- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
This paper proposes OR-Diffusion, a systematic optimization modeling method: it first uses a diffusion model to generate diverse data, then passes the data through a fine-grained evaluator, and finally adopts more targeted training methods for data with different feedback. Strengths: Innovative use of diffusion: the use of diffusion is well integrated with the motivation of insufficient data diversity, and it makes good use of the characteristics of diffusion models and the properties of the task itself. Isolating executability from other properties makes sense: the authors argue that during DPO, learning from comparing non-executable programs with correct answers is inefficient. Weaknesses: The baseline setup has problems: one of the main contributions of this paper is to address the lack of diversity in existing data generation, such as ReSocratic. However, only in the ablation experiments is there a comparison between OR-Diffusion and ReSocratic on some literal statistics, which is not convincing. The experiments in Table 1 should include the performance of a model trained with the ReSocratic method; the ablation experiments in Table 3 are too thin and insufficient. Strong teacher distillation confound. Reverse and Rewrite both use DeepSeek-V4-Flash, used by latter SFT/DPO. The improvement of the 8B model may partly come from distilling a stronger model, rather than entirely from diffusion/HVE. The paper does not report the zero-shot performance of DeepSeek-V4-Flash on these benchmarks, nor does it provide a “pure distillation with the same data” control. Lack of connection between contributions: the two core contributions—data augmentation and training-objective optimization—lack an intrinsic connection and look more like a concatenation to strengthen the contribution, without leaving a deeper or more insightful contribution in either aspect. For example, the discussion of the lexicographic protocol or HVE is actually quite shallow. Although the authors claim it is a four-layer hierarchical reward, V1 and V2 are not important enough (Table 3 also shows that V3 is the most critical layer). So why can’t the reward be directly simplified into a binary classification (V1–V3 as one class and V4 as the other)?
Critical strengths
Innovative use of diffusion: the use of diffusion is well integrated with the motivation of insufficient data diversity, and it makes good use of the characteristics of diffusion models and the properties of the task itself. Isolating executability from other properties makes sense: the authors argue that during DPO, learning from comparing non-executable programs with correct answers is inefficient.
Critical weaknesses
-
A central contribution is addressing the limited diversity of existing data-generation methods such as ReSocratic. However, the comparison with ReSocratic is restricted to an ablation and some surface-level textual statistics, which is not convincing. Table 1 should include a model trained with the ReSocratic method, and the ablations in Table 3 are too limited to substantiate the claims.
-
Both Reverse and Rewrite use DeepSeek-V4-Flash, and the resulting data later supervise SFT/DPO. Therefore, part of the improvement in the 8B model may come from distilling a stronger model rather than solely from diffusion/HVE. The paper does not report the zero-shot performance of DeepSeek-V4-Flash on these benchmarks, nor does it provide a distillation-only control under matched data and training budgets.
-
Data augmentation and training-objective optimization lack an intrinsic connection and read more like a concatenation intended to strengthen the contribution, without yielding deeper insight into either aspect. For example, the discussion of the lexicographic protocol/HVE is relatively shallow. Although the authors describe a four-layer hierarchical reward, the distinct value of V1 and V2 remains unclear; Table 3 suggests that V3 is the most critical layer. Why can the reward not be simplified into a binary classification, with V1–V3 as one class and V4 as the other? This concern bears directly on whether the hierarchical design, rather than the generated corpus or teacher distillation, drives the reported gains.
Questions for authors
- The paper positions diffusion-based structural synthesis against text-level synthesis such as ReSocratic, but Table 1 does not include a ReSocratic-trained model. Can the authors report the downstream Opt@1 of a ReSocratic-style baseline trained with the same backbone, corpus size, and SFT/DPO protocol? If that is not feasible, can they explain why the current textual-synthesis control (A2) is a sufficient proxy for evaluating the claimed advantage of diffusion?
- Reverse and Rewrite both use DeepSeek-V4-Flash, and the resulting data later supervise the 8B model. Can the authors report the zero-shot performance of DeepSeek-V4-Flash on the six benchmarks, and provide a distillation-only control that uses the same teacher-generated data but without diffusion-based synthesis or HVE-based routing/gating? This would help determine how much of the gain is attributable to OR-Diffusion rather than to strong-teacher supervision.
- Table 3 suggests that V3 is the most critical layer, while the independent contributions of V1 and V2 remain unclear. Can the authors provide an ablation that collapses V1–V3 into a single binary admission condition—i.e., pass V1–V3 versus fail V1–V3, followed by V4 optimization? Would performance remain comparable? This directly tests whether the lexicographic hierarchy is necessary or whether a simpler binary gate suffices.
- The paper claims the two components are strongly complementary, but the mechanism is not fully isolated. Can the authors provide more direct evidence that diffusion-generated instances, rather than seed instances or teacher-generated data, specifically elicit the feasible-but-suboptimal rollouts that gated DPO refines? For example, what fraction of DPO pairs are formed on diffusion versus seed corpora, and how does that mediate the reported gain?
7 comments
Response to Reviewer xkys (Part2, W2 & W3 & Q2)
2. Disentangling Teacher Distillation from OR-Diffusion—and Synergy Between Synthesis and HVE (Weakness 2, Weakness 3 & Question 2)
To cleanly separate strong-teacher supervision (DeepSeek-V4-Flash) from our structural synthesis and hierarchical training, we evaluated (i) the standalone zero-shot accuracy of DeepSeek-V4-Flash across all six benchmarks and (ii) a Distillation-Only Control that applies DeepSeek-V4-Flash (Rewrite) to the seed problems () and trains Qwen3-8B with standard unweighted cross-entropy SFT (no diffusion synthesis and no HVE routing/gating):
| Model / Control Setting | Diffusion | HVE Routing/Gating | NL4Opt | EasyLP | ComplexLP | NLP4LP | CompOR | IndOR | Avg. Opt@1 |
|---|---|---|---|---|---|---|---|---|---|
Base Qwen3-8B (Zero-shot, Table 1) | × | × | 76.2 | 95.6 | 14.4 | 88.2 | 22.2 | 50.0 | 57.8 |
DeepSeek-V3 (671B, Zero-shot, Table 1) | — | — | 79.9 | 96.7 | 55.0 | 94.9 | 44.4 | 69.0 | 73.3 |
DeepSeek-V4-Flash (Synthesis Teacher, Zero-shot, revised Table 1) | — | — | 75.7 | 94.7 | 57.7 | 94.4 | 50.0 | 69.0 | 73.6 |
| Distillation-Only Control (Seed + Teacher , Plain SFT, revised Table 1) | × | × | 87.4 | 92.3 | 62.2 | 88.8 | 38.9 | 45.2 | 69.1 |
Diffusion + Teacher , Plain SFT (B2 in Table 3) | ✓ | × | 84.6 | 89.0 | 36.9 | 89.9 | 27.8 | 45.2 | 62.2 |
Seed + Weakest-Layer SFT + Gated DPO (A1 in Table 3) | × | ✓ | 88.3 | 95.4 | 63.1 | 93.3 | 50.0 | 54.8 | 74.1 |
Full OR-Diffusion (Qwen3-8B, Table 1) | ✓ | ✓ | 96.3 | 97.4 | 65.8 | 97.2 | 55.6 | 57.1 | 78.2 |
Three key findings:
- Quantitative decomposition of the +20.4-point total gain (57.8% 78.2%): Standard supervised distillation from
DeepSeek-V4-Flashon seed problems liftsQwen3-8Bfrom 57.8% to 69.1% (Distillation-Only, +11.3 points), which remains 4.5 points below the teacher itself (73.6%). Under the exact same teacher, corpus size (), backbone, and training budget, coupling discrete graph diffusion with HVE routing and gated DPO contributes an additional +9.1 percentage points (69.1% 78.2%). BothDeepSeek-V4-Flashand theDistillation-Only Controlare now reported in Table 1 of the revised manuscript. - Why our 8B student surpasses its own synthesis teacher (
DeepSeek-V4-Flash, 78.2% vs. 73.6%, +4.6 points) as well as 671BDeepSeek-V3(73.3%): Two ingredients work together. First, witness-point construction supplies an exact solver-certified ground-truth optimum for every synthesized graph, allowing solver-checkedRewrite(with up to two retries; Appendix G.4) to filter out teacher errors exceeding (retaining 1,910 of 3,000 candidates; submission Table 7 / revised Table 9) so is clean. Second—asDistillation-Only(69.1% 73.6%) confirms—clean teacher labels alone do not surpass the teacher: exceeding the teacher requires expanding structural coverage via graph diffusion (42.9% unseen signatures) and resolving failure modes via HVE routing and gated DPO (+9.1 points). - Strong synergy rather than simple concatenation—why diffusion requires HVE (
62.2%without HVE vs.78.2%with HVE): Comparing both with and without HVE reveals a striking interaction:- Without HVE (Plain SFT): Training on novel diffused topologies (
B2, 62.2%, dropping to 36.9% onComplexLP) actually underperforms plain SFT on familiar seeds (Distillation-Only, 69.1%) by 6.9 points, because 42.9% unfamiliar multi-constraint topologies initially trigger frequent syntax and structural mistakes (87.6% of non-optimal base rollouts) that swamp unweighted updates. - With HVE (Matched in submission Table 5 / revised Table 7): Once balanced pool routing stabilizes Stage 1 (
B1, 72.5% vs. routedSeed SFT71.7%) and gated DPO refines numerical binding in Stage 2, the diffusion corpus yields a +5.7-point DPO gain versus +2.4 points on seeds (A1, 74.1%), flipping diffusion to +4.1 points ahead overall (78.2%, and +28.7 points on held-out unseen constraint signatures, 79.0% vs. 50.3% in revised Table 16). Structural synthesis and hierarchical training are therefore mutually reinforcing rather than a loose concatenation.
- Without HVE (Plain SFT): Training on novel diffused topologies (
Response to Reviewer xkys (Part3, W3 & Q3)
3. Deeper Formulation of HVE and Necessity of All Four Layers vs. Binary Collapse of – (W3 & Q3)
Clarifying HVE's role (routing & gating vs. scalar reward) and why B3 had the largest drop in original Table 3: First, we have substantially expanded Section 3.3 (Page 6) in the revised PDF to clarify the theoretical role of the four-layer hierarchy. Importantly, HVE is not trained as a four-layer weighted scalar reward (indeed, collapsing layers into an un-gated scalar reward in B4 degrades Opt@1 to 69.8%, below SFT-only at 72.5%). As formalized in Section 3.3, while the four-layer lexicographic order carries the exact same diagnostic information as a non-overlapping milestone rubric, its empirical power comes from driving two distinct stage-specific training rules: inverse-size pool routing in SFT and hard candidate exclusion in DPO. Moreover, the large drop when removing pool (B3, 66.7%, -11.5 points) in original Table 3 did not mean and were unimportant—rather, errors are rare (4.8%) compared to (28.4%) and (59.2%), so without explicit four-layer separation is immediately drowned out.
To test your exact hypothesis—collapsing – into a single binary condition (Pass V1–V3 vs. Fail V1–V3) followed by optimization, alongside individual ablations of and —we ran four new controlled ablations under the exact same backbone (Qwen3-8B), corpus size ( = 3,461), and budget (now added as rows B5–B8 in revised Table 3):
| Hierarchy / Ablation Variant ( = 3,461, Qwen3-8B, revised Table 3) | NL4Opt | EasyLP | ComplexLP | NLP4LP | CompOR | IndOR | Avg. Opt@1 |
|---|---|---|---|---|---|---|---|
| Full 4-Layer HVE (– routing + gated DPO) | 96.3 | 97.4 | 65.8 | 97.2 | 55.6 | 57.1 | 78.2 |
B5 Binary Collapse in SFT (merge – into one "Fail –" pool + Gated DPO) | 89.7 | 94.5 | 53.2 | 90.4 | 38.9 | 50.0 | 69.5 |
B8 Solver-Only Binary Gate in DPO (drop ; gate only on ) | 92.5 | 96.5 | 61.3 | 92.7 | 50.0 | 50.0 | 73.8 |
B6 w/o (executability) pool in SFT (+ Gated DPO) | 91.6 | 95.8 | 59.5 | 91.6 | 50.0 | 50.0 | 73.1 |
B7 w/o (structural fidelity) pool in SFT (+ Gated DPO) | 89.7 | 94.1 | 52.3 | 89.9 | 38.9 | 45.2 | 68.4 |
B3 w/o (feasibility) pool in SFT (+ Gated DPO) | 88.8 | 93.6 | 50.5 | 88.8 | 33.3 | 45.2 | 66.7 |
Why both stages require explicit separation of , , and :
- In Stage 1 (SFT Pool Routing), collapsing
Fail(V1–V3)into one binary pool drops Opt@1 by -8.7 points (B5, 69.5%), and both and contribute substantially (B6,B7): Across non-optimal base-policy rollouts on , 28.4% fail at (, execution crashes), 59.2% fail at (, wrong constraint structure), only 4.8% pass – yet fail solver feasibility (, conflicting parameter bindings), and 7.6% pass – with misbound objectives (). Merging – into a single binary "Fail –" SFT pool (B5) causes syntax and structural mistakes to occupy 94.8% of failed updates (), starving and dropping macro Opt@1 from 78.2% to 69.5% (-8.7 points). Moreover, individually dropping the pool (B6, 73.1%, -5.1 points) or the pool (B7, 68.4%, -9.8 points—nearly matchingB3's -11.5-point drop) confirms the strong independent contribution of both and . - In Stage 2 (DPO Admission Gate), our gate is a binary
Pass(V1–V3)vs.Fail(V1–V3)filter—but computingPass(V1–V3)still requires HVE's explicit structural check (-4.4 points without ,B8): We agree that DPO itself uses – conjunctively as a binary admission condition () before optimizing . Crucially, however, an off-the-shelf solver can only verify execution () and feasibility (); no solver can check whether the candidate encoded the right constraint structure (). On (Table 2), post-SFT feasibility (91.7%) far exceeds structural fidelity (60.0%) because omitting a constraint or flipping an inequality often leaves a feasible LP. Simplifying the DPO gate to a solver-only binary check (, dropping ;B8) lets structurally wrong programs leak into , dropping macro Opt@1 from 78.2% to 73.8% (-4.4 points). Thus, all four explicit verification layers are indispensable across the two stages.
Response to Reviewer xkys (Part1, W1 & Q1)
We sincerely thank Reviewer xkys for recognizing our core motivation—using discrete graph diffusion to overcome the structural diversity bottleneck in training data and isolating executability and feasibility from objective-matching accuracy during preference learning. Following your suggestions, we ran all four requested sets of controlled experiments under the exact same backbone (
Qwen3-8B), corpus size (N=3,461), and training budget, and have incorporated all new baselines, ablations, and mechanism diagnostics into the revised manuscript (highlighted in blue in the uploaded PDF). Below, we address each weakness and question in detail.1. Direct Downstream Comparison with ReSocratic (Weakness 1 & Question 1)
Downstream benchmark comparison under matched protocols. Following your recommendation, we implemented the sequential reverse-synthesis pipeline of ReSocratic—composing optimization expressions and application scenarios in text space and back-translating them into natural-language word problems—under the exact same corpus size (N=3,461), teacher model (
DeepSeek-V4-Flash), student backbone (Qwen3-8B), and HVE training protocol (evaluating both SFT-only and SFT + Gated DPO):A1, Table 3)A2, SFT only, Table 3)A3in revised Table 3)B1in Table 3)Why OR-Diffusion substantially outperforms ReSocratic under strictly matched protocols (+7.8 points under SFT-only; +12.0 points under SFT + Gated DPO): ReSocratic composes constraints sequentially in text/expression space without a formal feasibility construction. Under our matched budget, sequentially recombining seed-like constraint clauses and filtering for solver feasibility leaves surviving variable-type/degree signatures inside the existing OR-Instruct seed support (0.0% unseen constraint signatures under the subsampling protocol of revised Table 12—below even surface paraphrasing
A2at 2.9%). Consequently, ReSocratic reaches 64.7% under SFT-only (7.8 points belowOR-DiffusionSFT-onlyB1at 72.5%) and 66.2% under SFT + Gated DPO (A3, 12.0 points below fullOR-Diffusionat 78.2%), with the sharpest deficits on structurally complex benchmarks (ComplexLP: 55.0% vs. 65.8%;CompOR: 33.3%–38.9% vs. 55.6%). Without a witness-point anchor, combining multiple tight constraints in text space can produce mutually inconsistent systems that fail solver verification, leaving survivors concentrated on simple seed skeletons. Notably, gated DPO adds only +1.5 points on ReSocratic (64.7% → 66.2%) versus +5.7 points on OR-Diffusion (72.5% → 78.2%), confirming that preference optimization cannot compensate for narrow structural diversity. In contrast, OR-Diffusion denoises bipartite topologies jointly and instantiates right-hand sides around an explicit integral, non-negative witness point xw, guaranteeing 100% solver feasibility by construction (submission Table 7 / revised Table 9) and preserving complex multi-constraint topologies. We have added the ReSocratic baselines directly to Table 1 (main comparison), Table 3 (A3ablation), and Table 12 (submission Table 9, structural diversity) in the revised manuscript.