round trip (RGB-to-depth, then depth-to-RGB); (3) during training, a background depth overlay adds the round-trip depth estimate of a random no-contact real touch to each simulated depth map; (4) contact geometry variation erodes or dilates and blurs the simulated indentation. Each ablation removes one part. 'No CycleGAN' renders with Taxim during training and evaluates on raw real images. The metric is accuracy after the last touch, averaged over evaluations between 2.0M and 2.5M steps, reported as mean ± std over the 5 seeds."],"claim":"With the full sim-to-real recipe, APPLE-SAC trained only in simulation reaches 65.7 ± 4.3% final classification accuracy on real TactileMNIST test digits. Removing the CycleGAN gives 12.4 ± 1.7%, removing the real-image round trip gives 19.0 ± 2.1%, and removing the background depth overlay gives 31.9 ± 7.7%. Simulated accuracy stays near or above 68% in every configuration, so simulated accuracy does not predict transfer.","source":"Figure 42, page 44 (numbers in the Appendix H text, pages 44-45)","metric":"Final-step classification accuracy (%) on TactileMNISTRealSnap test split (higher is better), averaged over evaluations between 2.0M and 2.5M env steps; ± is std over 5 seeds. Chance is 10%.","comparison":"Full recipe vs. four single-component ablations (no CycleGAN / no real-image round trip / no background depth overlay / no contact geometry variation), all APPLE-SAC, same training budget; real vs. simulated accuracy for each.","falsification":"The full recipe's real-data accuracy falls to or below the 'no real-image round trip' or 'no background depth overlay' ablation (or stays within about one std of either), or the full recipe does not clearly beat chance (10%) on the real test split.","caption":"Figure 42: Final classification accuracy of APPLE-SAC on real (TactileMNISTRealSnap, test split; solid) and simulated tactile data (TactileMNISTSnap; light) for the full sim-to-real recipe and its ablations. Each ablation removes one component from the full recipe; without the CycleGAN, we render with Taxim during training and, since the real-image round trip requires the CycleGAN, evaluate without it on the raw real images. Bars show the mean over 5 seeds of the accuracy averaged between 2.0M and 2.5M environment steps; lines show one standard deviation. The dashed line marks chance level (10%).","image":"img/2506.06361.png","kind":"figure","image_w":1144,"image_h":483,"pwc":false}
 
1
+ {"id":"2506.06361","title":"Tactile MNIST: Benchmarking Active Tactile Perception","authors":"Tim Schneider; Guillaume Duret; Cristiana de Farias; Roberto Calandra; Liming Chen; Jan Peters","openreview_id":"3yMkqerEOt","presentation":"spotlight","track":"evaluations_datasets","verdict":"borderline","background":"Tactile sensors such as the GelSight Mini give rich but very local information, so an agent has to choose where to touch and combine many touches over time to recognise an object (active tactile perception). The field has no standard benchmark. The paper introduces ap_gym (a Gymnasium-compatible active perception framework) and the Tactile MNIST Benchmark Suite (TMBS), which contains simulated tactile tasks, 13,580 3D-printable MNIST digit meshes, and a real dataset of 153,600 GelSight touches on 600 3D-printed digits. A CycleGAN trained on the real data renders realistic tactile images in simulation. Whether policies trained in simulation work on real tactile images is a key question for this kind of benchmark.","method":["The task is TactileMNIST classification with 16 touches per episode. Training uses TactileMNISTSnap, where each step the commanded sensor position snaps to the closest of 16 uniformly sampled positions. Real evaluation uses TactileMNISTRealSnap, which replays recorded real touches from the test split: the commanded position snaps to the closest of the next 16 prerecorded touches of the current digit. The agent is APPLE-SAC (a transformer actor-critic with a vision encoder), trained for 2.5M environment steps on 5 seeds per configuration.","The full sim-to-real recipe has four parts: (1) during training, simulated depth maps are rendered with the CycleGAN depth-to-RGB generator; (2) at test time, each real tactile image is passed through a CycleGAN round trip (RGB-to-depth, then depth-to-RGB); (3) during training, a background depth overlay adds the round-trip depth estimate of a random no-contact real touch to each simulated depth map; (4) contact geometry variation erodes or dilates and blurs the simulated indentation. Each ablation removes one part. 'No CycleGAN' renders with Taxim during training and evaluates on raw real images. The metric is accuracy after the last touch, averaged over evaluations between 2.0M and 2.5M steps, reported as mean ± std over the 5 seeds."],"claim":"With the full sim-to-real recipe, APPLE-SAC trained only in simulation reaches 65.7 ± 4.3% final classification accuracy on real TactileMNIST test digits. Removing the CycleGAN gives 12.4 ± 1.7%, removing the real-image round trip gives 19.0 ± 2.1%, and removing the background depth overlay gives 31.9 ± 7.7%. Simulated accuracy stays near or above 68% in every configuration, so simulated accuracy does not predict transfer.","source":"Figure 42, page 44 (numbers in the Appendix H text, pages 44-45)","metric":"Final-step classification accuracy (%) on TactileMNISTRealSnap test split (higher is better), averaged over evaluations between 2.0M and 2.5M env steps; ± is std over 5 seeds. Chance is 10%.","comparison":"Full recipe vs. four single-component ablations (no CycleGAN / no real-image round trip / no background depth overlay / no contact geometry variation), all APPLE-SAC, same training budget; real vs. simulated accuracy for each.","falsification":"The full recipe's real-data accuracy falls to or below the 'no real-image round trip' or 'no background depth overlay' ablation (or stays within about one std of either), or the full recipe does not clearly beat chance (10%) on the real test split.","caption":"Figure 42: Final classification accuracy of APPLE-SAC on real (TactileMNISTRealSnap, test split; solid) and simulated tactile data (TactileMNISTSnap; light) for the full sim-to-real recipe and its ablations. Each ablation removes one component from the full recipe; without the CycleGAN, we render with Taxim during training and, since the real-image round trip requires the CycleGAN, evaluate without it on the raw real images. Bars show the mean over 5 seeds of the accuracy averaged between 2.0M and 2.5M environment steps; lines show one standard deviation. The dashed line marks chance level (10%).","image":"img/2506.06361.png","kind":"figure","image_w":1144,"image_h":483,"pwc":true}
site/p/2507.14777.json CHANGED
@@ -1 +1 @@
1
- {"id":"2507.14777","title":"Understanding the Interplay between Memorization and Learning in Large Language Models","authors":"Bishwamittra Ghosh; Soumi Das; Qinyuan Wu; Mohammad Aflah Khan; Deepak Garg; Krishna Gummadi; Evimaria Terzi","openreview_id":"oG5alWKa3Z","presentation":"poster","track":"evaluations_datasets","verdict":"borderline","background":"Memorization in LLMs is usually measured by how well a model recalls a training string, which mixes up rote memorization with recall that comes from having learned the language (for example, counting to 1000). The paper proposes 'contextual memorization': a training string counts as memorized only once its training loss falls below the lowest loss it would reach if it were left out of training. It studies this measure in a controlled setting where models are fine-tuned on strings sampled from probabilistic context-free grammars, so string frequencies and the language entropy are known exactly. It uses the measure to motivate a training scheme that reduces memorization without hurting learning.","method":["Mistral-7B is fine-tuned for 50 epochs (batch size 8, learning rate 5e-6, linear schedule with 0.05 warm-up) on strings sampled from a formal language (L2, L4 or L9, defined in Appendix F) and evaluated on 1024 held-out test strings from the same language. The 'optimal learning epoch' is the epoch with the lowest mean test cross-entropy loss, and all metrics are read at that epoch. Contextual memorization of a training string s is (min-over-epochs loss of s when s is not trained on − training loss of s) / that minimum, clipped at 0. The leave-one-out loss is approximated by the lowest test loss of test strings in the same frequency group. 'Weighted memorization' is the mean of this score over training strings.","Memorization-aware training (MAT): at each epoch, compute x = the fraction of test strings whose loss went up since the previous epoch, then drop the x fraction of training strings with the lowest training loss for the next epoch. Dropped strings come back if their loss rises again. Baselines are base training (a fixed training set every epoch) and deduplication (each distinct string kept once). MAT* is a more expensive variant that drops strings based on contextual loss. Results are averaged over multiple runs (three runs according to Section 3)."],"claim":"When Mistral-7B is fine-tuned on formal language L2, MAT gives lower weighted contextual memorization at the optimal-learning epoch than base training (about 0.02 vs. 0.21) and lower test-loss std (about 0.090 vs. 0.094), with mean test loss no worse (about 0.177 vs. 0.188). MAT's memorization is also below base training on L4 with Mistral-7B (about 0.043 vs. 0.081) and on L9, where the per-seed figures show Pythia-1B was used (about 0.055 vs. 0.14). On every language MAT's memorization is also below deduplication's (L2: about 0.02 vs. 0.085).","source":"Figure 61, page 64 (appendix J.1); summary version for L2 in Figure 9, page 10","metric":"Weighted contextual memorization at the optimal-learning epoch (lower is better). Also reported: std and mean of test cross-entropy loss over 1024 test strings at that epoch (lower is better), and cumulative number of training strings. Points are averages over runs; no error bars or ± values are reported.","comparison":"MAT vs. base training and deduplication (and MAT*), same model, data and hyperparameters, on languages L2, L4 and L9","falsification":"Re-running base training and MAT on L2 with the released code and several seeds, MAT's weighted contextual memorization at the optimal epoch is not clearly lower than base training's (for example, within seed variability, or not at least halved), or MAT's mean test loss is clearly higher than base training's.","caption":"Figure 61: Summary results on learning, memorization, and training cost efficiency on different formal languages, averaged across multiple runs. Lower values are better in all dimensions.","image":"img/2507.14777.png","kind":"figure","image_w":1144,"image_h":1027,"pwc":false}
 
1
+ {"id":"2507.14777","title":"Understanding the Interplay between Memorization and Learning in Large Language Models","authors":"Bishwamittra Ghosh; Soumi Das; Qinyuan Wu; Mohammad Aflah Khan; Deepak Garg; Krishna Gummadi; Evimaria Terzi","openreview_id":"oG5alWKa3Z","presentation":"poster","track":"evaluations_datasets","verdict":"borderline","background":"Memorization in LLMs is usually measured by how well a model recalls a training string, which mixes up rote memorization with recall that comes from having learned the language (for example, counting to 1000). The paper proposes 'contextual memorization': a training string counts as memorized only once its training loss falls below the lowest loss it would reach if it were left out of training. It studies this measure in a controlled setting where models are fine-tuned on strings sampled from probabilistic context-free grammars, so string frequencies and the language entropy are known exactly. It uses the measure to motivate a training scheme that reduces memorization without hurting learning.","method":["Mistral-7B is fine-tuned for 50 epochs (batch size 8, learning rate 5e-6, linear schedule with 0.05 warm-up) on strings sampled from a formal language (L2, L4 or L9, defined in Appendix F) and evaluated on 1024 held-out test strings from the same language. The 'optimal learning epoch' is the epoch with the lowest mean test cross-entropy loss, and all metrics are read at that epoch. Contextual memorization of a training string s is (min-over-epochs loss of s when s is not trained on − training loss of s) / that minimum, clipped at 0. The leave-one-out loss is approximated by the lowest test loss of test strings in the same frequency group. 'Weighted memorization' is the mean of this score over training strings.","Memorization-aware training (MAT): at each epoch, compute x = the fraction of test strings whose loss went up since the previous epoch, then drop the x fraction of training strings with the lowest training loss for the next epoch. Dropped strings come back if their loss rises again. Baselines are base training (a fixed training set every epoch) and deduplication (each distinct string kept once). MAT* is a more expensive variant that drops strings based on contextual loss. Results are averaged over multiple runs (three runs according to Section 3)."],"claim":"When Mistral-7B is fine-tuned on formal language L2, MAT gives lower weighted contextual memorization at the optimal-learning epoch than base training (about 0.02 vs. 0.21) and lower test-loss std (about 0.090 vs. 0.094), with mean test loss no worse (about 0.177 vs. 0.188). MAT's memorization is also below base training on L4 with Mistral-7B (about 0.043 vs. 0.081) and on L9, where the per-seed figures show Pythia-1B was used (about 0.055 vs. 0.14). On every language MAT's memorization is also below deduplication's (L2: about 0.02 vs. 0.085).","source":"Figure 61, page 64 (appendix J.1); summary version for L2 in Figure 9, page 10","metric":"Weighted contextual memorization at the optimal-learning epoch (lower is better). Also reported: std and mean of test cross-entropy loss over 1024 test strings at that epoch (lower is better), and cumulative number of training strings. Points are averages over runs; no error bars or ± values are reported.","comparison":"MAT vs. base training and deduplication (and MAT*), same model, data and hyperparameters, on languages L2, L4 and L9","falsification":"Re-running base training and MAT on L2 with the released code and several seeds, MAT's weighted contextual memorization at the optimal epoch is not clearly lower than base training's (for example, within seed variability, or not at least halved), or MAT's mean test loss is clearly higher than base training's.","caption":"Figure 61: Summary results on learning, memorization, and training cost efficiency on different formal languages, averaged across multiple runs. Lower values are better in all dimensions.","image":"img/2507.14777.png","kind":"figure","image_w":1144,"image_h":1027,"pwc":true}
site/p/2508.13530.json CHANGED
@@ -1 +1 @@
1
- {"id":"2508.13530","title":"CrafterDojo: A Suite of Foundation Models for Building Open-Ended Embodied Agents in Crafter","authors":"Junyeong Park; Hyeonseo Cho; Sungjin Ahn","openreview_id":"Q7Y8eh12e2","presentation":"poster","track":"evaluations_datasets","verdict":"include","background":"Minecraft is the main testbed for open-ended embodied agents because foundation models trained on internet-scale data exist for it: VPT (behavior prior), MineCLIP (video-language grounding) and Steve-1 (instruction following). Minecraft is slow, crash-prone and hard to modify. Crafter is a lightweight 2D, pure-Python game with similar mechanics (procedural maps, resource collection, crafting, survival), but it has no comparable foundation models, so Crafter research has mostly used reinforcement learning from scratch on its 22 fixed achievements. CrafterDojo builds Crafter versions of the three Minecraft models from synthetic data so that hierarchical, instruction-following agents can be prototyped cheaply.","method":["A PPO-RNN expert trained for 10B steps in Craftax-Classic (symbolic) generates CrafterPlay, 20,000 pixel-observation episodes. C-VPT (ResNet + TransformerXL + policy head, VPT architecture) is behavior-cloned on CrafterPlay. A rule-based captioner (15 rules, 61 templates, each expanded with 40 LLM paraphrases) labels 6-frame clips to form CrafterCaption, on which C-CLIP (MineCLIP-style video/text encoders initialized from CLIP) is trained contrastively. C-Steve-1 fine-tunes C-VPT with LoRA as a goal-conditioned policy (goal embedding added after the TransformerXL) using event-based packed hindsight relabeling with C-CLIP video embeddings as goals, plus a CVAE prior mapping C-CLIP text embeddings to goal embeddings. At inference it uses classifier-free guidance (lambda = 1.5).","For long-horizon tasks, PPO-Steve is a hierarchical agent: a PPO high-level planner chooses one of the 61 template captions every 10 steps and passes it to frozen C-Steve-1 as the low-level controller. Baselines are PPO-RNN trained from scratch on pixels, C-VPT-RL (C-VPT base fine-tuned with PPO + LoRA r = 16, alpha = 16, with a 0.1-weighted KL to the pretrained policy), and C-Steve-1 prompted with a single instruction for the whole task (no training, shown as a dashed line). Tasks have sparse rewards given only for completing sub-tasks in the right order, with a 1,000-step episode limit: T1 eat plant (simplified env without health decay or mobs), T2 place plant then place table, T3 make wood pickaxe then obtain sapling, and T4 obtain coal, craft wood pickaxe, then obtain stone. Each agent is trained for 1M environment steps with 3 seeds."],"claim":"On the long-horizon Crafter task T4 (obtain coal -> make wood pickaxe -> obtain stone), after 1M environment steps the hierarchical PPO-Steve agent (PPO planner over C-Steve-1) reaches a success rate of about 0.6, compared with about 0.17 for PPO fine-tuning of C-VPT (C-VPT-RL), about 0 for PPO-RNN from scratch, and 0.33 for single-instruction C-Steve-1. On T3, PPO-Steve reaches about 0.96 vs. about 0.77 for C-VPT-RL.","source":"Figure 3, page 6","metric":"Episode success rate during training (higher is better) vs. environment steps; lines are means over 3 seeds and shaded bands are standard deviations across seeds. C-Steve-1 single-instruction is a fixed evaluation shown as a dashed line. Values are read off the plot and approximate.","comparison":"PPO-Steve vs. C-VPT-RL, PPO-RNN and single-instruction C-Steve-1, all on the same tasks with the same 1M-step training budget","falsification":"Retraining PPO-Steve and C-VPT-RL on T4 for 1M environment steps with 3 seeds, PPO-Steve's final success rate is not clearly above C-VPT-RL's (e.g. the gap is under ~0.2 or the seed bands overlap), or PPO-Steve does not beat single-instruction C-Steve-1 on T4.","caption":"Figure 3: Training performance of four models on four long-horizon tasks. The first task requires the agents to obtain sapling, place plant, and then eat plant. The second task requires the agents to place plant and then place table. The third task requires the agents to make wood pickaxe and then obtain sapling. The last task requires the agents to obtain coal, craft wood pickaxe, and then obtain stone. For all tasks, the agents receive a reward for completing the given tasks in the right order. We report training performance with success rates and standard deviations among three seeds for each agent.","image":"img/2508.13530.png","kind":"figure","image_w":1430,"image_h":550,"pwc":false}
 
1
+ {"id":"2508.13530","title":"CrafterDojo: A Suite of Foundation Models for Building Open-Ended Embodied Agents in Crafter","authors":"Junyeong Park; Hyeonseo Cho; Sungjin Ahn","openreview_id":"Q7Y8eh12e2","presentation":"poster","track":"evaluations_datasets","verdict":"include","background":"Minecraft is the main testbed for open-ended embodied agents because foundation models trained on internet-scale data exist for it: VPT (behavior prior), MineCLIP (video-language grounding) and Steve-1 (instruction following). Minecraft is slow, crash-prone and hard to modify. Crafter is a lightweight 2D, pure-Python game with similar mechanics (procedural maps, resource collection, crafting, survival), but it has no comparable foundation models, so Crafter research has mostly used reinforcement learning from scratch on its 22 fixed achievements. CrafterDojo builds Crafter versions of the three Minecraft models from synthetic data so that hierarchical, instruction-following agents can be prototyped cheaply.","method":["A PPO-RNN expert trained for 10B steps in Craftax-Classic (symbolic) generates CrafterPlay, 20,000 pixel-observation episodes. C-VPT (ResNet + TransformerXL + policy head, VPT architecture) is behavior-cloned on CrafterPlay. A rule-based captioner (15 rules, 61 templates, each expanded with 40 LLM paraphrases) labels 6-frame clips to form CrafterCaption, on which C-CLIP (MineCLIP-style video/text encoders initialized from CLIP) is trained contrastively. C-Steve-1 fine-tunes C-VPT with LoRA as a goal-conditioned policy (goal embedding added after the TransformerXL) using event-based packed hindsight relabeling with C-CLIP video embeddings as goals, plus a CVAE prior mapping C-CLIP text embeddings to goal embeddings. At inference it uses classifier-free guidance (lambda = 1.5).","For long-horizon tasks, PPO-Steve is a hierarchical agent: a PPO high-level planner chooses one of the 61 template captions every 10 steps and passes it to frozen C-Steve-1 as the low-level controller. Baselines are PPO-RNN trained from scratch on pixels, C-VPT-RL (C-VPT base fine-tuned with PPO + LoRA r = 16, alpha = 16, with a 0.1-weighted KL to the pretrained policy), and C-Steve-1 prompted with a single instruction for the whole task (no training, shown as a dashed line). Tasks have sparse rewards given only for completing sub-tasks in the right order, with a 1,000-step episode limit: T1 eat plant (simplified env without health decay or mobs), T2 place plant then place table, T3 make wood pickaxe then obtain sapling, and T4 obtain coal, craft wood pickaxe, then obtain stone. Each agent is trained for 1M environment steps with 3 seeds."],"claim":"On the long-horizon Crafter task T4 (obtain coal -> make wood pickaxe -> obtain stone), after 1M environment steps the hierarchical PPO-Steve agent (PPO planner over C-Steve-1) reaches a success rate of about 0.6, compared with about 0.17 for PPO fine-tuning of C-VPT (C-VPT-RL), about 0 for PPO-RNN from scratch, and 0.33 for single-instruction C-Steve-1. On T3, PPO-Steve reaches about 0.96 vs. about 0.77 for C-VPT-RL.","source":"Figure 3, page 6","metric":"Episode success rate during training (higher is better) vs. environment steps; lines are means over 3 seeds and shaded bands are standard deviations across seeds. C-Steve-1 single-instruction is a fixed evaluation shown as a dashed line. Values are read off the plot and approximate.","comparison":"PPO-Steve vs. C-VPT-RL, PPO-RNN and single-instruction C-Steve-1, all on the same tasks with the same 1M-step training budget","falsification":"Retraining PPO-Steve and C-VPT-RL on T4 for 1M environment steps with 3 seeds, PPO-Steve's final success rate is not clearly above C-VPT-RL's (e.g. the gap is under ~0.2 or the seed bands overlap), or PPO-Steve does not beat single-instruction C-Steve-1 on T4.","caption":"Figure 3: Training performance of four models on four long-horizon tasks. The first task requires the agents to obtain sapling, place plant, and then eat plant. The second task requires the agents to place plant and then place table. The third task requires the agents to make wood pickaxe and then obtain sapling. The last task requires the agents to obtain coal, craft wood pickaxe, and then obtain stone. For all tasks, the agents receive a reward for completing the given tasks in the right order. We report training performance with success rates and standard deviations among three seeds for each agent.","image":"img/2508.13530.png","kind":"figure","image_w":1430,"image_h":550,"pwc":true}
site/p/2508.18741.json CHANGED
@@ -1 +1 @@
1
- {"id":"2508.18741","title":"Statistical Complexity of Soft Bellman Residual Minimization","authors":"Enoch H. Kang; Kyoungseok Jang","openreview_id":"QJrpoS9mBV","presentation":"poster","track":"main","verdict":"borderline","background":"Offline RL and inverse RL fit value functions or rewards from a fixed batch of logged transitions. Bellman residual minimization (BRM) fits a Q-function by minimizing the mean-squared Bellman error (MSBE). With entropy regularization ('soft' BRM) and a bi-conjugate reformulation, BRM becomes a min-max problem. Prior work showed that stochastic gradient descent-ascent (SGDA) converges globally on this problem under Polyak-Lojasiewicz (PL) geometry. How many offline samples SGDA needs was still open: earlier BRM sample-complexity results either assume an exact (oracle) solver or give O(n^-1/2) rates.","method":["Soft BRM is written as min over Q_w, max over an auxiliary function zeta_v, of the empirical average of a per-sample saddle loss f(w, v; s, a, s') = (TD loss of Q_w) - beta^2 (V_Q(s') - zeta_v(s, a))^2 (Eq. 3-5). It is solved by minibatch SGDA (Algorithm 1): a descent step on w and an ascent step on v with dual stepsize ratio kappa_v. Under PL, quadratic-growth and smoothness conditions, the paper proves O(1/n) on-average primal argument stability (Theorem 2) and an O(1/n + 1/T) excess population risk / excess MSBE for the SGDA output (Theorem 5).","Numerical check (Section 4, Appendix E): a soft MDP with |S| = 5, |A| = 3, beta = 0.9, eta = 1, rewards ~ Unif[0,1], P(.|s,a) = 1/2 delta_s + 1/2 Dirichlet(1), and a uniform behavior policy. Q_w (one-hot state -> |A| outputs) and zeta_v (one-hot state-action -> scalar) are two-layer softplus networks of width H in {8, 16, 32} with NTK parameterization. SGDA uses B = 32, eta_t = 400/(2000 + t), kappa_v = 2 and T = 40,000. The data are n in {100, ..., 6400} transitions, either i.i.d. from the stationary distribution mu or from a single stationary trajectory. The excess population MSBE of the output is computed exactly from P; since Q* is representable, it equals the excess risk. Each (n, H) setting uses 50 repetitions, and the slope is a least-squares fit of log(mean) on log n over n >= 200, with a 95% bootstrap interval over repetitions."],"claim":"On the 5-state, 3-action soft MDP with i.i.d. data and width-16 networks, the excess population MSBE of the SGDA output decays as n^-1: the fitted log-log slope over n = 200..6400 is -1.00 (95% bootstrap interval [-1.05, -0.95]), n x excess MSBE stays between 0.385 and 0.443, and every width and data regime has a slope interval that excludes -1/2.","source":"Table 6, page 28 (Appendix E); also Figure 1, page 10","metric":"n x excess population MSBE (mean +/- standard error over 50 repetitions) and the fitted slope of log(excess MSBE) vs log n over n >= 200, with a 95% bootstrap interval over repetitions","comparison":"Observed decay rate of the SGDA output's excess MSBE vs the theorem's predicted n^-1 rate and the n^-1/2 rate of prior bounds, across widths H = 8/16/32, i.i.d. vs single-trajectory data, and against the exact empirical minimizer","falsification":"Rerunning the stated setup gives a fitted slope (n >= 200) whose 95% bootstrap interval excludes -1 or includes -1/2 for H = 16 i.i.d. data, or n x excess MSBE grows steadily with n (for example, a clear upward trend from 0.44 at n = 200 to well above 1 at n = 6400).","caption":"Table 6: n times the excess population MSBE of the SGDA output (mean ± standard error), and of the exact empirical minimizer on the datasets of the i.i.d. H = 16 runs.","image":"img/2508.18741.png","kind":"table","image_w":1144,"image_h":455,"pwc":false}
 
1
+ {"id":"2508.18741","title":"Statistical Complexity of Soft Bellman Residual Minimization","authors":"Enoch H. Kang; Kyoungseok Jang","openreview_id":"QJrpoS9mBV","presentation":"poster","track":"main","verdict":"borderline","background":"Offline RL and inverse RL fit value functions or rewards from a fixed batch of logged transitions. Bellman residual minimization (BRM) fits a Q-function by minimizing the mean-squared Bellman error (MSBE). With entropy regularization ('soft' BRM) and a bi-conjugate reformulation, BRM becomes a min-max problem. Prior work showed that stochastic gradient descent-ascent (SGDA) converges globally on this problem under Polyak-Lojasiewicz (PL) geometry. How many offline samples SGDA needs was still open: earlier BRM sample-complexity results either assume an exact (oracle) solver or give O(n^-1/2) rates.","method":["Soft BRM is written as min over Q_w, max over an auxiliary function zeta_v, of the empirical average of a per-sample saddle loss f(w, v; s, a, s') = (TD loss of Q_w) - beta^2 (V_Q(s') - zeta_v(s, a))^2 (Eq. 3-5). It is solved by minibatch SGDA (Algorithm 1): a descent step on w and an ascent step on v with dual stepsize ratio kappa_v. Under PL, quadratic-growth and smoothness conditions, the paper proves O(1/n) on-average primal argument stability (Theorem 2) and an O(1/n + 1/T) excess population risk / excess MSBE for the SGDA output (Theorem 5).","Numerical check (Section 4, Appendix E): a soft MDP with |S| = 5, |A| = 3, beta = 0.9, eta = 1, rewards ~ Unif[0,1], P(.|s,a) = 1/2 delta_s + 1/2 Dirichlet(1), and a uniform behavior policy. Q_w (one-hot state -> |A| outputs) and zeta_v (one-hot state-action -> scalar) are two-layer softplus networks of width H in {8, 16, 32} with NTK parameterization. SGDA uses B = 32, eta_t = 400/(2000 + t), kappa_v = 2 and T = 40,000. The data are n in {100, ..., 6400} transitions, either i.i.d. from the stationary distribution mu or from a single stationary trajectory. The excess population MSBE of the output is computed exactly from P; since Q* is representable, it equals the excess risk. Each (n, H) setting uses 50 repetitions, and the slope is a least-squares fit of log(mean) on log n over n >= 200, with a 95% bootstrap interval over repetitions."],"claim":"On the 5-state, 3-action soft MDP with i.i.d. data and width-16 networks, the excess population MSBE of the SGDA output decays as n^-1: the fitted log-log slope over n = 200..6400 is -1.00 (95% bootstrap interval [-1.05, -0.95]), n x excess MSBE stays between 0.385 and 0.443, and every width and data regime has a slope interval that excludes -1/2.","source":"Table 6, page 28 (Appendix E); also Figure 1, page 10","metric":"n x excess population MSBE (mean +/- standard error over 50 repetitions) and the fitted slope of log(excess MSBE) vs log n over n >= 200, with a 95% bootstrap interval over repetitions","comparison":"Observed decay rate of the SGDA output's excess MSBE vs the theorem's predicted n^-1 rate and the n^-1/2 rate of prior bounds, across widths H = 8/16/32, i.i.d. vs single-trajectory data, and against the exact empirical minimizer","falsification":"Rerunning the stated setup gives a fitted slope (n >= 200) whose 95% bootstrap interval excludes -1 or includes -1/2 for H = 16 i.i.d. data, or n x excess MSBE grows steadily with n (for example, a clear upward trend from 0.44 at n = 200 to well above 1 at n = 6400).","caption":"Table 6: n times the excess population MSBE of the SGDA output (mean ± standard error), and of the exact empirical minimizer on the datasets of the i.i.d. H = 16 runs.","image":"img/2508.18741.png","kind":"table","image_w":1144,"image_h":455,"pwc":true}
site/p/2510.03016.json CHANGED
@@ -1 +1 @@
1
- {"id":"2510.03016","title":"DELTA: Robustly Training Label-Conditional Diffusion Models with Weak Annotations","authors":"Dong-Dong Wu; Jiacheng Cui; Wei Wang; Zhiqiang Shen; Masashi Sugiyama","openreview_id":"1fJGfUpQRC","presentation":"poster","track":"main","verdict":"include","background":"Label-conditional diffusion models are trained on (image, class label) pairs and assume the labels are clean. In practice labels are often noisy (wrong), ambiguous (a candidate set that contains the true class), or missing for most of the data. Training a conditional diffusion model directly on such weak annotations produces samples that do not match their class, collapse to a few modes per class, or degrade in quality. Earlier robust methods each handle only one kind of weak label and need extra information such as a known noise transition matrix or confidence scores.","method":["DELTA treats the true label as a latent variable and trains with an EM-style objective. In the generative term, the network's conditional score outputs for all classes are mixed with weights equal to the estimated posterior p(y | x_t, z) given the weak annotation z, and this mixture is regressed onto the standard denoising score-matching target (posterior-weighted score matching). The posterior is computed by the diffusion model itself (an EMA copy) as a diffusion classifier: Bayes' rule over class-conditional ELBOs, evaluated only on a short window of noise levels centred at the median of the EDM log-normal weight (log sigma = -1.2, window length 6.4) to make it cheap enough to run during training. For ambiguous labels the posterior is renormalized over the candidate set. A classification term (posterior-weighted cross-entropy) refines the model's posterior.","Vanilla is the internal ablation: the same joint objective and training setup, but the generative term trains directly on the weak pairs (x, z) without posterior reweighting. Both use EDM with a DDPM++ U-Net (about 55.7M parameters), trained from scratch for 200k iterations with Adam (lr 1e-3) and batch size 64 on CIFAR-10. In the 'Random' ambiguous setting each non-target label is added to the candidate set with probability 0.5. Samples are generated per class and scored with the DLSM evaluation code: FID, IS, Density and Coverage on the pooled samples, plus class-wise CW-FID, CW-Density and CW-Coverage averaged over classes."],"claim":"On CIFAR-10 with random ambiguous labels (each wrong class added to the candidate set with probability 0.5), DELTA reaches CW-FID 10.65 and FID 2.26, versus 27.18 and 7.76 for the Vanilla ablation trained identically without posterior-weighted score matching (clean-label upper bound: 9.83 and 2.05).","source":"Table 1, page 6","metric":"Class-wise FID (CW-FID, intra-class FID averaged over the 10 classes) and unconditional FID, lower is better; single runs, no error bars reported","comparison":"DELTA vs. Vanilla (same architecture, data, weak labels and 200k-iteration budget; only the generative term differs), with a model trained on clean labels as reference","falsification":"Training both models on CIFAR-10 with random ambiguous labels (q = 0.5) under the same budget, DELTA's CW-FID is not clearly below Vanilla's (e.g. less than about 1.5x lower, or the gap is within the spread of a few seeds), or DELTA's FID is not clearly below Vanilla's.","caption":"Table 1: Generative results under various weak annotations. The gray-shaded 'Clean' column represents the best performance trained on the fully clean dataset. 'uncond'/'cond' denote unconditional/conditional metrics. Bold numbers indicate better performance.","image":"img/2510.03016.png","kind":"table","image_w":1150,"image_h":469,"pwc":false}
 
1
+ {"id":"2510.03016","title":"DELTA: Robustly Training Label-Conditional Diffusion Models with Weak Annotations","authors":"Dong-Dong Wu; Jiacheng Cui; Wei Wang; Zhiqiang Shen; Masashi Sugiyama","openreview_id":"1fJGfUpQRC","presentation":"poster","track":"main","verdict":"include","background":"Label-conditional diffusion models are trained on (image, class label) pairs and assume the labels are clean. In practice labels are often noisy (wrong), ambiguous (a candidate set that contains the true class), or missing for most of the data. Training a conditional diffusion model directly on such weak annotations produces samples that do not match their class, collapse to a few modes per class, or degrade in quality. Earlier robust methods each handle only one kind of weak label and need extra information such as a known noise transition matrix or confidence scores.","method":["DELTA treats the true label as a latent variable and trains with an EM-style objective. In the generative term, the network's conditional score outputs for all classes are mixed with weights equal to the estimated posterior p(y | x_t, z) given the weak annotation z, and this mixture is regressed onto the standard denoising score-matching target (posterior-weighted score matching). The posterior is computed by the diffusion model itself (an EMA copy) as a diffusion classifier: Bayes' rule over class-conditional ELBOs, evaluated only on a short window of noise levels centred at the median of the EDM log-normal weight (log sigma = -1.2, window length 6.4) to make it cheap enough to run during training. For ambiguous labels the posterior is renormalized over the candidate set. A classification term (posterior-weighted cross-entropy) refines the model's posterior.","Vanilla is the internal ablation: the same joint objective and training setup, but the generative term trains directly on the weak pairs (x, z) without posterior reweighting. Both use EDM with a DDPM++ U-Net (about 55.7M parameters), trained from scratch for 200k iterations with Adam (lr 1e-3) and batch size 64 on CIFAR-10. In the 'Random' ambiguous setting each non-target label is added to the candidate set with probability 0.5. Samples are generated per class and scored with the DLSM evaluation code: FID, IS, Density and Coverage on the pooled samples, plus class-wise CW-FID, CW-Density and CW-Coverage averaged over classes."],"claim":"On CIFAR-10 with random ambiguous labels (each wrong class added to the candidate set with probability 0.5), DELTA reaches CW-FID 10.65 and FID 2.26, versus 27.18 and 7.76 for the Vanilla ablation trained identically without posterior-weighted score matching (clean-label upper bound: 9.83 and 2.05).","source":"Table 1, page 6","metric":"Class-wise FID (CW-FID, intra-class FID averaged over the 10 classes) and unconditional FID, lower is better; single runs, no error bars reported","comparison":"DELTA vs. Vanilla (same architecture, data, weak labels and 200k-iteration budget; only the generative term differs), with a model trained on clean labels as reference","falsification":"Training both models on CIFAR-10 with random ambiguous labels (q = 0.5) under the same budget, DELTA's CW-FID is not clearly below Vanilla's (e.g. less than about 1.5x lower, or the gap is within the spread of a few seeds), or DELTA's FID is not clearly below Vanilla's.","caption":"Table 1: Generative results under various weak annotations. The gray-shaded 'Clean' column represents the best performance trained on the fully clean dataset. 'uncond'/'cond' denote unconditional/conditional metrics. Bold numbers indicate better performance.","image":"img/2510.03016.png","kind":"table","image_w":1150,"image_h":469,"pwc":true}
site/p/2510.05849.json CHANGED
@@ -1 +1 @@
1
- {"id":"2510.05849","title":"ESS-Flow: training-free guidance as Bayesian inference in source space","authors":"Adhithyan Kalaivanan; Zheng Zhao; Jens Sjölund; Fredrik Lindsten","openreview_id":"qokTHEVYpp","presentation":"poster","track":"main","verdict":"include","background":"Training-free guidance steers a pretrained diffusion or flow model toward samples with a target property without retraining it. Most such methods (DPS, D-Flow, PnP-Flow, DAPS) need gradients of the reward or likelihood, and sometimes of the generative map too. In scientific design problems those gradients are often unavailable, for example when the generator rounds discrete atom types or the reward comes from an external simulator. Gradient-based guidance can also get stuck in disconnected modes of the data distribution, such as different atomic compositions.","method":["ESS-Flow treats a flow model with a Gaussian source as a deterministic transport map x = T(z), z ~ N(0, I), and samples the target pi(z) proportional to g(T(z)) N(z; 0, I) directly in source space using elliptical slice sampling (ESS). Each MCMC step draws nu ~ N(0, I), proposes z' = z cos(theta) + nu sin(theta) along the ellipse, and accepts if log g(T(z')) exceeds log g(T(z)) + log u. On rejection it shrinks the angle bracket and tries again. The method only needs forward evaluations of T and g, with no gradients and no tuning.","For materials, the prior is FlowMM (Riemannian flow matching on MP-20). The atom types are rounded to discrete values, so the generator is not differentiable. The potential is g = exp(-(P - y)^2 / 2 sigma^2), where P is predicted by ALIGNN models trained on JARVIS-DFT. Targets are bulk modulus 300 GPa (4 atoms, sigma = 10 GPa) and shear modulus 200 GPa (4 atoms, sigma = 10 GPa), both above the prior's 99th percentile. ESS-Flow uses 50 ODE steps and 1000 parallel chains run for 201 MCMC steps, keeping the final state of each chain (200 burn-in). The baselines are D-Flow, PnP-Flow, and DAPS. They use soft atom embeddings (D-Flow, PnP-Flow) or Metropolis-Hastings bit flips for atom types (DAPS), with hyperparameters given in Appendix Table 7. The metric is the absolute error between the predicted property and the target over the generated samples."],"claim":"When guiding FlowMM toward bulk modulus 300 GPa and shear modulus 200 GPa with ALIGNN-predicted properties, ESS-Flow reaches a mean absolute error of 8.99 ± 6.69 GPa and 10.53 ± 9.21 GPa. The best gradient-based baselines reach 39.14 ± 26.47 (DAPS) and 75.48 ± 33.45 (PnP-Flow).","source":"Table 2, page 7","metric":"Mean absolute error between the ALIGNN-predicted property of generated materials and the target, in GPa (lower is better). ± is the standard deviation across generated samples (1000 samples, one run), not across seeds.","comparison":"ESS-Flow vs. D-Flow, PnP-Flow, and DAPS. All use the same FlowMM prior, the same ALIGNN predictor, and the same target and atom count.","falsification":"With the same FlowMM prior, ALIGNN predictor and settings, ESS-Flow's mean absolute error on bulk or shear modulus is not clearly lower than the best of DAPS or PnP-Flow. For example, the best baseline comes within ~2x of ESS-Flow, or ESS-Flow's bulk-modulus MAE exceeds ~20 GPa.","caption":"Table 2: Mean and standard deviation of absolute errors between the sample properties and targets, and of energy above hull values.","image":"img/2510.05849.png","kind":"table","image_w":1144,"image_h":250,"pwc":false}
 
1
+ {"id":"2510.05849","title":"ESS-Flow: training-free guidance as Bayesian inference in source space","authors":"Adhithyan Kalaivanan; Zheng Zhao; Jens Sjölund; Fredrik Lindsten","openreview_id":"qokTHEVYpp","presentation":"poster","track":"main","verdict":"include","background":"Training-free guidance steers a pretrained diffusion or flow model toward samples with a target property without retraining it. Most such methods (DPS, D-Flow, PnP-Flow, DAPS) need gradients of the reward or likelihood, and sometimes of the generative map too. In scientific design problems those gradients are often unavailable, for example when the generator rounds discrete atom types or the reward comes from an external simulator. Gradient-based guidance can also get stuck in disconnected modes of the data distribution, such as different atomic compositions.","method":["ESS-Flow treats a flow model with a Gaussian source as a deterministic transport map x = T(z), z ~ N(0, I), and samples the target pi(z) proportional to g(T(z)) N(z; 0, I) directly in source space using elliptical slice sampling (ESS). Each MCMC step draws nu ~ N(0, I), proposes z' = z cos(theta) + nu sin(theta) along the ellipse, and accepts if log g(T(z')) exceeds log g(T(z)) + log u. On rejection it shrinks the angle bracket and tries again. The method only needs forward evaluations of T and g, with no gradients and no tuning.","For materials, the prior is FlowMM (Riemannian flow matching on MP-20). The atom types are rounded to discrete values, so the generator is not differentiable. The potential is g = exp(-(P - y)^2 / 2 sigma^2), where P is predicted by ALIGNN models trained on JARVIS-DFT. Targets are bulk modulus 300 GPa (4 atoms, sigma = 10 GPa) and shear modulus 200 GPa (4 atoms, sigma = 10 GPa), both above the prior's 99th percentile. ESS-Flow uses 50 ODE steps and 1000 parallel chains run for 201 MCMC steps, keeping the final state of each chain (200 burn-in). The baselines are D-Flow, PnP-Flow, and DAPS. They use soft atom embeddings (D-Flow, PnP-Flow) or Metropolis-Hastings bit flips for atom types (DAPS), with hyperparameters given in Appendix Table 7. The metric is the absolute error between the predicted property and the target over the generated samples."],"claim":"When guiding FlowMM toward bulk modulus 300 GPa and shear modulus 200 GPa with ALIGNN-predicted properties, ESS-Flow reaches a mean absolute error of 8.99 ± 6.69 GPa and 10.53 ± 9.21 GPa. The best gradient-based baselines reach 39.14 ± 26.47 (DAPS) and 75.48 ± 33.45 (PnP-Flow).","source":"Table 2, page 7","metric":"Mean absolute error between the ALIGNN-predicted property of generated materials and the target, in GPa (lower is better). ± is the standard deviation across generated samples (1000 samples, one run), not across seeds.","comparison":"ESS-Flow vs. D-Flow, PnP-Flow, and DAPS. All use the same FlowMM prior, the same ALIGNN predictor, and the same target and atom count.","falsification":"With the same FlowMM prior, ALIGNN predictor and settings, ESS-Flow's mean absolute error on bulk or shear modulus is not clearly lower than the best of DAPS or PnP-Flow. For example, the best baseline comes within ~2x of ESS-Flow, or ESS-Flow's bulk-modulus MAE exceeds ~20 GPa.","caption":"Table 2: Mean and standard deviation of absolute errors between the sample properties and targets, and of energy above hull values.","image":"img/2510.05849.png","kind":"table","image_w":1144,"image_h":250,"pwc":true}
site/p/2510.26401.json CHANGED
@@ -1 +1 @@
1
- {"id":"2510.26401","title":"Outlier-Robust Multi-Output Gaussian Processes","authors":"Joshua Rooijakkers; Leiv Rønneberg; Francois-Xavier Briol; Jeremias Knoblauch; Matias Altamirano","openreview_id":"4tfLhUSLnF","presentation":"poster","track":"main","verdict":"include","background":"Multi-output Gaussian processes (MOGPs) model several correlated response variables jointly and have a closed-form posterior when the noise is Gaussian. That assumption makes them fragile: a single outlier in one output distorts predictions for that output and, through the cross-output correlation, for the other outputs too. Existing robust alternatives swap in Student-t or Laplace likelihoods, which break conjugacy and need variational approximations that are much slower, on top of the already cubic O(N^3 T^3) cost of MOGPs.","method":["R-MOGP replaces the likelihood with a generalised-Bayes posterior built on a weighted Fisher divergence, extending robust conjugate GPs (RCGPs) to multiple outputs. Each observation y_{i,t} gets an inverse-multiquadric weight w_t = beta_t (1 + ((y_t - gamma_t)/c_t)^2)^(-1/2), where the centre gamma_t is the conditional mean of output t given the other outputs at the same input (borrowing strength across outputs), c_t is the (1 - eps_t) quantile of |y_{i,t} - gamma_t|, and beta_t = sigma_t/sqrt(2). The posterior stays Gaussian and closed form (same cost as an MOGP). Hyperparameters (kernel, noise) are fit with a weighted leave-one-out CV objective in which each LOO term is weighted by (w_{i,t}/beta_t)^2.","On the UCI Energy Efficiency dataset (d = 8 inputs, T = 2 outputs: heating and cooling load, standardised; 75/25 split, N = 576 train) an ICM kernel with a squared-exponential base kernel is used. 10% of heating-load training values are contaminated: Uniform (add or subtract U(6,9), half each), Asymmetric (add U(6,9)), or Student-t (add t_3(0,5)) noise. The baseline MOGP uses the same kernel, shared noise sigma_t = sigma, prior mean set to the output's mean, and an unweighted LOO-CV objective. Metrics are test RMSE and NLPD over 10 random train/test splits per scenario."],"claim":"On UCI Energy Efficiency with 10% of heating-load observations contaminated, R-MOGP reaches test RMSE 0.17 ± 0.02 (uniform outliers) and 0.16 ± 0.02 (asymmetric outliers) versus 2.27 ± 0.29 and 2.18 ± 0.18 for a standard MOGP with the same kernel, while matching the MOGP without outliers (0.12 ± 0.01 for both).","source":"Table 2, page 9","metric":"Test RMSE (lower is better) and NLPD (lower is better) on standardised outputs; ± is the standard deviation over 10 random train/test splits","comparison":"R-MOGP (IMQ weights, w-LOO-CV) vs. standard MOGP (Gaussian likelihood, LOO-CV), same ICM kernel and data splits; t-MOGP and L-MOGP (variational GPflow) also reported","falsification":"Under uniform or asymmetric 10% outliers, R-MOGP's RMSE is not clearly below the MOGP's (e.g. not at least 2x lower, or within a few std of each other), or R-MOGP's RMSE in the no-outlier setting is clearly worse than the MOGP's (outside ~1 std of 0.12).","caption":"Table 2: Average and standard deviation of RMSE and NLPD on the Energy Efficiency dataset under different outlier scenarios (10 seeds per scenario). Best results are highlighted in bold. Figure 4: Average runtime per function evaluation on the Energy Efficiency dataset across 40 runs.","image":"img/2510.26401.png","kind":"table","image_w":1122,"image_h":666,"pwc":false}
 
1
+ {"id":"2510.26401","title":"Outlier-Robust Multi-Output Gaussian Processes","authors":"Joshua Rooijakkers; Leiv Rønneberg; Francois-Xavier Briol; Jeremias Knoblauch; Matias Altamirano","openreview_id":"4tfLhUSLnF","presentation":"poster","track":"main","verdict":"include","background":"Multi-output Gaussian processes (MOGPs) model several correlated response variables jointly and have a closed-form posterior when the noise is Gaussian. That assumption makes them fragile: a single outlier in one output distorts predictions for that output and, through the cross-output correlation, for the other outputs too. Existing robust alternatives swap in Student-t or Laplace likelihoods, which break conjugacy and need variational approximations that are much slower, on top of the already cubic O(N^3 T^3) cost of MOGPs.","method":["R-MOGP replaces the likelihood with a generalised-Bayes posterior built on a weighted Fisher divergence, extending robust conjugate GPs (RCGPs) to multiple outputs. Each observation y_{i,t} gets an inverse-multiquadric weight w_t = beta_t (1 + ((y_t - gamma_t)/c_t)^2)^(-1/2), where the centre gamma_t is the conditional mean of output t given the other outputs at the same input (borrowing strength across outputs), c_t is the (1 - eps_t) quantile of |y_{i,t} - gamma_t|, and beta_t = sigma_t/sqrt(2). The posterior stays Gaussian and closed form (same cost as an MOGP). Hyperparameters (kernel, noise) are fit with a weighted leave-one-out CV objective in which each LOO term is weighted by (w_{i,t}/beta_t)^2.","On the UCI Energy Efficiency dataset (d = 8 inputs, T = 2 outputs: heating and cooling load, standardised; 75/25 split, N = 576 train) an ICM kernel with a squared-exponential base kernel is used. 10% of heating-load training values are contaminated: Uniform (add or subtract U(6,9), half each), Asymmetric (add U(6,9)), or Student-t (add t_3(0,5)) noise. The baseline MOGP uses the same kernel, shared noise sigma_t = sigma, prior mean set to the output's mean, and an unweighted LOO-CV objective. Metrics are test RMSE and NLPD over 10 random train/test splits per scenario."],"claim":"On UCI Energy Efficiency with 10% of heating-load observations contaminated, R-MOGP reaches test RMSE 0.17 ± 0.02 (uniform outliers) and 0.16 ± 0.02 (asymmetric outliers) versus 2.27 ± 0.29 and 2.18 ± 0.18 for a standard MOGP with the same kernel, while matching the MOGP without outliers (0.12 ± 0.01 for both).","source":"Table 2, page 9","metric":"Test RMSE (lower is better) and NLPD (lower is better) on standardised outputs; ± is the standard deviation over 10 random train/test splits","comparison":"R-MOGP (IMQ weights, w-LOO-CV) vs. standard MOGP (Gaussian likelihood, LOO-CV), same ICM kernel and data splits; t-MOGP and L-MOGP (variational GPflow) also reported","falsification":"Under uniform or asymmetric 10% outliers, R-MOGP's RMSE is not clearly below the MOGP's (e.g. not at least 2x lower, or within a few std of each other), or R-MOGP's RMSE in the no-outlier setting is clearly worse than the MOGP's (outside ~1 std of 0.12).","caption":"Table 2: Average and standard deviation of RMSE and NLPD on the Energy Efficiency dataset under different outlier scenarios (10 seeds per scenario). Best results are highlighted in bold. Figure 4: Average runtime per function evaluation on the Energy Efficiency dataset across 40 runs.","image":"img/2510.26401.png","kind":"table","image_w":1122,"image_h":666,"pwc":true}
site/p/2511.11934.json CHANGED
@@ -1 +1 @@
1
- {"id":"2511.11934","title":"A Systematic Analysis of Out-of-Distribution Detection Under Representation and Training Paradigm Shifts","authors":"Claudio César Claros-Olivares; Austin Brockmeier","openreview_id":"xZm0LvUahe","presentation":"poster","track":"evaluations_datasets","verdict":"include","background":"Out-of-distribution (OOD) detectors, or confidence scoring functions (CSFs), assign a score to each prediction so that unreliable inputs can be flagged. Rankings of these scores are unstable across datasets, architectures and training pipelines. The paper argues that which detector works depends mainly on the geometry of the learned penultimate-layer representation, and proposes a simple intervention on that geometry: reconstructing features from a low-rank PCA subspace before scoring.","method":["Projection filtering: fit PCA on the training-set penultimate-layer activations H of a trained classifier, keep the top-k eigenvectors P (k chosen by an explained-variance threshold tuned on an ID-only validation split), and replace each feature h by ĥ = P Pᵀ(h − µ) + µ before computing the CSF. 'Global' fits one P on all training features; 'class pred' (CPP) uses a per-class projector fit on that class's features, selected by the model's predicted class. Mahalanobis (Maha) is the class-conditional Mahalanobis-distance score to the class centroids on penultimate features, using the empirical covariance matrix Σ (Appendix C.3.5).","Models follow the FD-Shifts protocol: VGG-13 trained from scratch on CIFAR-10, CIFAR-100, SuperCIFAR-100 and TinyImageNet under three training paradigms (ConfidNet / vanilla CE, DeVries, Deep Gamblers), 5 seeds per cell, with OOD sets iSUN, LSUN-crop, LSUN-resize, SVHN, Places365 and Textures. OOD detection compares OOD samples against correctly classified ID samples and is scored with AUGRC (lower is better). For each base CSF and projection variant, the paper computes paired AUGRC differences (base − variant) across source × OOD dataset × paradigm × run combinations and tests them with a two-sided Wilcoxon signed-rank test."],"claim":"On VGG-13 (FD-Shifts setup), applying global PCA projection filtering to penultimate features before Mahalanobis scoring lowers AUGRC by a mean of 24.1 (median 16.6), winning 50 of 64 paired comparisons (Wilcoxon p < 0.001).","source":"Table 15, page 34 (Appendix F); summarized in Section 4.3, page 8","metric":"AUGRC (lower is better; the paper does not state the scale, likely ×10³). ΔAUGRC = mean(base − variant) over paired source × OOD dataset × paradigm × run combinations; positive means the projected variant is better. Uncertainty is a two-sided Wilcoxon signed-rank p-value with win/loss counts; no std or CI is reported.","comparison":"Mahalanobis on raw penultimate features vs. Mahalanobis on globally PCA-projected features, paired over the same source × OOD dataset × paradigm × run combinations of VGG-13 with the same ID/OOD data; note that the training-time dropout setting is selected separately for each (CSF, projection variant) on validation AUGRC (Appendix A, Table 1), so base and projected Maha can come from differently trained checkpoints in some cells (e.g. CIFAR-10 ConfidNet: base uses dropout, global does not)","falsification":"Re-training VGG-13 models on the FD-Shifts sources and evaluating Mahalanobis with and without global projection on the same OOD sets, the projected variant does not lower AUGRC in a clear majority of paired comparisons (e.g. win rate near 50% or a paired Wilcoxon test that is not significant), or the mean gain is near zero or negative.","caption":"Table 15: Full paired AUGRC comparison: base CSF vs. projection variants on VGG-13 (Conv). ∆AUGRC = mean(base − variant); positive = variant better. Significant improvements (p < 0.05, variant better) are marked with ✓.","image":"img/2511.11934.png","kind":"table","image_w":1144,"image_h":1655,"pwc":false}
 
1
+ {"id":"2511.11934","title":"A Systematic Analysis of Out-of-Distribution Detection Under Representation and Training Paradigm Shifts","authors":"Claudio César Claros-Olivares; Austin Brockmeier","openreview_id":"xZm0LvUahe","presentation":"poster","track":"evaluations_datasets","verdict":"include","background":"Out-of-distribution (OOD) detectors, or confidence scoring functions (CSFs), assign a score to each prediction so that unreliable inputs can be flagged. Rankings of these scores are unstable across datasets, architectures and training pipelines. The paper argues that which detector works depends mainly on the geometry of the learned penultimate-layer representation, and proposes a simple intervention on that geometry: reconstructing features from a low-rank PCA subspace before scoring.","method":["Projection filtering: fit PCA on the training-set penultimate-layer activations H of a trained classifier, keep the top-k eigenvectors P (k chosen by an explained-variance threshold tuned on an ID-only validation split), and replace each feature h by ĥ = P Pᵀ(h − µ) + µ before computing the CSF. 'Global' fits one P on all training features; 'class pred' (CPP) uses a per-class projector fit on that class's features, selected by the model's predicted class. Mahalanobis (Maha) is the class-conditional Mahalanobis-distance score to the class centroids on penultimate features, using the empirical covariance matrix Σ (Appendix C.3.5).","Models follow the FD-Shifts protocol: VGG-13 trained from scratch on CIFAR-10, CIFAR-100, SuperCIFAR-100 and TinyImageNet under three training paradigms (ConfidNet / vanilla CE, DeVries, Deep Gamblers), 5 seeds per cell, with OOD sets iSUN, LSUN-crop, LSUN-resize, SVHN, Places365 and Textures. OOD detection compares OOD samples against correctly classified ID samples and is scored with AUGRC (lower is better). For each base CSF and projection variant, the paper computes paired AUGRC differences (base − variant) across source × OOD dataset × paradigm × run combinations and tests them with a two-sided Wilcoxon signed-rank test."],"claim":"On VGG-13 (FD-Shifts setup), applying global PCA projection filtering to penultimate features before Mahalanobis scoring lowers AUGRC by a mean of 24.1 (median 16.6), winning 50 of 64 paired comparisons (Wilcoxon p < 0.001).","source":"Table 15, page 34 (Appendix F); summarized in Section 4.3, page 8","metric":"AUGRC (lower is better; the paper does not state the scale, likely ×10³). ΔAUGRC = mean(base − variant) over paired source × OOD dataset × paradigm × run combinations; positive means the projected variant is better. Uncertainty is a two-sided Wilcoxon signed-rank p-value with win/loss counts; no std or CI is reported.","comparison":"Mahalanobis on raw penultimate features vs. Mahalanobis on globally PCA-projected features, paired over the same source × OOD dataset × paradigm × run combinations of VGG-13 with the same ID/OOD data; note that the training-time dropout setting is selected separately for each (CSF, projection variant) on validation AUGRC (Appendix A, Table 1), so base and projected Maha can come from differently trained checkpoints in some cells (e.g. CIFAR-10 ConfidNet: base uses dropout, global does not)","falsification":"Re-training VGG-13 models on the FD-Shifts sources and evaluating Mahalanobis with and without global projection on the same OOD sets, the projected variant does not lower AUGRC in a clear majority of paired comparisons (e.g. win rate near 50% or a paired Wilcoxon test that is not significant), or the mean gain is near zero or negative.","caption":"Table 15: Full paired AUGRC comparison: base CSF vs. projection variants on VGG-13 (Conv). ∆AUGRC = mean(base − variant); positive = variant better. Significant improvements (p < 0.05, variant better) are marked with ✓.","image":"img/2511.11934.png","kind":"table","image_w":1144,"image_h":1655,"pwc":true}
site/p/2511.15856.json CHANGED
@@ -1 +1 @@
1
- {"id":"2511.15856","title":"GLOBE: Accurate Surrogates for Boundary-Driven PDEs via Domain-Inspired Architectures and Equivariance","authors":"Peter Sharpe","openreview_id":"YvUcvzejJe","presentation":"poster","track":"main","verdict":"borderline","background":"Steady RANS simulations of aerodynamic flows are expensive, so ML surrogates that predict flow fields from geometry are widely studied. Many such surrogates do not enforce basic physical structure: they break under rotation or reflection, depend on mesh discretization or units, or use local receptive fields even though elliptic PDEs need global information flow. Classical boundary-element and panel methods get large speedups by discretizing only the boundary and representing the solution as a superposition of Green's-function influences. This paper asks whether building that structure into a neural surrogate improves accuracy, particularly with little training data.","method":["GLOBE represents output fields at arbitrary query points as an area-weighted sum over boundary faces of a learned kernel K(source face -> target point). The kernel uses only relative positions nondimensionalized by reference lengths. It builds rotation-invariant features (smoothlog magnitudes, Legendre polynomials of pairwise angles), processes them with a Padé-approximant MLP of the form sgn(phi_n)|phi_n|^N/(1+|phi_d|^D), multiplies by a far-field decay envelope (~1/r^(d-1)), and reprojects vector outputs onto local basis vectors. This makes it translation-, rotation- and parity-equivariant, discretization-invariant and units-invariant. Several kernel branches with learned multiples of the reference lengths (chord and a viscous length) are summed, and H=2 communication hyperlayers pass latent scalars/vectors and source strengths between boundary faces before the final evaluation at query points. The AirFRANS model has 117,282 parameters (hidden sizes [64,64,64], 12 latent scalars and 6 latent vectors per hyperlayer, Padé N=D=2).","On AirFRANS (steady incompressible RANS over NACA airfoils), the Scarce split trains on 200 cases and tests on the official 200 test cases. Inputs are the airfoil boundary mesh, freestream direction and reference lengths; there is no signed-distance field. Training uses a Huber loss on nondimensional fields at 2,048 query points per iteration with Muon + RAdamW. Evaluation follows the released AirFRANS evaluator: pointwise MSE at all volume-mesh nodes (surface pressure on surface nodes), averaged equally over cases, divided by one point-pooled training-set variance per field."],"claim":"On the AirFRANS Scarce split (200 training cases), GLOBE's normalized test MSE (×10^-2) is 0.0238 (u_x), 0.0148 (u_y), 0.0496 (p) and 0.1889 (surface p), versus 0.1520, 0.1130, 0.2400 and 2.7000 reported for MARIO, the best published surrogate, i.e. 4.8-14.3x lower on every field.","source":"Table 2, page 9","metric":"Mean squared error on official AirFRANS Scarce test cases, normalized by training-set field variance (released AirFRANS evaluator); lower is better. Single-run point estimates, no error bars or seeds reported.","comparison":"GLOBE (trained by the authors) vs. published values for MARIO, Transolver, MLP, GraphSAGE, PointNet and GUNet (copied from Catalani et al. 2025b and Bonnet et al. 2023, not rerun)","falsification":"A GLOBE model trained on the 200 Scarce training cases and scored with the official AirFRANS evaluator does not beat MARIO's reported MSE on all four fields, or the margin shrinks to under ~2x on volume pressure (p).","caption":"Table 2: AirFRANS Scarce (low-data) task: mean squared error on the official test cases, normalized by the training-set field variance as in Table 1; lower is better. External values are as reported in the cited works (not rerun). GLOBE values are single-run point estimates.","image":"img/2511.15856.png","kind":"table","image_w":1122,"image_h":394,"pwc":false}
 
1
+ {"id":"2511.15856","title":"GLOBE: Accurate Surrogates for Boundary-Driven PDEs via Domain-Inspired Architectures and Equivariance","authors":"Peter Sharpe","openreview_id":"YvUcvzejJe","presentation":"poster","track":"main","verdict":"borderline","background":"Steady RANS simulations of aerodynamic flows are expensive, so ML surrogates that predict flow fields from geometry are widely studied. Many such surrogates do not enforce basic physical structure: they break under rotation or reflection, depend on mesh discretization or units, or use local receptive fields even though elliptic PDEs need global information flow. Classical boundary-element and panel methods get large speedups by discretizing only the boundary and representing the solution as a superposition of Green's-function influences. This paper asks whether building that structure into a neural surrogate improves accuracy, particularly with little training data.","method":["GLOBE represents output fields at arbitrary query points as an area-weighted sum over boundary faces of a learned kernel K(source face -> target point). The kernel uses only relative positions nondimensionalized by reference lengths. It builds rotation-invariant features (smoothlog magnitudes, Legendre polynomials of pairwise angles), processes them with a Padé-approximant MLP of the form sgn(phi_n)|phi_n|^N/(1+|phi_d|^D), multiplies by a far-field decay envelope (~1/r^(d-1)), and reprojects vector outputs onto local basis vectors. This makes it translation-, rotation- and parity-equivariant, discretization-invariant and units-invariant. Several kernel branches with learned multiples of the reference lengths (chord and a viscous length) are summed, and H=2 communication hyperlayers pass latent scalars/vectors and source strengths between boundary faces before the final evaluation at query points. The AirFRANS model has 117,282 parameters (hidden sizes [64,64,64], 12 latent scalars and 6 latent vectors per hyperlayer, Padé N=D=2).","On AirFRANS (steady incompressible RANS over NACA airfoils), the Scarce split trains on 200 cases and tests on the official 200 test cases. Inputs are the airfoil boundary mesh, freestream direction and reference lengths; there is no signed-distance field. Training uses a Huber loss on nondimensional fields at 2,048 query points per iteration with Muon + RAdamW. Evaluation follows the released AirFRANS evaluator: pointwise MSE at all volume-mesh nodes (surface pressure on surface nodes), averaged equally over cases, divided by one point-pooled training-set variance per field."],"claim":"On the AirFRANS Scarce split (200 training cases), GLOBE's normalized test MSE (×10^-2) is 0.0238 (u_x), 0.0148 (u_y), 0.0496 (p) and 0.1889 (surface p), versus 0.1520, 0.1130, 0.2400 and 2.7000 reported for MARIO, the best published surrogate, i.e. 4.8-14.3x lower on every field.","source":"Table 2, page 9","metric":"Mean squared error on official AirFRANS Scarce test cases, normalized by training-set field variance (released AirFRANS evaluator); lower is better. Single-run point estimates, no error bars or seeds reported.","comparison":"GLOBE (trained by the authors) vs. published values for MARIO, Transolver, MLP, GraphSAGE, PointNet and GUNet (copied from Catalani et al. 2025b and Bonnet et al. 2023, not rerun)","falsification":"A GLOBE model trained on the 200 Scarce training cases and scored with the official AirFRANS evaluator does not beat MARIO's reported MSE on all four fields, or the margin shrinks to under ~2x on volume pressure (p).","caption":"Table 2: AirFRANS Scarce (low-data) task: mean squared error on the official test cases, normalized by the training-set field variance as in Table 1; lower is better. External values are as reported in the cited works (not rerun). GLOBE values are single-run point estimates.","image":"img/2511.15856.png","kind":"table","image_w":1122,"image_h":394,"pwc":true}
site/p/2601.04539.json CHANGED
@@ -1 +1 @@
1
- {"id":"2601.04539","title":"Paradoxical noise preference in RNNs","authors":"Noah Eckstein; Manoj Srinivasan","openreview_id":"","presentation":"poster","track":"journal:TMLR","verdict":"include","background":"Recurrent neural networks used as models of biological neural populations are usually trained with injected noise, both to mimic neural variability and as a regularizer. The usual assumption is that removing the noise at test time should keep or improve performance. This paper reports the opposite for continuous-time RNNs (CTRNNs) when the noise is injected inside the activation function, as is common in computational neuroscience: the trained networks perform best at a nonzero test noise close to the training noise, because the noise shifts the stationary states of the dynamics and the network learns to rely on that shift.","method":["The CTRNN is Euler-discretized as h_{t+1} = (1 - γ) h_t + γ f(W_rec h_t + W_in u_t + B_in + N(0, σ_in²)) for noise-in networks, and h_{t+1} = (1 - γ) h_t + γ (f(W_rec h_t + W_in u_t + B_in) + N(0, σ_out²)) for noise-out networks, with γ = Δt/τ = 0.02/0.1 = 0.2, output z = W_out h + B_out, 100 neurons, and ReLU or SoftPlus (α = 10) activation. Noise is sampled independently per neuron per step. Weights start from N(0, 0.8/√n), biases at zero, and the initial state is a trained vector starting at zero.","Task: compute g(x) = sin(x) for x ∈ [0, 2π] (also tanh(x) on [-4, 4]); the constant input x is on for 100 steps and the output is read at the last step (fixed readout) or, for random readout, at a step drawn uniformly from [70, 100] (in training, the random-readout loss uses two readout times: the final step and one random step). Training uses an RMSE loss, Adam (lr 0.001, half-life 2,500 batches), 10,000 batches of 32 trials, training noise std 0.1, and no other regularization. The trained networks are then tested over a range of noise std from 0 to about 0.15 (same placement as in training), and RMSE, mean error (bias) and error std are reported over many noise realizations."],"claim":"For 100-neuron CTRNNs trained on sin(x) (fixed readout) with noise std 0.1, noise-in networks reach their lowest test RMSE (about 0.03) at test noise ≈ 0.1 and have about 0.23 RMSE at zero test noise, whereas noise-out networks have their lowest RMSE (about 0.015) at zero test noise, rising to about 0.04 at 0.1.","source":"Figure 2a and 2d, page 5","metric":"Test RMSE between network output and target over inputs and noise realizations (lower is better), split into mean error (bias) and error std; no error bars or seed variation shown in Figure 2 (Figure A1 shows four independent runs qualitatively); values are read off the plot and approximate","comparison":"Noise-in vs. noise-out CTRNNs trained identically (same task, size and training noise 0.1), each tested from zero noise up to above the training noise; and within noise-in networks, test noise 0 vs. test noise at the training level","falsification":"Retrained noise-in sin(x) networks (training noise 0.1) have test RMSE at zero noise that is not clearly higher than at noise ≈ 0.1 (e.g. less than 2x), or the minimum of RMSE vs. test noise is at or near zero; or noise-out networks trained the same way show a minimum at a clearly nonzero test noise.","caption":"Figure 2: Simple function computation: Noise-in networks prefer the training noise level, but noise-out networks prefer zero noise. a) Root mean squared error, mean error, and error standard deviations for noise-in function computation networks. The best performance occurs when testing at or near the training noise level, and this preference is driven by a non-monotonic dependence between the noise level and the mean error (i.e., the systematic error). [...] b) Noise-preference results for fixed readout time sine computation networks trained with various training noise levels. Noise preference becomes more pronounced for greater training noise, with very low training noise resulting in zero noise preference. [...] d) Root mean squared error, mean error, and error standard deviations for noise-out function computation networks. Unlike the noise-in networks, the noise-out networks perform best when tested with zero noise, and the systematic error appears unaffected by changes in the noise.","image":"img/2601.04539.png","kind":"figure","image_w":1319,"image_h":1780,"pwc":false}
 
1
+ {"id":"2601.04539","title":"Paradoxical noise preference in RNNs","authors":"Noah Eckstein; Manoj Srinivasan","openreview_id":"","presentation":"poster","track":"journal:TMLR","verdict":"include","background":"Recurrent neural networks used as models of biological neural populations are usually trained with injected noise, both to mimic neural variability and as a regularizer. The usual assumption is that removing the noise at test time should keep or improve performance. This paper reports the opposite for continuous-time RNNs (CTRNNs) when the noise is injected inside the activation function, as is common in computational neuroscience: the trained networks perform best at a nonzero test noise close to the training noise, because the noise shifts the stationary states of the dynamics and the network learns to rely on that shift.","method":["The CTRNN is Euler-discretized as h_{t+1} = (1 - γ) h_t + γ f(W_rec h_t + W_in u_t + B_in + N(0, σ_in²)) for noise-in networks, and h_{t+1} = (1 - γ) h_t + γ (f(W_rec h_t + W_in u_t + B_in) + N(0, σ_out²)) for noise-out networks, with γ = Δt/τ = 0.02/0.1 = 0.2, output z = W_out h + B_out, 100 neurons, and ReLU or SoftPlus (α = 10) activation. Noise is sampled independently per neuron per step. Weights start from N(0, 0.8/√n), biases at zero, and the initial state is a trained vector starting at zero.","Task: compute g(x) = sin(x) for x ∈ [0, 2π] (also tanh(x) on [-4, 4]); the constant input x is on for 100 steps and the output is read at the last step (fixed readout) or, for random readout, at a step drawn uniformly from [70, 100] (in training, the random-readout loss uses two readout times: the final step and one random step). Training uses an RMSE loss, Adam (lr 0.001, half-life 2,500 batches), 10,000 batches of 32 trials, training noise std 0.1, and no other regularization. The trained networks are then tested over a range of noise std from 0 to about 0.15 (same placement as in training), and RMSE, mean error (bias) and error std are reported over many noise realizations."],"claim":"For 100-neuron CTRNNs trained on sin(x) (fixed readout) with noise std 0.1, noise-in networks reach their lowest test RMSE (about 0.03) at test noise ≈ 0.1 and have about 0.23 RMSE at zero test noise, whereas noise-out networks have their lowest RMSE (about 0.015) at zero test noise, rising to about 0.04 at 0.1.","source":"Figure 2a and 2d, page 5","metric":"Test RMSE between network output and target over inputs and noise realizations (lower is better), split into mean error (bias) and error std; no error bars or seed variation shown in Figure 2 (Figure A1 shows four independent runs qualitatively); values are read off the plot and approximate","comparison":"Noise-in vs. noise-out CTRNNs trained identically (same task, size and training noise 0.1), each tested from zero noise up to above the training noise; and within noise-in networks, test noise 0 vs. test noise at the training level","falsification":"Retrained noise-in sin(x) networks (training noise 0.1) have test RMSE at zero noise that is not clearly higher than at noise ≈ 0.1 (e.g. less than 2x), or the minimum of RMSE vs. test noise is at or near zero; or noise-out networks trained the same way show a minimum at a clearly nonzero test noise.","caption":"Figure 2: Simple function computation: Noise-in networks prefer the training noise level, but noise-out networks prefer zero noise. a) Root mean squared error, mean error, and error standard deviations for noise-in function computation networks. The best performance occurs when testing at or near the training noise level, and this preference is driven by a non-monotonic dependence between the noise level and the mean error (i.e., the systematic error). [...] b) Noise-preference results for fixed readout time sine computation networks trained with various training noise levels. Noise preference becomes more pronounced for greater training noise, with very low training noise resulting in zero noise preference. [...] d) Root mean squared error, mean error, and error standard deviations for noise-out function computation networks. Unlike the noise-in networks, the noise-out networks perform best when tested with zero noise, and the systematic error appears unaffected by changes in the noise.","image":"img/2601.04539.png","kind":"figure","image_w":1319,"image_h":1780,"pwc":true}
site/p/2601.09859.json CHANGED
@@ -1 +1 @@
1
- {"id":"2601.09859","title":"Tune-Up Open-Weight CLIP: Optimization Framework for Self-Supervised Fine-tuning of CLIP","authors":"Anant Mehta; Xiyuan Wei; Xingyu Chen; Tianbao Yang","openreview_id":"F6LR7aEwxS","presentation":"spotlight","track":"main","verdict":"borderline","background":"Improving CLIP models usually means pretraining again from scratch on billions of image-text pairs, which few groups can afford. Supervised fine-tuning on a labelled target dataset improves that one task but tends to hurt robustness and transfer. The paper studies self-supervised fine-tuning (SSFT): continuing contrastive training of an already-released CLIP checkpoint on a generic web image-text corpus to improve it across many tasks at once. The authors and prior work (CLIP-Refine) observe that simply resuming standard contrastive training from a well-pretrained checkpoint makes the model worse, with a sharp drop in the first epoch that is not recovered later.","method":["TuneCLIP has two stages. Stage I, Optimizer Statistics Recovery (OSR), runs the SogCLR/FastCLIP global-contrastive-loss update for 5 epochs with the model weights frozen at the pretrained checkpoint, only to accumulate the Adam first and second moments (m, v) and the per-sample running estimates u_{i,x}, u_{i,z} of the softmax normalisers. The authors argue that the usual zero or single-batch initialisation of these statistics (cold-start bias) causes the early performance drop. Stage II initialises the optimizer and u-estimates from OSR and fine-tunes for 5 more epochs with SogCLR + AdamW on a Hinged Global Contrastive Loss (HGCL), where the pairwise term s_ij - s_ii is replaced by max(s_ij - s_ii + m, 0)^2 with margin m = 0.1, so negatives already m below the positive stop receiving gradient (meant to limit over-penalising false negatives).","Main setting: OpenAI CLIP ViT-B/16 fine-tuned on DFN-12M, temperature fixed to the checkpoint's value, AdamW (0.9, 0.98), lr 1e-5 with cosine schedule, weight decay 0.02, global batch 2048 (256 x 8 GPUs), 224x224 images. Baselines run on the same data and model: FastCLIP (GCL + SogCLR, 5 epochs), OpenCLIP (mini-batch contrastive loss, 5 epochs) and CLIP-Refine (1 epoch, as its authors recommend). Evaluation follows the DataComp protocol: mean zero-shot top-1 over ImageNet-1k and six distribution-shift variants, mean Recall@1 over MSCOCO and Flickr30k image and text retrieval, and the 38-task DataComp average."],"claim":"Fine-tuning OpenAI CLIP ViT-B/16 on DFN-12M, TuneCLIP raises mean retrieval Recall@1 from 57.46 to 64.12, while the best matched baseline (CLIP-Refine) reaches 58.82; TuneCLIP also improves ImageNet & variants (57.67 -> 59.36) and DataComp (56.26 -> 58.62), where FastCLIP, OpenCLIP and CLIP-Refine all fall below the base model.","source":"Table 1, page 8","metric":"Mean zero-shot accuracy on ImageNet-1k + 6 variants, mean Recall@1 on MSCOCO/Flickr30k retrieval (I->T and T->I), DataComp 38-task average; all higher is better. Single runs, no seeds or error bars reported.","comparison":"TuneCLIP (OSR + HGCL) vs. the untuned base checkpoint and vs. FastCLIP, OpenCLIP and CLIP-Refine fine-tuned on the same DFN-12M data from the same OpenAI ViT-B/16 checkpoint","falsification":"On OpenAI ViT-B/16 fine-tuned on DFN-12M, TuneCLIP's mean retrieval Recall@1 is not clearly above both the base model (57.46) and the best baseline (CLIP-Refine, 58.82), e.g. less than about 2 points above, or TuneCLIP falls below the base model on ImageNet & variants or DataComp.","caption":"Table 1: Summary of mean zero-shot performance across ImageNet & distributional variants, retrieval benchmarks, and the DataComp benchmark, together with wall-clock training time per GPU.","image":"img/2601.09859.png","kind":"table","image_w":1144,"image_h":483,"pwc":false}
 
1
+ {"id":"2601.09859","title":"Tune-Up Open-Weight CLIP: Optimization Framework for Self-Supervised Fine-tuning of CLIP","authors":"Anant Mehta; Xiyuan Wei; Xingyu Chen; Tianbao Yang","openreview_id":"F6LR7aEwxS","presentation":"spotlight","track":"main","verdict":"borderline","background":"Improving CLIP models usually means pretraining again from scratch on billions of image-text pairs, which few groups can afford. Supervised fine-tuning on a labelled target dataset improves that one task but tends to hurt robustness and transfer. The paper studies self-supervised fine-tuning (SSFT): continuing contrastive training of an already-released CLIP checkpoint on a generic web image-text corpus to improve it across many tasks at once. The authors and prior work (CLIP-Refine) observe that simply resuming standard contrastive training from a well-pretrained checkpoint makes the model worse, with a sharp drop in the first epoch that is not recovered later.","method":["TuneCLIP has two stages. Stage I, Optimizer Statistics Recovery (OSR), runs the SogCLR/FastCLIP global-contrastive-loss update for 5 epochs with the model weights frozen at the pretrained checkpoint, only to accumulate the Adam first and second moments (m, v) and the per-sample running estimates u_{i,x}, u_{i,z} of the softmax normalisers. The authors argue that the usual zero or single-batch initialisation of these statistics (cold-start bias) causes the early performance drop. Stage II initialises the optimizer and u-estimates from OSR and fine-tunes for 5 more epochs with SogCLR + AdamW on a Hinged Global Contrastive Loss (HGCL), where the pairwise term s_ij - s_ii is replaced by max(s_ij - s_ii + m, 0)^2 with margin m = 0.1, so negatives already m below the positive stop receiving gradient (meant to limit over-penalising false negatives).","Main setting: OpenAI CLIP ViT-B/16 fine-tuned on DFN-12M, temperature fixed to the checkpoint's value, AdamW (0.9, 0.98), lr 1e-5 with cosine schedule, weight decay 0.02, global batch 2048 (256 x 8 GPUs), 224x224 images. Baselines run on the same data and model: FastCLIP (GCL + SogCLR, 5 epochs), OpenCLIP (mini-batch contrastive loss, 5 epochs) and CLIP-Refine (1 epoch, as its authors recommend). Evaluation follows the DataComp protocol: mean zero-shot top-1 over ImageNet-1k and six distribution-shift variants, mean Recall@1 over MSCOCO and Flickr30k image and text retrieval, and the 38-task DataComp average."],"claim":"Fine-tuning OpenAI CLIP ViT-B/16 on DFN-12M, TuneCLIP raises mean retrieval Recall@1 from 57.46 to 64.12, while the best matched baseline (CLIP-Refine) reaches 58.82; TuneCLIP also improves ImageNet & variants (57.67 -> 59.36) and DataComp (56.26 -> 58.62), where FastCLIP, OpenCLIP and CLIP-Refine all fall below the base model.","source":"Table 1, page 8","metric":"Mean zero-shot accuracy on ImageNet-1k + 6 variants, mean Recall@1 on MSCOCO/Flickr30k retrieval (I->T and T->I), DataComp 38-task average; all higher is better. Single runs, no seeds or error bars reported.","comparison":"TuneCLIP (OSR + HGCL) vs. the untuned base checkpoint and vs. FastCLIP, OpenCLIP and CLIP-Refine fine-tuned on the same DFN-12M data from the same OpenAI ViT-B/16 checkpoint","falsification":"On OpenAI ViT-B/16 fine-tuned on DFN-12M, TuneCLIP's mean retrieval Recall@1 is not clearly above both the base model (57.46) and the best baseline (CLIP-Refine, 58.82), e.g. less than about 2 points above, or TuneCLIP falls below the base model on ImageNet & variants or DataComp.","caption":"Table 1: Summary of mean zero-shot performance across ImageNet & distributional variants, retrieval benchmarks, and the DataComp benchmark, together with wall-clock training time per GPU.","image":"img/2601.09859.png","kind":"table","image_w":1144,"image_h":483,"pwc":true}
site/p/2601.12654.json CHANGED
@@ -1 +1 @@
1
- {"id":"2601.12654","title":"Explanation Multiplicity in SHAP: Characterization and Assessment","authors":"Hyunseung Hwang; Seungeun Lee; Lucas Rosenblatt; Steven Whang; Julia Stoyanovich","openreview_id":"ANaxIH1evO","presentation":"poster","track":"evaluations_datasets","verdict":"include","background":"SHAP feature attributions are used to justify individual decisions in lending, healthcare and public services, usually as a ranked list of the most important features. Model-agnostic SHAP estimators such as Kernel SHAP approximate Shapley values by filling 'absent' features from a small sampled background dataset (often 100 rows, as in the SHAP tutorials), so rerunning the same explainer on the same trained model and instance can give a different explanation. The paper calls this explanation multiplicity and argues that the common magnitude-based stability check (l2 distance between attribution vectors) hides how often the top-ranked features change.","method":["For each dataset and model (Decision Tree, Random Forest, XGBoost, MLP, FT-Transformer, tuned by grid search on the training fold), the model is trained once and Kernel SHAP is run R = 5 times on every test instance of the first held-out fold of a stratified five-fold split, each run using a different background seed to draw K = 100 background rows. Background strategies are Random row sampling, Kernel SHAP K-Means summarization, prediction-stratified sampling (PredStrat) and Compress-Then-Explain (CTE). Kernel SHAP uses min(2^d, 1024) coalitions; GMSC has d = 10 and N = 9,000 test instances.","For each instance, the 10 run pairs are compared with three distances: l2 between attribution vectors, top-3 Jaccard distance between the sets of the three features with largest |phi|, and 1 - RBO. Each distance is thresholded at the minimum of a randomized null-model band: a Dirichlet model (total mass T = 0.4, r = floor(d/3), rho in {0.6, 0.7, 0.8}, kappa in {5..15}) gives an RMS l2 reference, and a Mallows model (q in {0.3, 0.4, 0.5}) gives the rank-metric references. The reported number is the percentage of the 10N run-pair-instance comparisons whose distance strictly exceeds the threshold (Eq. 7). For top-3 Jaccard the threshold lies below the smallest positive value (0.5), so exceedance equals the frequency with which the top-3 feature set changes between two runs."],"claim":"On Give Me Some Credit with Kernel SHAP and K = 100 background rows, 0% of run-pair-instance comparisons exceed the l2 randomized baseline for every model and sampling method, whereas under Random, PredStrat and CTE backgrounds the top-3 feature set differs between two reruns in 10-65% of comparisons (Random sampling: DT 10%, RF 29%, XGB 42%, MLP 28%, FTT 49%; PredStrat 37-65%; CTE 29-44%; K-Means is lower at 2-15%).","source":"Figure 4, page 8 (other datasets in Figure 12, Appendix E)","metric":"Percentage of 10N run-pair-instance comparisons (N = 9,000) whose distance strictly exceeds the minimum of the randomized null-model band; higher means more disagreement. No error bars, seeds or confidence intervals are reported; one trained model per configuration.","comparison":"l2-distance exceedance vs. top-3 Jaccard (and 1-RBO) exceedance on the same rerun pairs, across 5 models x 4 background sampling methods.","falsification":"Re-running Kernel SHAP (K = 100, 5 background seeds) on GMSC with a Random-sampled background for a tuned XGBoost or Random Forest gives a top-3 set change rate far below the reported value (e.g. under 10% where the paper reports 29-42%), or a substantial fraction (more than a few percent) of l2 distances exceeds the Dirichlet RMS reference.","caption":"Figure 4: Metric divergence for Kernel SHAP on GMSC. Percentage of run-pair-instance comparisons exceeding the randomized baseline, by method and model (Eq. (7)). Each cell uses 10N comparisons. Low l2 exceedance coexists with Top-3 membership churn and RBO disagreement.","image":"img/2601.12654.png","kind":"figure","image_w":1144,"image_h":794,"pwc":false}
 
1
+ {"id":"2601.12654","title":"Explanation Multiplicity in SHAP: Characterization and Assessment","authors":"Hyunseung Hwang; Seungeun Lee; Lucas Rosenblatt; Steven Whang; Julia Stoyanovich","openreview_id":"ANaxIH1evO","presentation":"poster","track":"evaluations_datasets","verdict":"include","background":"SHAP feature attributions are used to justify individual decisions in lending, healthcare and public services, usually as a ranked list of the most important features. Model-agnostic SHAP estimators such as Kernel SHAP approximate Shapley values by filling 'absent' features from a small sampled background dataset (often 100 rows, as in the SHAP tutorials), so rerunning the same explainer on the same trained model and instance can give a different explanation. The paper calls this explanation multiplicity and argues that the common magnitude-based stability check (l2 distance between attribution vectors) hides how often the top-ranked features change.","method":["For each dataset and model (Decision Tree, Random Forest, XGBoost, MLP, FT-Transformer, tuned by grid search on the training fold), the model is trained once and Kernel SHAP is run R = 5 times on every test instance of the first held-out fold of a stratified five-fold split, each run using a different background seed to draw K = 100 background rows. Background strategies are Random row sampling, Kernel SHAP K-Means summarization, prediction-stratified sampling (PredStrat) and Compress-Then-Explain (CTE). Kernel SHAP uses min(2^d, 1024) coalitions; GMSC has d = 10 and N = 9,000 test instances.","For each instance, the 10 run pairs are compared with three distances: l2 between attribution vectors, top-3 Jaccard distance between the sets of the three features with largest |phi|, and 1 - RBO. Each distance is thresholded at the minimum of a randomized null-model band: a Dirichlet model (total mass T = 0.4, r = floor(d/3), rho in {0.6, 0.7, 0.8}, kappa in {5..15}) gives an RMS l2 reference, and a Mallows model (q in {0.3, 0.4, 0.5}) gives the rank-metric references. The reported number is the percentage of the 10N run-pair-instance comparisons whose distance strictly exceeds the threshold (Eq. 7). For top-3 Jaccard the threshold lies below the smallest positive value (0.5), so exceedance equals the frequency with which the top-3 feature set changes between two runs."],"claim":"On Give Me Some Credit with Kernel SHAP and K = 100 background rows, 0% of run-pair-instance comparisons exceed the l2 randomized baseline for every model and sampling method, whereas under Random, PredStrat and CTE backgrounds the top-3 feature set differs between two reruns in 10-65% of comparisons (Random sampling: DT 10%, RF 29%, XGB 42%, MLP 28%, FTT 49%; PredStrat 37-65%; CTE 29-44%; K-Means is lower at 2-15%).","source":"Figure 4, page 8 (other datasets in Figure 12, Appendix E)","metric":"Percentage of 10N run-pair-instance comparisons (N = 9,000) whose distance strictly exceeds the minimum of the randomized null-model band; higher means more disagreement. No error bars, seeds or confidence intervals are reported; one trained model per configuration.","comparison":"l2-distance exceedance vs. top-3 Jaccard (and 1-RBO) exceedance on the same rerun pairs, across 5 models x 4 background sampling methods.","falsification":"Re-running Kernel SHAP (K = 100, 5 background seeds) on GMSC with a Random-sampled background for a tuned XGBoost or Random Forest gives a top-3 set change rate far below the reported value (e.g. under 10% where the paper reports 29-42%), or a substantial fraction (more than a few percent) of l2 distances exceeds the Dirichlet RMS reference.","caption":"Figure 4: Metric divergence for Kernel SHAP on GMSC. Percentage of run-pair-instance comparisons exceeding the randomized baseline, by method and model (Eq. (7)). Each cell uses 10N comparisons. Low l2 exceedance coexists with Top-3 membership churn and RBO disagreement.","image":"img/2601.12654.png","kind":"figure","image_w":1144,"image_h":794,"pwc":true}
site/p/2601.13562.json CHANGED
@@ -1 +1 @@
1
- {"id":"2601.13562","title":"RoSeViT: Role-Separated Vision Transformers for ARC Visual Reasoning","authors":"Zhiguang Liu; Yi Shang","openreview_id":"McXZ63160k","presentation":"poster","track":"main","verdict":"borderline","background":"ARC (Abstraction and Reasoning Corpus) tasks give a few input-output grid pairs and ask a solver to infer the transformation and apply it to a new grid. VARC recently showed that a plain Vision Transformer, trained offline on ARC-1 training tasks (expanded with RE-ARC), then adapted per task with test-time training and evaluated with multi-view voting, is a strong ARC solver. In a dense ViT every patch token carries both local grid state and the global transformation rule. This paper asks whether giving the rule its own route through the network, separate from the visual workspace, improves rule induction.","method":["RoSeViT keeps the VARC frontend (64x64 augmented canvas, 2x2 patches, so 1024 workspace tokens, width 512, depth 10) and adds controller tokens: one learned task token plus one context token per demonstration pair, computed by an MLP on the pooled difference of output and input patch embeddings. Each block (variant d) runs a normal dense self-attention pass, then a structured masked-attention pass where controller tokens attend to everything while workspace tokens attend only to the controller tokens, themselves and their 8 spatial neighbours, then an MLP. Nonlocal workspace-to-workspace communication in the structured pass therefore goes through the controller. d-1 means one step, with no recurrence.","The controlled test (Table 1) keeps the VARC protocol fixed: 100 offline epochs on ARC-1 training tasks with RE-ARC expansion, VARC default test-time adaptation (TTT1, 100 epochs per task, fresh task token), 510 augmented views with majority voting, and pass@2 on the 400 ARC-1 evaluation tasks. Baselines are the 18M VARC-ViT and a dense VARC-ViT widened to 28M (embed 640, MLP 720) to match RoSeViT d-1's 28.4M parameters, run with the official VARC code."],"claim":"Under the identical VARC training/TTT1/510-view-voting protocol, RoSeViT d-1 (28M) reaches 56.6% ARC-1 pass@2 (mean over 4 seeds), versus 53.6 ± 1.0% for a parameter-matched 28M dense VARC-ViT (8 seeds, best seed 54.9) and 54.5% for the 18M VARC-ViT.","source":"Table 1, page 7 (matched-control details in Section 5.1 and Appendix D.1, page 16)","metric":"ARC-1 pass@2 accuracy on the 400 evaluation tasks, %, higher is better. RoSeViT: mean over 4 seeds, no spread reported. Matched dense control: mean ± std over 8 adaptation seeds (range 52.1-54.9). VARC-ViT 18M: single number, no error.","comparison":"RoSeViT d-1 (role-separated, 28M) vs. width-matched dense VARC-ViT (28M) and default dense VARC-ViT (18M), all with VARC default adaptation (TTT1)","falsification":"Re-running both arms under the same protocol, RoSeViT d-1's mean ARC-1 pass@2 is not above the matched dense control's best seed (around 54.9), or the gap between the means falls within about one seed std (~1 point).","caption":"Table 1: Same-training/same-testing backbone comparison (VARC default adaptation, ARC-1 pass@2) with a parameter-matched dense control and the measured forward cost (batch 8, H100). Theoretical GFLOPs count the sparse structured pass; actual GFLOPs are those of our masked dense implementation.","image":"img/2601.13562.png","kind":"table","image_w":1144,"image_h":294,"pwc":false}
 
1
+ {"id":"2601.13562","title":"RoSeViT: Role-Separated Vision Transformers for ARC Visual Reasoning","authors":"Zhiguang Liu; Yi Shang","openreview_id":"McXZ63160k","presentation":"poster","track":"main","verdict":"borderline","background":"ARC (Abstraction and Reasoning Corpus) tasks give a few input-output grid pairs and ask a solver to infer the transformation and apply it to a new grid. VARC recently showed that a plain Vision Transformer, trained offline on ARC-1 training tasks (expanded with RE-ARC), then adapted per task with test-time training and evaluated with multi-view voting, is a strong ARC solver. In a dense ViT every patch token carries both local grid state and the global transformation rule. This paper asks whether giving the rule its own route through the network, separate from the visual workspace, improves rule induction.","method":["RoSeViT keeps the VARC frontend (64x64 augmented canvas, 2x2 patches, so 1024 workspace tokens, width 512, depth 10) and adds controller tokens: one learned task token plus one context token per demonstration pair, computed by an MLP on the pooled difference of output and input patch embeddings. Each block (variant d) runs a normal dense self-attention pass, then a structured masked-attention pass where controller tokens attend to everything while workspace tokens attend only to the controller tokens, themselves and their 8 spatial neighbours, then an MLP. Nonlocal workspace-to-workspace communication in the structured pass therefore goes through the controller. d-1 means one step, with no recurrence.","The controlled test (Table 1) keeps the VARC protocol fixed: 100 offline epochs on ARC-1 training tasks with RE-ARC expansion, VARC default test-time adaptation (TTT1, 100 epochs per task, fresh task token), 510 augmented views with majority voting, and pass@2 on the 400 ARC-1 evaluation tasks. Baselines are the 18M VARC-ViT and a dense VARC-ViT widened to 28M (embed 640, MLP 720) to match RoSeViT d-1's 28.4M parameters, run with the official VARC code."],"claim":"Under the identical VARC training/TTT1/510-view-voting protocol, RoSeViT d-1 (28M) reaches 56.6% ARC-1 pass@2 (mean over 4 seeds), versus 53.6 ± 1.0% for a parameter-matched 28M dense VARC-ViT (8 seeds, best seed 54.9) and 54.5% for the 18M VARC-ViT.","source":"Table 1, page 7 (matched-control details in Section 5.1 and Appendix D.1, page 16)","metric":"ARC-1 pass@2 accuracy on the 400 evaluation tasks, %, higher is better. RoSeViT: mean over 4 seeds, no spread reported. Matched dense control: mean ± std over 8 adaptation seeds (range 52.1-54.9). VARC-ViT 18M: single number, no error.","comparison":"RoSeViT d-1 (role-separated, 28M) vs. width-matched dense VARC-ViT (28M) and default dense VARC-ViT (18M), all with VARC default adaptation (TTT1)","falsification":"Re-running both arms under the same protocol, RoSeViT d-1's mean ARC-1 pass@2 is not above the matched dense control's best seed (around 54.9), or the gap between the means falls within about one seed std (~1 point).","caption":"Table 1: Same-training/same-testing backbone comparison (VARC default adaptation, ARC-1 pass@2) with a parameter-matched dense control and the measured forward cost (batch 8, H100). Theoretical GFLOPs count the sparse structured pass; actual GFLOPs are those of our masked dense implementation.","image":"img/2601.13562.png","kind":"table","image_w":1144,"image_h":294,"pwc":true}
site/p/2601.18842.json CHANGED
@@ -1 +1 @@
1
- {"id":"2601.18842","title":"GUIGuard-Bench: Toward a General Evaluation for Privacy-Preserving GUI Agents","authors":"Yanxi Wang; Zhiling Zhang; Wenbo Zhou; Weiming Zhang; Jie Zhang; Qiannan Zhu; Yu Shi; Shuxin Zheng; Jiyan He","openreview_id":"FnL0OI4r5R","presentation":"poster","track":"evaluations_datasets","verdict":"borderline","background":"GUI agents send screenshots to remote vision-language models for planning, and those screenshots often contain names, accounts, messages, locations and other private data. Existing visual-privacy datasets use natural photos, and GUI-agent benchmarks measure grounding or task success without privacy labels. Before a local module can mask private content, it has to find it on the screen. GUIGuard-Bench asks how well current VLMs can find and label private regions in real GUI-agent trajectories.","method":["The benchmark has 241 real GUI-agent trajectories (4,080 screenshots) from Android and PC. Each private region is annotated with a bounding box, its extracted text, a risk level (low/medium/high), one of six semantic privacy categories, and whether it is necessary for the task. 121 trajectories (2,002 screenshots) are public; the other 120 are private and evaluated only by the authors on request. Eight VLMs (Gemini 3 Pro, Claude Sonnet 4.5, GPT-5.1, Grok 4, Qwen3-VL-235B-A22B-Thinking, Doubao-Seed-1.6-Vision, GLM-4.5V, DeepSeek-VL2) each get the screenshot plus a single privacy prompt (Appendix L.1) that defines the taxonomy and output format, and they return privacy regions with text, box and labels.","Three metrics are computed. Binary privacy detection: whether the model correctly says if the screenshot contains any private content. Privacy recall: the fraction of ground-truth privacy elements matched, where a match needs both relaxed text matching (at least 90% character coverage in either direction) and box IoU >= 0.6. Full match: an element counts only if it is matched and its risk, category and task-necessity labels are all correct. Figure 5 reports each metric averaged over the eight models, separately for Android and PC."],"claim":"Averaged over eight VLMs on GUIGuard-Bench, privacy recognition falls sharply from screenshot-level binary detection to element-level recall to strict full match: 89.0% -> 52.9% -> 8.8% on Android and 63.3% -> 13.5% -> 0.6% on PC.","source":"Figure 5, page 7 (per-model values in Figure 10, page 30)","metric":"Accuracy / recall in % (higher is better), averaged over 8 models; no error bars, seeds or confidence intervals reported","comparison":"Binary privacy detection vs. element-level privacy recall vs. strict full-match accuracy, on Android vs. PC screenshots","falsification":"With the released prompt and matching rules on the public split, a comparable set of VLMs gets average element recall within about 10 points of binary detection, or full-match accuracy above about 25% on Android, or PC is not clearly below Android on all three metrics.","caption":"Figure 5: Average privacy-recognition performance on Android and PC, comparing binary privacy detection, privacy recall rate, and strict full-match accuracy.","image":"img/2601.18842.png","kind":"figure","image_w":572,"image_h":477,"pwc":false}
 
1
+ {"id":"2601.18842","title":"GUIGuard-Bench: Toward a General Evaluation for Privacy-Preserving GUI Agents","authors":"Yanxi Wang; Zhiling Zhang; Wenbo Zhou; Weiming Zhang; Jie Zhang; Qiannan Zhu; Yu Shi; Shuxin Zheng; Jiyan He","openreview_id":"FnL0OI4r5R","presentation":"poster","track":"evaluations_datasets","verdict":"borderline","background":"GUI agents send screenshots to remote vision-language models for planning, and those screenshots often contain names, accounts, messages, locations and other private data. Existing visual-privacy datasets use natural photos, and GUI-agent benchmarks measure grounding or task success without privacy labels. Before a local module can mask private content, it has to find it on the screen. GUIGuard-Bench asks how well current VLMs can find and label private regions in real GUI-agent trajectories.","method":["The benchmark has 241 real GUI-agent trajectories (4,080 screenshots) from Android and PC. Each private region is annotated with a bounding box, its extracted text, a risk level (low/medium/high), one of six semantic privacy categories, and whether it is necessary for the task. 121 trajectories (2,002 screenshots) are public; the other 120 are private and evaluated only by the authors on request. Eight VLMs (Gemini 3 Pro, Claude Sonnet 4.5, GPT-5.1, Grok 4, Qwen3-VL-235B-A22B-Thinking, Doubao-Seed-1.6-Vision, GLM-4.5V, DeepSeek-VL2) each get the screenshot plus a single privacy prompt (Appendix L.1) that defines the taxonomy and output format, and they return privacy regions with text, box and labels.","Three metrics are computed. Binary privacy detection: whether the model correctly says if the screenshot contains any private content. Privacy recall: the fraction of ground-truth privacy elements matched, where a match needs both relaxed text matching (at least 90% character coverage in either direction) and box IoU >= 0.6. Full match: an element counts only if it is matched and its risk, category and task-necessity labels are all correct. Figure 5 reports each metric averaged over the eight models, separately for Android and PC."],"claim":"Averaged over eight VLMs on GUIGuard-Bench, privacy recognition falls sharply from screenshot-level binary detection to element-level recall to strict full match: 89.0% -> 52.9% -> 8.8% on Android and 63.3% -> 13.5% -> 0.6% on PC.","source":"Figure 5, page 7 (per-model values in Figure 10, page 30)","metric":"Accuracy / recall in % (higher is better), averaged over 8 models; no error bars, seeds or confidence intervals reported","comparison":"Binary privacy detection vs. element-level privacy recall vs. strict full-match accuracy, on Android vs. PC screenshots","falsification":"With the released prompt and matching rules on the public split, a comparable set of VLMs gets average element recall within about 10 points of binary detection, or full-match accuracy above about 25% on Android, or PC is not clearly below Android on all three metrics.","caption":"Figure 5: Average privacy-recognition performance on Android and PC, comparing binary privacy detection, privacy recall rate, and strict full-match accuracy.","image":"img/2601.18842.png","kind":"figure","image_w":572,"image_h":477,"pwc":true}
site/p/2601.23163.json CHANGED
@@ -1 +1 @@
1
- {"id":"2601.23163","title":"Probing the Trajectories of Reasoning Traces in Large Language Models","authors":"Marthe Ballon; Brecht F Verbeken; Vincent Ginis; Andres Algaba","openreview_id":"SqFPT2YE5W","presentation":"poster","track":"evaluations_datasets","verdict":"include","background":"Reasoning LLMs emit a long 'thinking' trace before answering, and these traces improve benchmark accuracy. It is debated whether the gain comes from the specific content of the trace or from generic effects such as extra context length, a reasoning-like format, or the presence of relevant words. Separating these matters for deciding whether traces can be truncated, compressed or reused across models, and for how much to trust them in monitoring.","method":["For each question in GPQA Diamond (198 questions, options A-D) and MMLU-Pro (12,032 questions, options A-J), the model (Qwen3-4B/8B/14B, gpt-oss-20b/120b) is prompted in thinking mode with a system instruction to answer with a single letter, and its full reasoning trace is generated (Qwen3: temperature 0.6, top_p 0.95, top_k 20; gpt-oss: temperature 1, top_p 1, top_k 50). The trace is tokenized and truncated to the first ceil(d/100 * L) tokens for deciles d = 0 (empty thinking block), 10, ..., 100. Each prefix is re-injected into the same model's prompt followed by an early-stopping suffix (Qwen3: 'Considering the limited time by the user, I have to give the solution based on the thinking directly now.\\n</think>\\n\\n'; gpt-oss: close the analysis channel and open the final channel). One token is generated, and the answer is the argmax of the raw softmax next-token probabilities over the bare letter tokens.","Three length-matched controls replace the injected prefix: Random (uniformly random non-special tokens of the same token count), Trace-swap (the first k tokens of a different question's trace, sampled from traces at least as long), and Permutation/Shuffle (a random permutation of the original prefix's tokens, with the <think> marker kept in place). Accuracy at decile 100 is compared between the original trace and each control, pooled over 3 independent generation runs, with McNemar's test."],"claim":"At full reasoning (decile 100) on GPQA Diamond with Qwen3-4B, injecting the model's own trace gives 54.4% accuracy, beating length-matched controls by +19.4 points (random tokens), +20.5 (another question's trace) and +14.3 (shuffled own tokens); for all 5 models and both benchmarks, the original trace beats every control by at least 13.1 points (all p < 0.001, McNemar).","source":"Table A2 (Appendix), page 32; summarized in Figure 3, page 6","metric":"Multiple-choice accuracy at decile 100 (higher is better), argmax over answer-letter next-token probabilities, pooled over 3 runs (n = 594 question-run pairs on GPQA, 36,096 on MMLU-Pro). Table A2 reports no ± values; Table A1 gives run-to-run std of original-trace accuracy at decile 100 (Qwen3-4B GPQA: 54.4 ± 0.6; max 2.9 for gpt-oss-120b GPQA). Significance via McNemar's test.","comparison":"Original reasoning trace vs. Random-token, Trace-swap and Permutation controls of identical token length, same model, same prompt and early-stopping suffix","falsification":"For Qwen3-4B on GPQA Diamond (or any model/benchmark in the table), the original-trace accuracy at decile 100 exceeds one of the three length-matched controls by less than about 5 points, or the permutation control reaches accuracy comparable to the original trace.","caption":"Table A2: Summary statistics and significance tests. Gain shows accuracy improvement from decile 0 to 100. \"Original vs Control\" columns show the accuracy advantage of original traces over each control at decile 100. \"Control vs d=0\" columns show the accuracy change of each control relative to the no-reasoning baseline. n is the number of question-run pairs (pooled across 3 runs). Significance: * p < 0.05, ** p < 0.01, *** p < 0.001 (McNemar's test).","image":"img/2601.23163.png","kind":"table","image_w":1333,"image_h":450,"pwc":false}
 
1
+ {"id":"2601.23163","title":"Probing the Trajectories of Reasoning Traces in Large Language Models","authors":"Marthe Ballon; Brecht F Verbeken; Vincent Ginis; Andres Algaba","openreview_id":"SqFPT2YE5W","presentation":"poster","track":"evaluations_datasets","verdict":"include","background":"Reasoning LLMs emit a long 'thinking' trace before answering, and these traces improve benchmark accuracy. It is debated whether the gain comes from the specific content of the trace or from generic effects such as extra context length, a reasoning-like format, or the presence of relevant words. Separating these matters for deciding whether traces can be truncated, compressed or reused across models, and for how much to trust them in monitoring.","method":["For each question in GPQA Diamond (198 questions, options A-D) and MMLU-Pro (12,032 questions, options A-J), the model (Qwen3-4B/8B/14B, gpt-oss-20b/120b) is prompted in thinking mode with a system instruction to answer with a single letter, and its full reasoning trace is generated (Qwen3: temperature 0.6, top_p 0.95, top_k 20; gpt-oss: temperature 1, top_p 1, top_k 50). The trace is tokenized and truncated to the first ceil(d/100 * L) tokens for deciles d = 0 (empty thinking block), 10, ..., 100. Each prefix is re-injected into the same model's prompt followed by an early-stopping suffix (Qwen3: 'Considering the limited time by the user, I have to give the solution based on the thinking directly now.\\n</think>\\n\\n'; gpt-oss: close the analysis channel and open the final channel). One token is generated, and the answer is the argmax of the raw softmax next-token probabilities over the bare letter tokens.","Three length-matched controls replace the injected prefix: Random (uniformly random non-special tokens of the same token count), Trace-swap (the first k tokens of a different question's trace, sampled from traces at least as long), and Permutation/Shuffle (a random permutation of the original prefix's tokens, with the <think> marker kept in place). Accuracy at decile 100 is compared between the original trace and each control, pooled over 3 independent generation runs, with McNemar's test."],"claim":"At full reasoning (decile 100) on GPQA Diamond with Qwen3-4B, injecting the model's own trace gives 54.4% accuracy, beating length-matched controls by +19.4 points (random tokens), +20.5 (another question's trace) and +14.3 (shuffled own tokens); for all 5 models and both benchmarks, the original trace beats every control by at least 13.1 points (all p < 0.001, McNemar).","source":"Table A2 (Appendix), page 32; summarized in Figure 3, page 6","metric":"Multiple-choice accuracy at decile 100 (higher is better), argmax over answer-letter next-token probabilities, pooled over 3 runs (n = 594 question-run pairs on GPQA, 36,096 on MMLU-Pro). Table A2 reports no ± values; Table A1 gives run-to-run std of original-trace accuracy at decile 100 (Qwen3-4B GPQA: 54.4 ± 0.6; max 2.9 for gpt-oss-120b GPQA). Significance via McNemar's test.","comparison":"Original reasoning trace vs. Random-token, Trace-swap and Permutation controls of identical token length, same model, same prompt and early-stopping suffix","falsification":"For Qwen3-4B on GPQA Diamond (or any model/benchmark in the table), the original-trace accuracy at decile 100 exceeds one of the three length-matched controls by less than about 5 points, or the permutation control reaches accuracy comparable to the original trace.","caption":"Table A2: Summary statistics and significance tests. Gain shows accuracy improvement from decile 0 to 100. \"Original vs Control\" columns show the accuracy advantage of original traces over each control at decile 100. \"Control vs d=0\" columns show the accuracy change of each control relative to the no-reasoning baseline. n is the number of question-run pairs (pooled across 3 runs). Significance: * p < 0.05, ** p < 0.01, *** p < 0.001 (McNemar's test).","image":"img/2601.23163.png","kind":"table","image_w":1333,"image_h":450,"pwc":true}
site/p/2602.01493.json CHANGED
@@ -1 +1 @@
1
- {"id":"2602.01493","title":"OpInf-LLM: Parametric PDE Solving with LLMs via Operator Inference","authors":"Zhuoyuan \"Jacob\" Wang; Hanjiang Hu; Xiyu Deng; Saviz Mowlavi; Yorie Nakahira","openreview_id":"","presentation":"poster","track":"journal:TMLR","verdict":"include","background":"LLM-based PDE solving currently splits into two camps: LLMs that write numerical solver code (e.g. CodePDE), which often fails to run or is numerically inaccurate, and transformer PDE foundation models with language inputs (e.g. MOL-LLM), which always produce an output but need lots of training data and generalize poorly to unseen parameters and boundary conditions. Operator inference (OpInf) is a classical data-driven reduced-order modeling method that learns compact, interpretable reduced dynamics from a few simulated trajectories, but it is normally hand-configured for each PDE family and has no natural-language interface.","method":["Offline, for each PDE family the authors simulate a small number of training trajectories at a few parameter values (lid-driven cavity: Re in {50, 75, 100, 125, 150}, 8 trajectories each, 34x34 grid, T = 2, snapshots every 0.02, randomized sinusoidal lid velocity). A shared POD basis (r = 20 modes for cavity) is computed from all snapshots, and for each training parameter a quadratic reduced model da/dt = A a + H (a kron a) + B u + c is fit by Tikhonov-regularized least squares (lambda = 3). Training totals 449 trajectories across three PDEs and about 30 s.","Online, an LLM (GPT-4.1, GPT-4o or Gemini-2.0-flash) reads a natural-language task description, parses the equation and parameter, and either calls fixed tools or writes code to regress the reduced operators to the unseen parameter (polynomial regression over the stored operators), integrate the reduced ODE with the given boundary input, and reconstruct the full field. Evaluation is average relative L2 error over [0, T] on 2 test trajectories at each unseen Re in {60, 80, 90, 110, 120, 140} (12 total). Baselines on identical data: CodePDE (5 generation rounds x 5 debug trials, best solver reported), MOL-LLM trained on the same 449 trajectories and on 15,000 trajectories, and direct LLM prediction of the field."],"claim":"On 2D lid-driven cavity flow at unseen Reynolds numbers, OpInf-LLM (tool calling) reaches relative L2 error 4.63e-2, versus 6.44e-1 for MOL-LLM trained on 15,000 trajectories, 7.24e-1 for MOL-LLM on the same 449 trajectories, and 1.41 for the best CodePDE solver (GPT-4.1).","source":"Table 1, page 9","metric":"Average relative L2 error of the reconstructed vorticity field over [0, T] on 12 test trajectories (lower is better); no error bars, seeds or spread are reported","comparison":"OpInf-LLM (tool, any of the three LLMs; identical result) vs. MOL-LLM (449 and 15,000 training trajectories), CodePDE (GPT-4.1 / GPT-4o / Gemini-2.0-flash) and direct LLM prediction, all on the same cavity test set","falsification":"Re-running OpInf with the paper's settings (r = 20, lambda = 3, operator regression across Re) on the cavity test set gives relative L2 error that is not clearly below MOL-LLM's ~0.64 (e.g. above ~0.2), or a retrained MOL-LLM or a CodePDE solver reaches error within a factor of ~2 of OpInf-LLM.","caption":"Table 1: Results summary. Average relative L2 error is reported for each equation. Bold denotes the best result, and underline denotes the second best. ↑ higher is better, ↓ lower is better.","image":"img/2602.01493.png","kind":"table","image_w":1333,"image_h":633,"pwc":false}
 
1
+ {"id":"2602.01493","title":"OpInf-LLM: Parametric PDE Solving with LLMs via Operator Inference","authors":"Zhuoyuan \"Jacob\" Wang; Hanjiang Hu; Xiyu Deng; Saviz Mowlavi; Yorie Nakahira","openreview_id":"","presentation":"poster","track":"journal:TMLR","verdict":"include","background":"LLM-based PDE solving currently splits into two camps: LLMs that write numerical solver code (e.g. CodePDE), which often fails to run or is numerically inaccurate, and transformer PDE foundation models with language inputs (e.g. MOL-LLM), which always produce an output but need lots of training data and generalize poorly to unseen parameters and boundary conditions. Operator inference (OpInf) is a classical data-driven reduced-order modeling method that learns compact, interpretable reduced dynamics from a few simulated trajectories, but it is normally hand-configured for each PDE family and has no natural-language interface.","method":["Offline, for each PDE family the authors simulate a small number of training trajectories at a few parameter values (lid-driven cavity: Re in {50, 75, 100, 125, 150}, 8 trajectories each, 34x34 grid, T = 2, snapshots every 0.02, randomized sinusoidal lid velocity). A shared POD basis (r = 20 modes for cavity) is computed from all snapshots, and for each training parameter a quadratic reduced model da/dt = A a + H (a kron a) + B u + c is fit by Tikhonov-regularized least squares (lambda = 3). Training totals 449 trajectories across three PDEs and about 30 s.","Online, an LLM (GPT-4.1, GPT-4o or Gemini-2.0-flash) reads a natural-language task description, parses the equation and parameter, and either calls fixed tools or writes code to regress the reduced operators to the unseen parameter (polynomial regression over the stored operators), integrate the reduced ODE with the given boundary input, and reconstruct the full field. Evaluation is average relative L2 error over [0, T] on 2 test trajectories at each unseen Re in {60, 80, 90, 110, 120, 140} (12 total). Baselines on identical data: CodePDE (5 generation rounds x 5 debug trials, best solver reported), MOL-LLM trained on the same 449 trajectories and on 15,000 trajectories, and direct LLM prediction of the field."],"claim":"On 2D lid-driven cavity flow at unseen Reynolds numbers, OpInf-LLM (tool calling) reaches relative L2 error 4.63e-2, versus 6.44e-1 for MOL-LLM trained on 15,000 trajectories, 7.24e-1 for MOL-LLM on the same 449 trajectories, and 1.41 for the best CodePDE solver (GPT-4.1).","source":"Table 1, page 9","metric":"Average relative L2 error of the reconstructed vorticity field over [0, T] on 12 test trajectories (lower is better); no error bars, seeds or spread are reported","comparison":"OpInf-LLM (tool, any of the three LLMs; identical result) vs. MOL-LLM (449 and 15,000 training trajectories), CodePDE (GPT-4.1 / GPT-4o / Gemini-2.0-flash) and direct LLM prediction, all on the same cavity test set","falsification":"Re-running OpInf with the paper's settings (r = 20, lambda = 3, operator regression across Re) on the cavity test set gives relative L2 error that is not clearly below MOL-LLM's ~0.64 (e.g. above ~0.2), or a retrained MOL-LLM or a CodePDE solver reaches error within a factor of ~2 of OpInf-LLM.","caption":"Table 1: Results summary. Average relative L2 error is reported for each equation. Bold denotes the best result, and underline denotes the second best. ↑ higher is better, ↓ lower is better.","image":"img/2602.01493.png","kind":"table","image_w":1333,"image_h":633,"pwc":true}
site/p/2602.02304.json CHANGED
@@ -1 +1 @@
1
- {"id":"2602.02304","title":"Comparing Explanations is not Enough,Explain the Change: New Standards are Needed to Explain Behavioral Shifts in Large Language Models","authors":"Martino Ciaperoni; Marzio Di Vece; Roberto Pellungrini; Luca Pappalardo; Fosca Giannotti; Francesco Giannini","openreview_id":"y0jWumzUxx","presentation":"poster","track":"position","verdict":"borderline","background":"LLMs often change behavior after interventions such as fine-tuning, RLHF or scaling, sometimes in harmful ways (for example, emergent misalignment after narrow fine-tuning). Standard explainability tools explain a single checkpoint, so they cannot tell which features or mechanisms are new after the intervention and which were already present. The paper argues that the shift between a reference model and an updated model should itself be what gets explained, proposes 'Comparative XAI' (XAI∆) with ten desiderata, and illustrates it on a model fine-tuned to give bad medical advice.","method":["Following the emergent-misalignment setup of Turner et al., the reference model M_pre is instruction-tuned Qwen2.5-0.5B. M_post is a LoRA fine-tune of it on a narrow dataset of incorrect/harmful medical advice (minimizing symptoms, discouraging timely care). A placebo model M_placebo is fine-tuned with the identical LoRA configuration on correct medical advice, so M_post and M_placebo differ only in the quality of the training advice.","Behavior is measured with a binary metric B(M, x): whether the response recommends urgent care, detected by keyword matching against a fixed list of urgency phrases with negation handling. It is evaluated on the 93 prompts (out of 150 medical-advice prompts) that describe urgent clinical scenarios. Comparisons are paired by prompt, using paired-bootstrap 95% CIs (10,000 resamples) and Holm-corrected one-sided exact McNemar tests. The paper sets a shift threshold of ε_B = 10 points."],"claim":"On 93 urgent clinical prompts, LoRA fine-tuning Qwen2.5-0.5B-Instruct on bad medical advice lowers the urgent-care recommendation rate from 29.0% to 10.8% (Δ = −18.3 pts, 95% CI [−29.0, −8.6], McNemar p = 1.5e-3), while an identically configured placebo fine-tune on correct medical advice gives no significant change (36.6%, Δ = +7.5, CI [−4.3, 18.3]).","source":"Table 1, page 7","metric":"Urgent-care recommendation rate B (%) over 93 urgent prompts, keyword-matched; higher means more urgent-care advice. Intervals are paired-bootstrap 95% CIs over prompts (not over fine-tuning seeds); p-values are Holm-corrected one-sided exact McNemar tests.","comparison":"M_post (fine-tuned on bad medical advice) vs. M_pre (base Qwen2.5-0.5B-Instruct), with M_placebo (same LoRA config, correct medical advice) vs. M_pre as the specificity control","falsification":"The bad-advice fine-tune does not lower B by a significant amount (the CI for Δ includes 0, or the drop is under the 10-point threshold), or the placebo fine-tune lowers B by a similar amount.","caption":"Table 1: Medical fine-tuning reduces urgent-care recommendations. B denotes the urgent-care recommendation rate, while ∆ measures its change from the reference (percentage points). We report paired-bootstrap 95% confidence intervals and Holm-corrected one-sided exact McNemar p-values.","image":"img/2602.02304.png","kind":"table","image_w":1144,"image_h":250,"pwc":false}
 
1
+ {"id":"2602.02304","title":"Comparing Explanations is not Enough,Explain the Change: New Standards are Needed to Explain Behavioral Shifts in Large Language Models","authors":"Martino Ciaperoni; Marzio Di Vece; Roberto Pellungrini; Luca Pappalardo; Fosca Giannotti; Francesco Giannini","openreview_id":"y0jWumzUxx","presentation":"poster","track":"position","verdict":"borderline","background":"LLMs often change behavior after interventions such as fine-tuning, RLHF or scaling, sometimes in harmful ways (for example, emergent misalignment after narrow fine-tuning). Standard explainability tools explain a single checkpoint, so they cannot tell which features or mechanisms are new after the intervention and which were already present. The paper argues that the shift between a reference model and an updated model should itself be what gets explained, proposes 'Comparative XAI' (XAI∆) with ten desiderata, and illustrates it on a model fine-tuned to give bad medical advice.","method":["Following the emergent-misalignment setup of Turner et al., the reference model M_pre is instruction-tuned Qwen2.5-0.5B. M_post is a LoRA fine-tune of it on a narrow dataset of incorrect/harmful medical advice (minimizing symptoms, discouraging timely care). A placebo model M_placebo is fine-tuned with the identical LoRA configuration on correct medical advice, so M_post and M_placebo differ only in the quality of the training advice.","Behavior is measured with a binary metric B(M, x): whether the response recommends urgent care, detected by keyword matching against a fixed list of urgency phrases with negation handling. It is evaluated on the 93 prompts (out of 150 medical-advice prompts) that describe urgent clinical scenarios. Comparisons are paired by prompt, using paired-bootstrap 95% CIs (10,000 resamples) and Holm-corrected one-sided exact McNemar tests. The paper sets a shift threshold of ε_B = 10 points."],"claim":"On 93 urgent clinical prompts, LoRA fine-tuning Qwen2.5-0.5B-Instruct on bad medical advice lowers the urgent-care recommendation rate from 29.0% to 10.8% (Δ = −18.3 pts, 95% CI [−29.0, −8.6], McNemar p = 1.5e-3), while an identically configured placebo fine-tune on correct medical advice gives no significant change (36.6%, Δ = +7.5, CI [−4.3, 18.3]).","source":"Table 1, page 7","metric":"Urgent-care recommendation rate B (%) over 93 urgent prompts, keyword-matched; higher means more urgent-care advice. Intervals are paired-bootstrap 95% CIs over prompts (not over fine-tuning seeds); p-values are Holm-corrected one-sided exact McNemar tests.","comparison":"M_post (fine-tuned on bad medical advice) vs. M_pre (base Qwen2.5-0.5B-Instruct), with M_placebo (same LoRA config, correct medical advice) vs. M_pre as the specificity control","falsification":"The bad-advice fine-tune does not lower B by a significant amount (the CI for Δ includes 0, or the drop is under the 10-point threshold), or the placebo fine-tune lowers B by a similar amount.","caption":"Table 1: Medical fine-tuning reduces urgent-care recommendations. B denotes the urgent-care recommendation rate, while ∆ measures its change from the reference (percentage points). We report paired-bootstrap 95% confidence intervals and Holm-corrected one-sided exact McNemar p-values.","image":"img/2602.02304.png","kind":"table","image_w":1144,"image_h":250,"pwc":true}
site/p/2602.07603.json CHANGED
@@ -1 +1 @@
1
- {"id":"2602.07603","title":"Closed-Form Implicit Neural Representations","authors":"Woojin Cho; Junghwan Park","openreview_id":"xgO3o4J75E","presentation":"poster","track":"main","verdict":"include","background":"Implicit neural representations (INRs) store a signal such as an image as a network that maps coordinates to values. Fitting one usually needs thousands of steps of gradient descent per signal, and MLPs fit high-frequency detail slowly (spectral bias), so fine textures tend to stay blurred. A closed-form fit that captures fine detail quickly would make INRs cheaper to use for images, satellite data and scientific fields.","method":["ELM-INR splits the image into overlapping square subdomains (default S_i = 32x32 = 1024 pixels each). In each subdomain, local pixel coordinates are normalized to [-1, 1], encoded with random Fourier features (F = 10, sigma = 1, as in FFN), and passed through one random, frozen hidden layer of m = 1024 ReLU units (an Extreme Learning Machine). Only the output weights are fitted, by a closed-form least-squares solve per subdomain, batched on the GPU. The local predictors are blended with partition-of-unity window functions into one global reconstruction. BEAM (adaptive partitioning) is not used for this result.","Baselines are SIREN, FFN, GaussNet and WIRE, trained with Adam for 1000 steps on the standard images, with the best configuration from a search over width {256, 512, 1024}, depth {3, 4, 5} and learning rate {5e-5, 1e-4}. PSNR is computed on the training pixels after normalizing by the ground-truth maximum, as -10 log10(MSE + 1e-8), which caps PSNR at 80 dB. Kodak images kodim05, kodim08 and kodim24 are 512x512 RGB and kodim20 is 256x256."],"claim":"On four Kodak images, closed-form ELM-INR (32x32 subdomains, m = 1024) reconstructs the image at 78.4-79.1 dB PSNR (e.g. kodim05: 79.03 ± 1.59), while the best of SIREN, FFN, GaussNet and WIRE reaches only 30.4-40.1 dB (kodim05: SIREN 31.12 ± 0.22), a gap of more than 38 dB on every image.","source":"Table 2, page 15 (Appendix B)","metric":"PSNR in dB on the fitted image, higher is better; mean ± std over 5 random seeds; PSNR = -10 log10(MSE + 1e-8) after normalizing by the ground-truth maximum, so 80 dB is the ceiling","comparison":"ELM-INR vs. SIREN, FFN, GaussNet and WIRE (Adam, 1000 steps, best of a small hyperparameter search) on the same images","falsification":"With the stated configuration (S_i = 32^2, m = 1024, RFF F = 10 sigma = 1, ReLU, partition-of-unity blending), ELM-INR fails to exceed about 60 dB PSNR on the Kodak images, or a tuned SIREN/WIRE baseline under the stated protocol comes within about 10 dB of it.","caption":"Table 2: All benchmarks are reported as mean ± std over 5 random seeds. PSNR is averaged over all spectral bands for the multispectral benchmarks.","image":"img/2602.07603.png","kind":"table","image_w":1138,"image_h":361,"pwc":false}
 
1
+ {"id":"2602.07603","title":"Closed-Form Implicit Neural Representations","authors":"Woojin Cho; Junghwan Park","openreview_id":"xgO3o4J75E","presentation":"poster","track":"main","verdict":"include","background":"Implicit neural representations (INRs) store a signal such as an image as a network that maps coordinates to values. Fitting one usually needs thousands of steps of gradient descent per signal, and MLPs fit high-frequency detail slowly (spectral bias), so fine textures tend to stay blurred. A closed-form fit that captures fine detail quickly would make INRs cheaper to use for images, satellite data and scientific fields.","method":["ELM-INR splits the image into overlapping square subdomains (default S_i = 32x32 = 1024 pixels each). In each subdomain, local pixel coordinates are normalized to [-1, 1], encoded with random Fourier features (F = 10, sigma = 1, as in FFN), and passed through one random, frozen hidden layer of m = 1024 ReLU units (an Extreme Learning Machine). Only the output weights are fitted, by a closed-form least-squares solve per subdomain, batched on the GPU. The local predictors are blended with partition-of-unity window functions into one global reconstruction. BEAM (adaptive partitioning) is not used for this result.","Baselines are SIREN, FFN, GaussNet and WIRE, trained with Adam for 1000 steps on the standard images, with the best configuration from a search over width {256, 512, 1024}, depth {3, 4, 5} and learning rate {5e-5, 1e-4}. PSNR is computed on the training pixels after normalizing by the ground-truth maximum, as -10 log10(MSE + 1e-8), which caps PSNR at 80 dB. Kodak images kodim05, kodim08 and kodim24 are 512x512 RGB and kodim20 is 256x256."],"claim":"On four Kodak images, closed-form ELM-INR (32x32 subdomains, m = 1024) reconstructs the image at 78.4-79.1 dB PSNR (e.g. kodim05: 79.03 ± 1.59), while the best of SIREN, FFN, GaussNet and WIRE reaches only 30.4-40.1 dB (kodim05: SIREN 31.12 ± 0.22), a gap of more than 38 dB on every image.","source":"Table 2, page 15 (Appendix B)","metric":"PSNR in dB on the fitted image, higher is better; mean ± std over 5 random seeds; PSNR = -10 log10(MSE + 1e-8) after normalizing by the ground-truth maximum, so 80 dB is the ceiling","comparison":"ELM-INR vs. SIREN, FFN, GaussNet and WIRE (Adam, 1000 steps, best of a small hyperparameter search) on the same images","falsification":"With the stated configuration (S_i = 32^2, m = 1024, RFF F = 10 sigma = 1, ReLU, partition-of-unity blending), ELM-INR fails to exceed about 60 dB PSNR on the Kodak images, or a tuned SIREN/WIRE baseline under the stated protocol comes within about 10 dB of it.","caption":"Table 2: All benchmarks are reported as mean ± std over 5 random seeds. PSNR is averaged over all spectral bands for the multispectral benchmarks.","image":"img/2602.07603.png","kind":"table","image_w":1138,"image_h":361,"pwc":true}
site/p/2602.10530.json CHANGED
@@ -1 +1 @@
1
- {"id":"2602.10530","title":"Generalized Robust Adaptive-Bandwidth Multi-View Manifold Learning in High Dimensions with Noise","authors":"Xiucai Ding; Chao Shen; Hau-Tieng Wu","openreview_id":"","presentation":"poster","track":"journal:AoS","verdict":"include","background":"Many scientific datasets record the same underlying system through several sensors ('views'), each with its own dimension, nonlinear distortion and noise level. Kernel methods such as diffusion maps can recover low-dimensional structure from one view, and multiview diffusion maps (MDM) fuse several views. But these methods choose the kernel bandwidth heuristically, assume clean data, and have little theory once there are more than two views or the noise is high-dimensional. A poor bandwidth in one noisy view can corrupt the fused embedding, so a principled per-view bandwidth rule could make multiview fusion much more robust.","method":["GRAB-MDM builds, for each view l, a Gaussian kernel matrix K^l with bandwidth eps_l. It forms a block matrix whose (l1, l2) off-diagonal block is K^l1 K^l2 and whose diagonal blocks are zero, so the random walk can only move across views, and row-normalizes it into a transition matrix A. Each sample is embedded by averaging its top non-trivial eigenvector coordinates (t = 1) across its K view copies (eq. 8). The bandwidth is eps_l = c * h_l (Algorithm 1). Here h_l is the omega_l-percentile of the pairwise squared distances in view l, after per-coordinate z-scoring. The global factor c comes from a grid search over [1e-3, 0.5] that picks the c whose spectrum of A is most stable, i.e. the c with the most other grid values whose squared eigenvalue distance is below a threshold delta. The embedding dimension m is chosen by an eigen-ratio elbow rule (m = 1 for this experiment), and omega_l and delta are chosen by procedures in Appendix E.","For Table 1, latent vectors w in R^{3d} (d = 10) for three clusters are drawn from uniform boxes (Setup 1) or truncated Gaussians (Setup 2), with cluster-specific parameters given in Section 6.1. View l takes block l of w, applies g1(x) = x^2 + x, g2(x) = 10 log(x + 2) or g3(x) = 0.8x entrywise, zero-pads to p_l = 100 dimensions, and adds isotropic Gaussian noise of variance v_l^2. With n = 200 and 3 views, each method's embedding is clustered into 3 clusters with k-means and scored by clustering accuracy (ACC). Baselines are mDM, mSNE, mLLE, mISOMAP, mUMAP and pairwise alternating diffusion maps (pADM), plus single-view DM, LLE, t-SNE, ISOMAP and UMAP on the concatenated views."],"claim":"In Setup (1) with heterogeneous high noise, GRAB-MDM reaches clustering accuracy 0.81 (0.02), 0.79 (0.02) and 0.77 (0.02) at noise variances (20, 10, 45), (10, 25, 40) and (10, 30, 50). The strongest baseline, multiview diffusion maps (mDM), gets 0.73, 0.72 and 0.70 (std 0.02), and every other baseline gets at most 0.67.","source":"Table 1, page 21","metric":"Clustering accuracy (ACC) of k-means on the embedding, higher is better; mean with standard deviation over 1,000 data replications in parentheses (per-replication spread, not standard error)","comparison":"GRAB-MDM (adaptive per-view bandwidths, zero within-view blocks, averaged embedding) vs. mDM, mSNE, mLLE, mISOMAP, mUMAP and pADM, and vs. DM, LLE, t-SNE, ISOMAP and UMAP on the concatenated views, all on the same simulated 3-view data","falsification":"In Setup (1) at noise variances (20, 10, 45), (10, 25, 40) or (10, 30, 50), a faithful reimplementation of GRAB-MDM does not beat mDM by more than about one per-replication std (0.02), or some other baseline matches or beats GRAB-MDM.","caption":"TABLE 1. Comparison of clustering accuracy across various methods. The reported values are the means with standard deviations (in parentheses), based on 1,000 replications.","image":"img/2602.10530.png","kind":"table","image_w":1138,"image_h":633,"pwc":false}
 
1
+ {"id":"2602.10530","title":"Generalized Robust Adaptive-Bandwidth Multi-View Manifold Learning in High Dimensions with Noise","authors":"Xiucai Ding; Chao Shen; Hau-Tieng Wu","openreview_id":"","presentation":"poster","track":"journal:AoS","verdict":"include","background":"Many scientific datasets record the same underlying system through several sensors ('views'), each with its own dimension, nonlinear distortion and noise level. Kernel methods such as diffusion maps can recover low-dimensional structure from one view, and multiview diffusion maps (MDM) fuse several views. But these methods choose the kernel bandwidth heuristically, assume clean data, and have little theory once there are more than two views or the noise is high-dimensional. A poor bandwidth in one noisy view can corrupt the fused embedding, so a principled per-view bandwidth rule could make multiview fusion much more robust.","method":["GRAB-MDM builds, for each view l, a Gaussian kernel matrix K^l with bandwidth eps_l. It forms a block matrix whose (l1, l2) off-diagonal block is K^l1 K^l2 and whose diagonal blocks are zero, so the random walk can only move across views, and row-normalizes it into a transition matrix A. Each sample is embedded by averaging its top non-trivial eigenvector coordinates (t = 1) across its K view copies (eq. 8). The bandwidth is eps_l = c * h_l (Algorithm 1). Here h_l is the omega_l-percentile of the pairwise squared distances in view l, after per-coordinate z-scoring. The global factor c comes from a grid search over [1e-3, 0.5] that picks the c whose spectrum of A is most stable, i.e. the c with the most other grid values whose squared eigenvalue distance is below a threshold delta. The embedding dimension m is chosen by an eigen-ratio elbow rule (m = 1 for this experiment), and omega_l and delta are chosen by procedures in Appendix E.","For Table 1, latent vectors w in R^{3d} (d = 10) for three clusters are drawn from uniform boxes (Setup 1) or truncated Gaussians (Setup 2), with cluster-specific parameters given in Section 6.1. View l takes block l of w, applies g1(x) = x^2 + x, g2(x) = 10 log(x + 2) or g3(x) = 0.8x entrywise, zero-pads to p_l = 100 dimensions, and adds isotropic Gaussian noise of variance v_l^2. With n = 200 and 3 views, each method's embedding is clustered into 3 clusters with k-means and scored by clustering accuracy (ACC). Baselines are mDM, mSNE, mLLE, mISOMAP, mUMAP and pairwise alternating diffusion maps (pADM), plus single-view DM, LLE, t-SNE, ISOMAP and UMAP on the concatenated views."],"claim":"In Setup (1) with heterogeneous high noise, GRAB-MDM reaches clustering accuracy 0.81 (0.02), 0.79 (0.02) and 0.77 (0.02) at noise variances (20, 10, 45), (10, 25, 40) and (10, 30, 50). The strongest baseline, multiview diffusion maps (mDM), gets 0.73, 0.72 and 0.70 (std 0.02), and every other baseline gets at most 0.67.","source":"Table 1, page 21","metric":"Clustering accuracy (ACC) of k-means on the embedding, higher is better; mean with standard deviation over 1,000 data replications in parentheses (per-replication spread, not standard error)","comparison":"GRAB-MDM (adaptive per-view bandwidths, zero within-view blocks, averaged embedding) vs. mDM, mSNE, mLLE, mISOMAP, mUMAP and pADM, and vs. DM, LLE, t-SNE, ISOMAP and UMAP on the concatenated views, all on the same simulated 3-view data","falsification":"In Setup (1) at noise variances (20, 10, 45), (10, 25, 40) or (10, 30, 50), a faithful reimplementation of GRAB-MDM does not beat mDM by more than about one per-replication std (0.02), or some other baseline matches or beats GRAB-MDM.","caption":"TABLE 1. Comparison of clustering accuracy across various methods. The reported values are the means with standard deviations (in parentheses), based on 1,000 replications.","image":"img/2602.10530.png","kind":"table","image_w":1138,"image_h":633,"pwc":true}
site/p/2602.12517.json CHANGED
@@ -1 +1 @@
1
- {"id":"2602.12517","title":"Bench-MFG: A Benchmark Suite for Learning in Stationary Mean Field Games","authors":"Lorenzo Magnino; Jiacheng Shen; Matthieu Geist; Olivier Pietquin; Mathieu Lauriere","openreview_id":"8CjE3pJACr","presentation":"poster","track":"evaluations_datasets","verdict":"include","background":"Mean Field Games (MFGs) approximate very large multi-agent systems by having one representative agent interact with the population distribution. The solution concept is a mean field Nash equilibrium, and how far a policy is from one is measured by its exploitability. Many learning algorithms for MFGs have been proposed (fixed-point iteration, fictitious play, policy-iteration variants, online mirror descent), but each is usually tested on a few bespoke toy environments. This makes it hard to tell which methods are robust. Bench-MFG offers a taxonomy of stationary discrete MFG classes with example environments, plus MF-Garnets, a random generator of MFG instances meant to support statistical comparisons between solvers.","method":["An MF-Garnet instance is specified by (|X|, |A|, branching factor, dynamics structure, reward structure). It has a sparse random base transition tensor P0 and Gaussian coupling coefficients Omega, which make the transitions depend on the mean field mu either additively (cp*P0 + rho_p*sum_y Omega*mu(y), then ReLU and normalize) or multiplicatively (P0 times ReLU(cp + rho_p*sum_y Omega*mu(y)), then normalize). The reward is a random base reward R0 combined with an interaction matrix M, again either additively (cr*R0 + rho_r*M mu) or multiplicatively (R0*(cr + rho_r*M mu)). Normalization and the finite noise grid used for sampling are defined in Appx. C.1.","Eight exact, model-based solvers are run on each instance: Fixed Point, Damped FP, Fictitious Play, Policy Iteration, Smooth PI, Boltzmann PI, Online Mirror Descent, and the paper's new MF-PSO (particle swarm search that minimizes exploitability, 200 particles). All methods start from the same initial policy: logits drawn uniformly from [-1, 1], passed through a Boltzmann policy with temperature 0.2. Hyperparameters are grid-searched per algorithm (Appx. H). For each algorithm and Garnet type, the final exploitability is averaged over 10 instances (Garnet seeds 0-9) and reported as mean ± std. The implementation is in JAX on an NVIDIA L40, and a pure-Python reference version is also provided."],"claim":"On 25-state, 10-action, branching-factor-10 MF-Garnets, Fixed Point, Damped FP, Fictitious Play, PI and Smooth PI all reach mean final exploitability of 0.003 or less in both the additive/additive and the multiplicative-dynamics/additive-reward settings. Examples: Damped FP 8.51e-04 ± 0.0017 (A/A) and 3.43e-05 ± 9.83e-05 (M/A). In the same settings, OMD (1.44 ± 0.39, 1.47 ± 0.45), Boltzmann PI (0.95 ± 0.29, 0.93 ± 0.32) and MF-PSO (3.86 ± 1.90, 4.02 ± 2.16) stay far from equilibrium.","source":"Table 2, page 9","metric":"Final exploitability (lower is better), mean ± std over 10 MF-Garnet instances (Garnet seeds 0-9)","comparison":"FP-family and greedy PI-family solvers (Fixed Point, Damped FP, Fictitious Play, PI, Smooth PI) vs. OMD, Boltzmann PI and MF-PSO, on the same randomly generated 25x10x10 MF-Garnet instances","falsification":"On 25x10x10 A/A or M/A MF-Garnets (10 instances), OMD, Boltzmann PI or MF-PSO with tuned hyperparameters reaches mean final exploitability within one std of the best FP-family method. Another refutation: FP, Damped FP or Fictitious Play fail to get below about 0.01 mean exploitability.","caption":"Table 2: MF-Garnet Results. Each MF-Garnet has 5 parameters |X| x |A| x Branching Factor (Dynamics/Reward Structure). We highlighted the three best exploitability values and the worst three standard deviations per game type. MF-Garnets show that different algorithms can perform significantly differently depending on the type of game structure and the dimensionality.","image":"img/2602.12517.png","kind":"table","image_w":1144,"image_h":411,"pwc":false}
 
1
+ {"id":"2602.12517","title":"Bench-MFG: A Benchmark Suite for Learning in Stationary Mean Field Games","authors":"Lorenzo Magnino; Jiacheng Shen; Matthieu Geist; Olivier Pietquin; Mathieu Lauriere","openreview_id":"8CjE3pJACr","presentation":"poster","track":"evaluations_datasets","verdict":"include","background":"Mean Field Games (MFGs) approximate very large multi-agent systems by having one representative agent interact with the population distribution. The solution concept is a mean field Nash equilibrium, and how far a policy is from one is measured by its exploitability. Many learning algorithms for MFGs have been proposed (fixed-point iteration, fictitious play, policy-iteration variants, online mirror descent), but each is usually tested on a few bespoke toy environments. This makes it hard to tell which methods are robust. Bench-MFG offers a taxonomy of stationary discrete MFG classes with example environments, plus MF-Garnets, a random generator of MFG instances meant to support statistical comparisons between solvers.","method":["An MF-Garnet instance is specified by (|X|, |A|, branching factor, dynamics structure, reward structure). It has a sparse random base transition tensor P0 and Gaussian coupling coefficients Omega, which make the transitions depend on the mean field mu either additively (cp*P0 + rho_p*sum_y Omega*mu(y), then ReLU and normalize) or multiplicatively (P0 times ReLU(cp + rho_p*sum_y Omega*mu(y)), then normalize). The reward is a random base reward R0 combined with an interaction matrix M, again either additively (cr*R0 + rho_r*M mu) or multiplicatively (R0*(cr + rho_r*M mu)). Normalization and the finite noise grid used for sampling are defined in Appx. C.1.","Eight exact, model-based solvers are run on each instance: Fixed Point, Damped FP, Fictitious Play, Policy Iteration, Smooth PI, Boltzmann PI, Online Mirror Descent, and the paper's new MF-PSO (particle swarm search that minimizes exploitability, 200 particles). All methods start from the same initial policy: logits drawn uniformly from [-1, 1], passed through a Boltzmann policy with temperature 0.2. Hyperparameters are grid-searched per algorithm (Appx. H). For each algorithm and Garnet type, the final exploitability is averaged over 10 instances (Garnet seeds 0-9) and reported as mean ± std. The implementation is in JAX on an NVIDIA L40, and a pure-Python reference version is also provided."],"claim":"On 25-state, 10-action, branching-factor-10 MF-Garnets, Fixed Point, Damped FP, Fictitious Play, PI and Smooth PI all reach mean final exploitability of 0.003 or less in both the additive/additive and the multiplicative-dynamics/additive-reward settings. Examples: Damped FP 8.51e-04 ± 0.0017 (A/A) and 3.43e-05 ± 9.83e-05 (M/A). In the same settings, OMD (1.44 ± 0.39, 1.47 ± 0.45), Boltzmann PI (0.95 ± 0.29, 0.93 ± 0.32) and MF-PSO (3.86 ± 1.90, 4.02 ± 2.16) stay far from equilibrium.","source":"Table 2, page 9","metric":"Final exploitability (lower is better), mean ± std over 10 MF-Garnet instances (Garnet seeds 0-9)","comparison":"FP-family and greedy PI-family solvers (Fixed Point, Damped FP, Fictitious Play, PI, Smooth PI) vs. OMD, Boltzmann PI and MF-PSO, on the same randomly generated 25x10x10 MF-Garnet instances","falsification":"On 25x10x10 A/A or M/A MF-Garnets (10 instances), OMD, Boltzmann PI or MF-PSO with tuned hyperparameters reaches mean final exploitability within one std of the best FP-family method. Another refutation: FP, Damped FP or Fictitious Play fail to get below about 0.01 mean exploitability.","caption":"Table 2: MF-Garnet Results. Each MF-Garnet has 5 parameters |X| x |A| x Branching Factor (Dynamics/Reward Structure). We highlighted the three best exploitability values and the worst three standard deviations per game type. MF-Garnets show that different algorithms can perform significantly differently depending on the type of game structure and the dimensionality.","image":"img/2602.12517.png","kind":"table","image_w":1144,"image_h":411,"pwc":true}
site/p/2602.18958.json CHANGED
@@ -1 +1 @@
1
- {"id":"2602.18958","title":"Structure-Adaptive Estimation of Heterogeneous Treatment Effects with Kernel Methods","authors":"Seok-Jin Kim","openreview_id":"B6if6B55a9","presentation":"poster","track":"main","verdict":"borderline","background":"Estimating the conditional average treatment effect (CATE) from observational data is hard when the baseline outcome functions are complex and when overlap is weak, meaning the propensity score gets close to 0 or 1. Doubly robust and R-learner style methods depend on estimating the propensity score, and inverse-propensity weighting becomes unstable under weak overlap. The CATE itself is often much simpler (smoother, sparse or low-dimensional) than the nuisance outcome functions. The paper asks whether a kernel method can reach error rates set by the simpler CATE class without estimating the propensity score.","method":["Algorithm 1 ('Ours') fits each treatment arm's outcome function with undersmoothed kernel ridge regression (KRR) in a nuisance RKHS F. It then forms switch-imputed pseudo-outcomes: m_i = y_i - f0_hat(x_i) for treated units and m_i = f1_hat(x_i) - y_i for control units. Finally it runs a second KRR of m_i on x_i in a simpler target kernel space H. No propensity score is estimated. In the experiments, the raw nuisance ridge is fixed at 0.01. The target kernel and ridge are chosen by Algorithm 2: candidates are fit on one split, nuisances are refit on a second split, candidates are scored against switch-imputed proxies on a third split, and predictions are averaged over the three cyclic fits.","Multivariate synthetic setup (Appendix C.2): X ~ Unif([-1,1]^d), f0(x) = (2/d) sum_{j=1}^d sin(x_j), h(x) = (0.3/p) sum_{j=1}^p x_j^2, A ~ Bernoulli(clip(sin(5||x||_2), kappa, 1-kappa)), Y = f0(X) + A h(X) + N(0, sigma^2), with defaults sigma = 1 and kappa = 0.1. The nuisance kernel is Matern nu = 1.5. Ours selects among Matern 1.5, Matern 2.5 and RBF target kernels, each on all coordinates or on x_{1:4}. Baselines are plug-in KRR (T-learner, 3-fold CV), the DR-learner (Kennedy 2023), oracle-propensity DR and the R-learner (Nie and Wager 2021). DR, oracle DR and R use Matern 2.5 target kernels with cross-fitting and receive the true support in sparse settings. Estimated propensities are clipped to [0.01, 0.99]. The metric is MISE on 3,000 held-out uniform test points."],"claim":"In the four multivariate sample-size settings ((d,p) = (10,10) and (10,4), n = 1000 and 2000, sigma = 1, kappa = 0.1, 100 replications), the two-stage switch-imputation KRR estimator has the lowest CATE MISE every time. For example, at (10,10), n = 1000 it gets 0.0099 (SE 0.0008), against 0.0127 (0.0012) for the R-learner, 0.0184 (0.0049) for DR, 0.0201 (0.0031) for oracle DR and 0.0478 (0.0018) for plug-in KRR.","source":"Table 2, page 18 (Appendix C.2)","metric":"Mean integrated squared error of the CATE estimate over 3,000 test points (lower is better); the value in parentheses is the Monte Carlo standard error of the mean over replications (100 for the sample-size rows, 30 for the sweeps)","comparison":"Ours (Algorithm 1 + Algorithm 2 selection) vs. plug-in KRR, DR-learner, oracle-propensity DR-learner and R-learner, on the same datasets and test points","falsification":"In the sample-size rows, the R-learner or DR-learner matches or beats Ours in mean MISE, or the gap between Ours and the best baseline falls within about one combined standard error (e.g. (10,10), n = 1000: R <= ~0.0113 with Ours ~0.0099).","caption":"Table 2: Multivariate synthetic results. Entries are mean MISE (SE), with the smallest mean in bold. Unless varied, sigma = 1 and kappa = 0.1; the sigma = 1 row also supplies the kappa = 0.1 reference. p = d is dense and p = 4 is sparse. Sample-size rows use fixed kernel lengths; other rows use median-calibrated lengths. The sigma = 1 row at (d, p) = (20, 4) also gives the default representation library.","image":"img/2602.18958.png","kind":"table","image_w":1144,"image_h":1194,"pwc":false}
 
1
+ {"id":"2602.18958","title":"Structure-Adaptive Estimation of Heterogeneous Treatment Effects with Kernel Methods","authors":"Seok-Jin Kim","openreview_id":"B6if6B55a9","presentation":"poster","track":"main","verdict":"borderline","background":"Estimating the conditional average treatment effect (CATE) from observational data is hard when the baseline outcome functions are complex and when overlap is weak, meaning the propensity score gets close to 0 or 1. Doubly robust and R-learner style methods depend on estimating the propensity score, and inverse-propensity weighting becomes unstable under weak overlap. The CATE itself is often much simpler (smoother, sparse or low-dimensional) than the nuisance outcome functions. The paper asks whether a kernel method can reach error rates set by the simpler CATE class without estimating the propensity score.","method":["Algorithm 1 ('Ours') fits each treatment arm's outcome function with undersmoothed kernel ridge regression (KRR) in a nuisance RKHS F. It then forms switch-imputed pseudo-outcomes: m_i = y_i - f0_hat(x_i) for treated units and m_i = f1_hat(x_i) - y_i for control units. Finally it runs a second KRR of m_i on x_i in a simpler target kernel space H. No propensity score is estimated. In the experiments, the raw nuisance ridge is fixed at 0.01. The target kernel and ridge are chosen by Algorithm 2: candidates are fit on one split, nuisances are refit on a second split, candidates are scored against switch-imputed proxies on a third split, and predictions are averaged over the three cyclic fits.","Multivariate synthetic setup (Appendix C.2): X ~ Unif([-1,1]^d), f0(x) = (2/d) sum_{j=1}^d sin(x_j), h(x) = (0.3/p) sum_{j=1}^p x_j^2, A ~ Bernoulli(clip(sin(5||x||_2), kappa, 1-kappa)), Y = f0(X) + A h(X) + N(0, sigma^2), with defaults sigma = 1 and kappa = 0.1. The nuisance kernel is Matern nu = 1.5. Ours selects among Matern 1.5, Matern 2.5 and RBF target kernels, each on all coordinates or on x_{1:4}. Baselines are plug-in KRR (T-learner, 3-fold CV), the DR-learner (Kennedy 2023), oracle-propensity DR and the R-learner (Nie and Wager 2021). DR, oracle DR and R use Matern 2.5 target kernels with cross-fitting and receive the true support in sparse settings. Estimated propensities are clipped to [0.01, 0.99]. The metric is MISE on 3,000 held-out uniform test points."],"claim":"In the four multivariate sample-size settings ((d,p) = (10,10) and (10,4), n = 1000 and 2000, sigma = 1, kappa = 0.1, 100 replications), the two-stage switch-imputation KRR estimator has the lowest CATE MISE every time. For example, at (10,10), n = 1000 it gets 0.0099 (SE 0.0008), against 0.0127 (0.0012) for the R-learner, 0.0184 (0.0049) for DR, 0.0201 (0.0031) for oracle DR and 0.0478 (0.0018) for plug-in KRR.","source":"Table 2, page 18 (Appendix C.2)","metric":"Mean integrated squared error of the CATE estimate over 3,000 test points (lower is better); the value in parentheses is the Monte Carlo standard error of the mean over replications (100 for the sample-size rows, 30 for the sweeps)","comparison":"Ours (Algorithm 1 + Algorithm 2 selection) vs. plug-in KRR, DR-learner, oracle-propensity DR-learner and R-learner, on the same datasets and test points","falsification":"In the sample-size rows, the R-learner or DR-learner matches or beats Ours in mean MISE, or the gap between Ours and the best baseline falls within about one combined standard error (e.g. (10,10), n = 1000: R <= ~0.0113 with Ours ~0.0099).","caption":"Table 2: Multivariate synthetic results. Entries are mean MISE (SE), with the smallest mean in bold. Unless varied, sigma = 1 and kappa = 0.1; the sigma = 1 row also supplies the kappa = 0.1 reference. p = d is dense and p = 4 is sparse. Sample-size rows use fixed kernel lengths; other rows use median-calibrated lengths. The sigma = 1 row at (d, p) = (20, 4) also gives the default representation library.","image":"img/2602.18958.png","kind":"table","image_w":1144,"image_h":1194,"pwc":true}
site/p/2602.19770.json CHANGED
@@ -1 +1 @@
1
- {"id":"2602.19770","title":"The Confusion is Real: GRAPHIC – A Network Science Approach to Confusion Matrices in Deep Learning","authors":"Johanna S. Fröhlich; Bastian Heinlein; Jan Claar; Hans Rosenberger; Vasileios Belagiannis; Ralf R Müller","openreview_id":"UP9bx1WJwR","presentation":"poster","track":"journal:TMLR","verdict":"borderline","background":"Explainability methods usually explain single predictions or rely on human-defined concepts. GRAPHIC instead works at the level of classes: it builds confusion matrices for intermediate layers of a network, treats each one as a directed graph over classes, and uses network-science tools (modularity-based community detection, assortativity) to follow how class confusions change across layers and training epochs. As a side effect of this probing setup, the paper also measures how linearly separable the classes are at each layer over training. There it reports that a vision transformer behaves differently from a CNN.","method":["During training, a linear classifier (softmax over W h + b) is fit at regular checkpoints on the features from each intermediate stage of the network under study. It uses the loss λ·CE(z, y) + (1−λ)·CE(z, ỹ), where y is the true label and ỹ is the network's own prediction. For the linear-separability analysis, λ = 1, so the probes see only true labels. The probes are trained on 80% of the CIFAR-100 training split. Accuracy is measured on the held-out 20% ('training set' curves) and on the CIFAR-100 validation split ('validation set' curves).","The network is the efficient vision transformer of Irandoust et al. (2022), called EffVit in the paper, trained from scratch on CIFAR-100 with batch size 64, learning rate 0.001 and AdamW for 1,000 epochs. A probe (learning rate 0.01) is trained every 10 epochs on the outputs of decoders 1, 3, 6, 9 and 12. For comparison, ResNet-50 is trained on CIFAR-100 (Adam, learning rate 0.003, batch size 1,024, 71 epochs) and probed every 5 epochs on the outputs of its four blocks. The paper reports that the ResNet-50 probe accuracies increase and then plateau, with no drop (Appendix Figure 11)."],"claim":"In EffVit trained on CIFAR-100, true-label linear-probe accuracy for early decoders peaks early and then declines: decoder 1 goes from about 0.22 (around epoch 200) to about 0.13 (epoch 1000), and decoder 3 from about 0.41 (epochs 200–350) to about 0.32. Over the same run, decoders 9 and 12 rise to about 0.79 and 0.85 validation accuracy and stay there.","source":"Figure 7, page 12","metric":"Accuracy of linear probes trained on true labels (λ = 1), on the held-out 20% of the training split and on the validation split; higher means more linearly separable. Single run with no error bars. The values here are read off the plot and are approximate.","comparison":"Probe accuracy at early decoders (1, 3) near their peak (~epoch 200) vs. at the end of training (epoch 1000); contrasted with the later decoders (9, 12) and with ResNet-50, whose probe accuracies increase monotonically (Appendix Figure 11).","falsification":"In a reproduction, linear-probe accuracy on decoders 1 and 3 does not fall substantially from its peak by the end of training (for example, decoder 1 stays above about 0.20 and decoder 3 stays above about 0.38 at epoch 1000), or the drop goes away when the probe hyperparameters or seed change.","caption":"Figure 7: Linear separability trends in EffVit. Accuracy of CMs generated by LCs trained on true labels for decoders 1, 3, 6, 9, and 12 of EffVit, shown over the training epochs.","image":"img/2602.19770.png","kind":"figure","image_w":1311,"image_h":550,"pwc":false}
 
1
+ {"id":"2602.19770","title":"The Confusion is Real: GRAPHIC – A Network Science Approach to Confusion Matrices in Deep Learning","authors":"Johanna S. Fröhlich; Bastian Heinlein; Jan Claar; Hans Rosenberger; Vasileios Belagiannis; Ralf R Müller","openreview_id":"UP9bx1WJwR","presentation":"poster","track":"journal:TMLR","verdict":"borderline","background":"Explainability methods usually explain single predictions or rely on human-defined concepts. GRAPHIC instead works at the level of classes: it builds confusion matrices for intermediate layers of a network, treats each one as a directed graph over classes, and uses network-science tools (modularity-based community detection, assortativity) to follow how class confusions change across layers and training epochs. As a side effect of this probing setup, the paper also measures how linearly separable the classes are at each layer over training. There it reports that a vision transformer behaves differently from a CNN.","method":["During training, a linear classifier (softmax over W h + b) is fit at regular checkpoints on the features from each intermediate stage of the network under study. It uses the loss λ·CE(z, y) + (1−λ)·CE(z, ỹ), where y is the true label and ỹ is the network's own prediction. For the linear-separability analysis, λ = 1, so the probes see only true labels. The probes are trained on 80% of the CIFAR-100 training split. Accuracy is measured on the held-out 20% ('training set' curves) and on the CIFAR-100 validation split ('validation set' curves).","The network is the efficient vision transformer of Irandoust et al. (2022), called EffVit in the paper, trained from scratch on CIFAR-100 with batch size 64, learning rate 0.001 and AdamW for 1,000 epochs. A probe (learning rate 0.01) is trained every 10 epochs on the outputs of decoders 1, 3, 6, 9 and 12. For comparison, ResNet-50 is trained on CIFAR-100 (Adam, learning rate 0.003, batch size 1,024, 71 epochs) and probed every 5 epochs on the outputs of its four blocks. The paper reports that the ResNet-50 probe accuracies increase and then plateau, with no drop (Appendix Figure 11)."],"claim":"In EffVit trained on CIFAR-100, true-label linear-probe accuracy for early decoders peaks early and then declines: decoder 1 goes from about 0.22 (around epoch 200) to about 0.13 (epoch 1000), and decoder 3 from about 0.41 (epochs 200–350) to about 0.32. Over the same run, decoders 9 and 12 rise to about 0.79 and 0.85 validation accuracy and stay there.","source":"Figure 7, page 12","metric":"Accuracy of linear probes trained on true labels (λ = 1), on the held-out 20% of the training split and on the validation split; higher means more linearly separable. Single run with no error bars. The values here are read off the plot and are approximate.","comparison":"Probe accuracy at early decoders (1, 3) near their peak (~epoch 200) vs. at the end of training (epoch 1000); contrasted with the later decoders (9, 12) and with ResNet-50, whose probe accuracies increase monotonically (Appendix Figure 11).","falsification":"In a reproduction, linear-probe accuracy on decoders 1 and 3 does not fall substantially from its peak by the end of training (for example, decoder 1 stays above about 0.20 and decoder 3 stays above about 0.38 at epoch 1000), or the drop goes away when the probe hyperparameters or seed change.","caption":"Figure 7: Linear separability trends in EffVit. Accuracy of CMs generated by LCs trained on true labels for decoders 1, 3, 6, 9, and 12 of EffVit, shown over the training epochs.","image":"img/2602.19770.png","kind":"figure","image_w":1311,"image_h":550,"pwc":true}