{"id":"6f54f72a-658f-4c92-822c-201cb5b5ec9c","arxiv_id":"2412.13573","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A self-feedback training framework that refines loss landscapes with dynamically generated soft labels finds more consistent flat minima and improves domain generalization accuracy across five benchmarks.","lead":"This paper proposes a training method called Self-Feedback Training (SFT) that makes a model's loss surface more similar across different training domains, so the model lands in flatter and more transferable minima. In tests on five standard domain generalization benchmarks, SFT improves average accuracy by about 2.6% over sharpness-aware minimization with ResNet-50 and 1.5% with ViT-B/16.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 2's transfer step treats fixed held-out domains as i.i.d. samples from W; under the DomainBed protocol this does not hold, so the claimed transfer of landscape consistency to unseen domains is unsupported.","rationale":"The reader identified the i.i.d. domain-shift assumption as the weakest point. I agree: this is the load-bearing step in the theoretical transfer argument. The empirical comparison with SAM may still stand, because the reported gains are measured directly on fixed benchmark test domains, but the paper's causal claim that consistent flat minima 'transfer' to unseen domains is only justified in a protocol where the held-out domains are random draws from W. The concrete test above would settle whether the proof can be repaired by adding an explicit domain-mismatch term. Since the reader already assigned CONDITIONAL largely because of this and related caveats, my analysis does not change the verdict.","tokens_in":29822,"tokens_out":16984,"duration_ms":176228,"concrete_test":"Independently re-derive Eq. (22) under the actual benchmark sampling scheme, where the p−1 held-out domains Dd′ and the test domain are fixed rather than sampled from W. Identify the additive divergence term Δ(W, {Dd′}, test) that must appear on the right-hand side for the bound to hold. Then check whether this Δ can be bounded on the five DomainBed benchmarks for any reasonable choice of W; if no such bound exists, Theorem 2 does not support transfer to fixed unseen domains.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central transfer claim rests on Theorem 2 in Supplementary Section 6.2. The proof's key step is Eq. (22), which applies McAllester's bound (Lemma 1) to the p−1 held-out domains Dd′ as though they were i.i.d. draws from the domain-shift distribution W. For this step to be valid, two conditions are needed: (i) Dd′ ∼ W, and (ii) the test distribution ω(Dd) has the same law as the Dd′. The DomainBed protocol satisfies neither: the Dd′ are the other fixed training domains, and the true test domain is a fixed disjoint domain. Consequently, the bound on ℓ(Q,ω) in Eq. (17) does not follow from the proof. What is missing is an explicit term measuring the mismatch between the empirical domain distribution and W, or between W and the actual test domain. The theorem's conclusion that 'the consistency of flat minima achieved in the training domains can be transferred to unseen test domains' is therefore not established by the supplied argument. This is exactly the assumption stated in Supplementary Section 6.1 ('We assume that the domain shifts ω follow a specific distribution W'), and the proof does not quantify the cost when this assumption is violated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Self-Feedback Training (SFT), an iterative two-phase method for domain generalization. In the feedback phase, SFT trains the model on one sampled domain with sharpness-aware minimization under soft labels, then measures landscape inconsistency as the absolute difference of sharpness between two training domains. In the refinement phase, it updates a landscape refiner g_phi using a loss that combines a projection cross-entropy term, a sharpness term, and the sharpness-difference consistency term. The authors claim that this procedure seeks flat minima that are consistent across domains, enabling better transfer to unseen test domains. They support this with a PAC-Bayesian theorem in the supplementary material, a toy experiment visualizing loss surfaces, and extensive DomainBed experiments with ResNet-50, ViT-B/16, and ViT-L/14, reporting average gains of +2.6% over SAM with ResNet-50 and +1.5% with ViT-B/16.","tokens_in":30059,"tokens_out":2925,"duration_ms":30197,"significance":"If the central claim holds, the paper makes a useful contribution: it reframes domain generalization as a problem of refining loss landscapes for consistency, introduces a differentiable soft-label refiner, and provides a complete, efficient algorithm for the projection cross-entropy optimization. The empirical evaluation is broad, covering five DomainBed benchmarks and three backbones, and the reported gains over SAM and its variants are consistent. The paper also ships a detailed reproducibility section and a self-contained derivation of Algorithm 2. However, the theoretical transfer guarantee, which is the main justification for the core claim that landscape consistency transfers to unseen domains, rests on an i.i.d. assumption about domain shifts that does not hold under the DomainBed protocol. The manuscript is therefore technically sound in its empirical parts but has a load-bearing gap in its theoretical argument.","major_comments":[{"comment":"The transfer step of Theorem 2 is not established under the DomainBed protocol. Eq. (22) applies McAllester's bound with the p-1 held-out training domains D_d' treated as i.i.d. draws from the domain-shift distribution W, and the test domain omega(D_d) as having the same law. In the actual experiments, the test domains are fixed, disjoint datasets, and the held-out domains are the other training domains, not random samples from a distribution over shifts. The proof does not include a term measuring the mismatch between the empirical domain distribution and W, or between W and the true test distribution. Consequently, the conclusion that 'the consistency of flat minima achieved in the training domains can be transferred to unseen test domains' does not follow from the supplied argument. The paper should either prove a bound with an explicit mismatch or distribution-shift term, or substantially weaken the claim to a conditional statement that is clearly satisfied by the experimental protocol.","section":"Supplementary Section 6.2, Eq. (22)"},{"comment":"The connection between the optimized objective and the theorem's premise is loose. Theorem 2 assumes that the expected sharpness difference is at most epsilon for all hyper-posterior distributions Q that satisfy Eq. (16). However, the training loss in Eq. (13) penalizes the sharpness difference for a single randomly selected domain pair with a finite weight lambda2, and also includes a separate sharpness term lambda1 that encourages flatness but does not directly enforce consistency. There is no argument that minimizing Eq. (13) drives the expectation in Eq. (16) below epsilon, nor is there a statement about how the finite lambda2 and the optimization trajectory relate to the theorem's uniform condition. The paper should provide an explicit optimization-to-theorem link, or state the condition as an assumption whose verification is left to the experiments.","section":"Section 3.2.2, Eq. (13) and Theorem 2 condition Eq. (16)"},{"comment":"McAllester's bound in Lemma 1 requires a bounded loss taking values in [0,1], but the sharpness quantity used throughout the paper, defined in Eq. (9) as a difference of cross-entropy losses, is not obviously bounded and can exceed one or become negative for soft labels. If the theorem is meant to bound the sharpness of the test loss, the paper must specify the normalization or clipping used to make the sharpness function satisfy the lemma's conditions, or prove that the sharpness difference lies in [0,1] under the stated assumptions. Without this, the PAC-Bayesian estimate terms in Eq. (17) do not formally apply.","section":"Supplementary Section 6.2, Lemma 1 and Eq. (9)"}],"minor_comments":[{"comment":"The caption of Table 10 says 'TerraIncognita' but the table reports DomainNet results; it should read 'DomainNet'.","section":"Table 10 caption"},{"comment":"The axis labels in Figure 3 are duplicated, and the hyperparameter in the legend appears as 'λ' rather than 'lambda2'; this makes the figure harder to read.","section":"Figure 3"},{"comment":"The expression 'p α 1' in the condition on line 5 is ambiguous in the main text; it should be written as p_1^alpha (or the product) with clear parentheses to avoid confusion with the subscript notation used elsewhere.","section":"Algorithm 2, line 5"},{"comment":"The description of the landscape refiner says it 'shares the same architecture with the model f_theta' but does not specify how its output is normalized or whether it has a separate output head for the N-class soft labels; a short clarification would improve reproducibility.","section":"Section 3.2.1"},{"comment":"In the KKT derivation, the notation n and N are both used for the number of classes (e.g., 'j = 2, ..., n' and '(q1, ..., qN)'); the paper should use one symbol consistently.","section":"Supplementary Section 7.1.1"}],"recommendation":"major_revision","confidential_remarks":"The empirical contribution is substantial and the paper is well-referenced in the sharpness-aware minimization literature. The main risk is that the theoretical section, as written, overclaims what is proven: the PAC-Bayesian transfer bound assumes a distribution over domain shifts that the DomainBed benchmark does not provide. This is fixable by rewriting the theorem as a conditional statement or adding a mismatch term, but it is central to the paper's stated justification. I would not reject on this basis, as the method's empirical results stand on their own; I would require the authors to either repair the theorem or explicitly delimit its scope before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is a useful, mostly empirical DG paper with a genuinely new mechanism, and the experiments are broad enough that the central claim — consistent 1-3% gains over SAM on five benchmarks — is likely correct. The theory in the supplement is the weak part: it claims a transfer bound that doesn't actually cover the fixed test domains used in DomainBed. But the method and the empirical work deserve a real referee.\n\nWhat's new: instead of just seeking flat minima in a static landscape, the paper trains a small refiner network to produce soft labels, and uses the difference in sharpness between two training domains as feedback to update that refiner. The refiner then reshapes the loss landscapes so the flat regions align across domains. That label-based landscape refinement is not in the SAM/SWAD/GAM/FAD/SAGM line, and I don't know of prior work doing exactly this. The projection cross-entropy loss with the exact KL-minimization algorithm is also a neat, self-contained piece of work — it's formal, and the algorithm's derivation is real.\n\nWhat it does well: five DomainBed benchmarks, three backbones (ResNet-50, ViT-B/16, ViT-L/14), consistent improvements over strong baselines. The ablation study on OfficeHome is informative — it shows the consistency term, the sharpness term, and the PCE loss all contribute. The toy experiment and the sharpness analysis back up the mechanism.\n\nWhere it's soft: the transfer theorem (Theorem 2 in the supplement) is the biggest gap. The proof applies McAllester's bound to the held-out training domains as if they were i.i.d. draws from the domain-shift distribution W, but under the DomainBed protocol the held-out domains are fixed and the test domain is a fixed disjoint domain. So the bound on the expected sharpness at the test distribution does not follow from the supplied argument. The main text's claim that consistency transfers to unseen domains is therefore not backed by the theory. This isn't fatal to the empirical contribution, but the authors should either fix the theorem or stop citing it as the reason the method works. Also, no code is released, and the method has four extra hyperparameters (ρ, λ1, λ2, α) tuned per dataset, which makes head-to-head comparisons a bit less clean.\n\nWho it's for: anyone working on flat-minima methods for DG or on label-based landscape shaping. It's worth a serious referee; I'd send it out with the expectation of major revision on the theory and a request for code.\n\nRecommendation: engage with it — but push on the theory and reproducibility.","headline":"Solid empirical DG paper with a genuinely new label-refining mechanism; the theory overclaims transfer, but the experiments stand on their own.","tokens_in":30572,"tokens_out":2483,"would_cite":true,"duration_ms":21983,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Consistent flat minima, enforced by refining loss landscapes with soft labels, improve domain generalization by up to 2.6 percentage points over sharpness-aware minimization.","keywords":["domain generalization","flat minima","loss landscape","sharpness-aware minimization","soft labels","self-feedback training","PAC-Bayesian analysis","DomainBed"],"falsifier":"Take the best SFT configuration and rerun training with the sharpness-difference term in Eq. (13) set to zero while keeping the soft-label refiner and the SAM step; if out-of-domain accuracy on a test domain whose shift was not present among training domains does not drop, then the consistency term is not the carrier of the reported gains.","tokens_in":29621,"feed_emoji":"🔁","tokens_out":6295,"duration_ms":57307,"temperature":0.7,"pith_summary":"This paper argues that for domain generalization, finding a flat minimum in the training loss is not enough: the minimum must be flat in the same way across all training domains, so that it also sits in a flat region of unseen test domains. It proposes Self-Feedback Training (SFT), an iterative loop that measures how much the loss-sharpness values of two training domains disagree, and then trains a small landscape refiner network to generate soft labels that reduce that disagreement while keeping labels correct. Over five DomainBed benchmarks, SFT raises average out-of-domain accuracy to 67.1% with ResNet-50, 2.6 points above sharpness-aware minimization (SAM), and to 77.8% with ViT-B/16, 1.5 points above SAM. The intended upshot is that landscape consistency, not flatness alone, is what transfers to unseen domains.","feed_headline":"Flat-minima consistency lifts domain generalization by 2.6 points","feed_subtitle":"A feedback loop aligns flat minima across domains by refining loss landscapes, beating SAM on five benchmarks.","key_machinery":"The machinery is the self-feedback loop between a model $f_\\theta$ and a landscape refiner $g_\\phi$. In the feedback phase, two training domains $D_d$ and $D_{d'}$ are picked; the model is updated on $D_d$ with a SAM-style loss using soft labels $\\tilde{y} = g_\\phi(x)$, and the difference in loss sharpness $|L^{\\mathrm{CE-SL}}_{D_d}(\\theta,\\phi) - L^{\\mathrm{CE-SL}}_{D_{d'}}(\\theta,\\phi)|$ is computed as the inconsistency signal. In the refinement phase, $g_\\phi$ is updated to minimize projection cross-entropy (PCE), which projects soft labels onto the label space $\\{q_1 \\geq \\alpha q_k,\\ \\sum_k q_k = 1\\}$ via a fast KL-minimization algorithm, plus penalties on sharpness and on the sharpness difference. This progressively reshapes the loss landscapes so that their flat minima coincide; a PAC-Bayesian bound (Theorem 2 in the supplementary) is used to argue that test-domain sharpness is controlled by training-domain sharpness when domain shifts are drawn independently from a fixed distribution.","core_discovery":"Flat minima help domain generalization only when they are consistent across domains. SFT's central discovery is that loss landscapes can be actively reshaped during training: a refiner network produces dynamic soft labels, and a feedback signal defined as the absolute difference between SAM sharpness values on two training domains tells the refiner how to change those labels so the landscapes in different domains become both flatter and more alike. The refined landscapes then share a flat basin, and the model trained inside that basin generalizes better out of domain. The paper reports consistent gains over SAM and over other sharpness-aware methods, including GAM, GSAM, FAD, and SAGM, on five DomainBed benchmarks, with the largest improvements on TerraIncognita and PACS.","pith_inferences":["A direct test of the mechanism would use the same soft-label refiner but with a random or sign-flipped feedback signal: if the gain persists, the consistency term itself is not what helps.","The refiner is a second network with the same architecture as the model; a smaller or shared refiner could reduce compute while retaining most of the consistency benefit.","The feedback signal uses zero-order sharpness; plugging first-order flatness measures, as in GAM or FAD, into the same loop might yield sharper reductions in landscape discrepancy.","The consistency-refinement view suggests a new interpretation of soft labels as data-dependent, geometry-aware label smoothing, which connects SFT to knowledge distillation and to label-noise-robust training."],"forward_implications":["SFT can be layered on top of sharpness-aware training and improves out-of-domain accuracy by 2.6 points with ResNet-50 and 1.5 points with ViT-B/16 averaged over five DomainBed benchmarks.","Because the consistency penalty is separate from the choice of sharpness optimizer, replacing the basic SAM step with stronger variants such as SAGM should give further gains; the paper names this as future work.","The projection cross-entropy algorithm converges in at most $N$ steps, where $N$ is the number of classes, and runs about 10 times faster than a generic convex solver, making landscape refinement cheap enough for large-scale models.","The theory predicts that reducing sharpness discrepancy among training domains bounds the sharpness on unseen test domains, which is the formal sense in which landscape consistency transfers."],"supporting_citations":[{"why":"Supplies the sharpness-aware minimization loss and sharpness definition that SFT builds on, and serves as the main baseline.","marker":"[16]"},{"why":"Provides the DomainBed benchmark and evaluation protocol used for all real-world domain generalization experiments.","marker":"[19]"},{"why":"Establishes the flat-minima approach to domain generalization that motivates seeking consistent flat minima.","marker":"[9]"},{"why":"Supplies the PAC-Bayesian bound used as Lemma 1 in the proof of Theorem 2 on transfer of sharpness consistency.","marker":"[74]"},{"why":"Provides the label smoothing baseline and comparison for the PCE soft-label objective.","marker":"[46]"},{"why":"Gives the loss landscape visualization methodology used to depict landscape consistency in the toy experiments.","marker":"[39]"},{"why":"Serves as a strong pre-trained-model domain generalization baseline that SFT outperforms.","marker":"[10]"}],"fun_headline_variants":["Feedback loop aligns flat minima to boost domain generalization","Refining loss landscapes for consistent flat minima in DG","SAM-beating method: consistent flat minima via landscape refinement","Steady flat minima across domains improve generalization by 2.6%","A feedback signal reshapes loss landscapes for better out-of-domain accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The theoretical transfer guarantee assumes that the shifts between training domains are independent draws from the same distribution $W$ that also produces the unseen test domains; on the real benchmarks the test domains are fixed, so the bound does not directly cover the evaluated setting.","fun_headline_variants_meta":{"raw":{"variants":["Feedback loop aligns flat minima to boost domain generalization","Refining loss landscapes for consistent flat minima in DG","SAM-beating method: consistent flat minima via landscape refinement","Steady flat minima across domains improve generalization by 2.6%","A feedback signal reshapes loss landscapes for better out-of-domain accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000735,"raw_usage":{"total_tokens":3277,"prompt_tokens":925,"completion_tokens":2352,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":541,"completion_tokens_details":{"reasoning_tokens":2269}},"tokens_in":541,"tokens_out":2352,"duration_ms":14718,"temperature":1.0,"reasoning_tokens":2269,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T13:00:10.224672+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the best SFT configuration and rerun training with the sharpness-difference term in Eq. (13) set to zero while keeping the soft-label refiner and the SAM step; if out-of-domain accuracy on a test domain whose shift was not present among training domains does not drop, then the consistency term is not the carrier of the reported gains.","supporting_citations":[{"cited_title":"Swad: Domain generalization by seeking flat minima","cited_arxiv_id":null,"evidence_quote":"Establishes the flat-minima approach to domain generalization that motivates seeking consistent flat minima."},{"cited_title":"Pac-bayesian model averaging","cited_arxiv_id":null,"evidence_quote":"Supplies the PAC-Bayesian bound used as Lemma 1 in the proof of Theorem 2 on transfer of sharpness consistency."},{"cited_title":"When does label smoothing help? Advances in neural infor- mation processing systems, 32, 2019","cited_arxiv_id":null,"evidence_quote":"Provides the label smoothing baseline and comparison for the PCE soft-label objective."},{"cited_title":"Visualizing the loss landscape of neural nets.Ad- vances in neural information processing systems , 31, 2018","cited_arxiv_id":null,"evidence_quote":"Gives the loss landscape visualization methodology used to depict landscape consistency in the toy experiments."},{"cited_title":"Domain Generalization by Mutual-Information Regularization with Pre-trained Models","cited_arxiv_id":"2203.10789","evidence_quote":"Serves as a strong pre-trained-model domain generalization baseline that SFT outperforms."}],"review_version":1}