Expanding the Decision Space of Pretrained Models through Structured Re-Decoding
Authors: Tinglan Wang, Qun Su, Shihang Wang, Chi Wang, Lin Wang, Shaobo Chen, Haixu Liu, Hai Wei, Yin Ma, Yu Kang, Qi Gao, Xiaojun Yao, Tingjun Hou, Kai Wen
Organizations: Beijing QBoson Quantum Technology Co., Ltd., Beijing, China · College of Pharmaceutical Sciences, Zhejiang University, Hangzhou, China. · Shanghai Innovation Institute, Shanghai, China. · Faculty of Applied Sciences, Macao Polytechnic University, Macao, China. · Department of Chemical Engineering, Tsinghua University, Beijing, China · Center for AI and Computational Biology, Institute of Systems Medicine, Chinese Academy of Medical Sciences, Suzhou, China.
Native decoders limit how pretrained models combine local predictions under new structural constraints. We introduce Structured Re-Decoding (SRD), a training-free framework that constructs local candidates from model scores, samples, or external specifications, jointly optimizes their selection, and verifies reconstructed outputs. We show that candidate recombination expands feasible support and can improve attainable reward under a fixed KL divergence budget. Across molecular generation, protein design, RNA structure prediction, and antibody modeling, SRD enables improved structured inference, including verified molecules absent from the original sample pool. External chemical templates further extend SRD to noncanonical amino acids, peptide cyclization, and grammar-guided molecular construction. Preliminary experiments on a physical coherent Ising machine demonstrate the feasibility of hardware-based optimization, suggesting broader opportunities for quantum and quantum-inspired approaches to structured inference and combinatorial optimization across scientific domains.
Figures & tables
Figure 1: Structured re-decoding of frozen pretrained models and its applications. (a) Model interface. Local candidates come from prediction scores or aligned complete samples. Light-blue c3 illustrates a score-derived alternative absent from the displayed samples. (b) Joint re-decoding. Model scores, pairwise compatibility, and constraint penalties form the QUBO matrix. Colored regions indicate interaction scopes; overlaps denote shared candidates. Green brackets group alternatives for each decision unit, with exactly one selected per group. The illustrated four groups of three states define 34 assignments before constraints. Dark-blue nodes mark the selected assignment. (c) Verified output. The assignment is reconstructed into a complete object and checked against the task’s acceptance policy. “New output” denotes an object outside the original complete-output pool, obtained by recombining observed states or using score-derived alternatives. The green check denotes acceptance. (d) Applications. The upper row presents molecular composition re-decoding (FLOWR), protein sequence-support analysis (ProteinMPNN), RNA pairing (KnotFold), and antibody side-chain coordination (DiffAb), from left to right. The lower row illustrates extensions to noncanonical amino-acid substitution, constrained peptide cyclization, and small-molecule-to-peptide decoding. NVL and HLU label the illustrated noncanonical residues. Structures illustrate the objects and operations studied.
Study
Candidate provenance
Joint decision
Baselines and attribution
Primary endpoint
FLOWR
First output’s terminal atom, charge, and bond scores
Bounded fixed-coordinate edits with composition targets
Best-of-four shares generation but changes the available decisions and uses screening. Local versus full-domain search shares support and objective.
Observed states from the same K feasible native draws
Exact feasible product support on four sites
Whole-pool versus recombined support shares samples, constraints, reward, and conditional reference.
Retained mass, minimum KL, and reward frontier
KnotFold
Frozen pairing scores; eligibility depends on decoder
Pair exclusivity and an optional stacking reward
Canonical matching provides the zero-reward objective reference. Local and full-domain stacking share support and objective.
Reference macro F1
DiffAb
Score-derived identities and constructed rotamers
Pairwise side-chain compatibility
Both arms share candidates. The pairwise arm adds a hard-pair fallback. A fixed-AA control holds identities fixed.
Specified steric endpoints before the optional gate
Table 1: The four primary evaluations and their attribution. Full-domain search refers to the entire declared candidate domain. Objective certificates and downstream verification are tracked separately.
Figure 2: Structured re-decoding (SRD) compared with task-specific baselines. (a) FLOWR single-target success requires composition and quality checks, over 1,800 requests with matched source-generation budgets. Local and full-domain SRD optimize the same objective with different search scopes. The gain is full-domain minus best-of-four success; the lower row reports quality passage under each fallback policy. (b) Exact ProteinMPNN support analysis using the same K=64 feasible draws. Bars show mean conditional mass; the lower row gives mean per-stream minimum KL across six four-site domains, each with four streams. The ratio compares mean masses. (c) Native KnotFold flow and full-domain stacking SRD are connected within each dataset, with absolute macro F1 listed alongside. Gray diamonds show canonical matching. Table 4 reports paired changes and 95% sequence-cluster intervals. (d) Mean DiffAb overlap before and after pairwise SRD. The baseline is unary selection. Below each group, Δ is SRD minus unary with its paired 95% complex-cluster interval (50 complexes, 3,600 cases). The AA-and-rotamer policy uses a hard-pair fallback; the fixed-AA control uses a sequence-preserving fallback.
Target
Method
Success, n (%)
Quality pass, n (%)
Single
Original output
143 (7.944)
1,619 (89.944)
Single
Best-of-4
419 (23.278)
1,797 (99.833)
Single
Local quadratic search
1,307 (72.611)
1,701 (94.500)
Single
Full-domain re-decoding
1,309 (72.722)
1,699 (94.389)
Dual
Original output
36 (2.000)
1,619 (89.944)
Dual
Best-of-4
119 (6.611)
1,797 (99.833)
Table 2: FLOWR delivery across all 1,800 scheduled requests. Success requires both the composition target and the specified quality checks; percentages use the full request denominator. The single target is N+O=4; the dual target also requires one halogen. Coordinates remain fixed with at most eight changed atom, charge, or bond decisions. Quality uses the declared RDKit and PoseBusters 0.6.5 mol_fast checks.
K
Support
Distinct states
pC mass (%)
Min. KL (nat)
4
Whole pool
3.96
2.0679
5.405006
4
Local recombination
37.46
8.1172
3.612204
16
Whole pool
15.58
6.1702
3.505393
16
Local recombination
826.96
33.9579
1.184234
64
Whole pool
57.67
14.7158
2.351939
64
Local recombination
8,041.79
77.4694
0.258796
Table 3: Support formed from the same K feasible native samples. Values average within each backbone and then equally across six backbones; the 24 streams are nested replicates. Minimum KL is the mean of per-stream −logpC(S) , not the negative logarithm of mean retained mass.
File (n)
Independent
Native flow
Canonical
Local stack
Full stack
Full − native (pp)
ArchiveII (393)
72.525
81.004
83.835
84.298
84.317
+3.313 [2.877, 3.739]
TS0 (1,305)
68.790
73.908
76.247
74.765
74.585
+0.677 [0.279, 1.073]
TS1 (60)
75.660
79.568
78.107
79.222
79.098
− 0.470 [ − 1.570, 0.800]
TS2 (38)
80.892
84.113
82.308
84.649
84.929
+0.816 [ − 1.129, 3.119]
TS3 (18)
66.365
68.241
66.855
68.808
70.438
+2.196 [ − 1.379, 6.826]
bpnew (5,401)
76.446
81.299
83.419
82.499
82.436
+1.137 [0.978, 1.297]
Table 4: RNA macro F1 (%) over all 7,215 records. The final column reports full-domain stacking minus native flow in percentage points with 95% sequence-hash-cluster bootstrap intervals. File sizes are shown in parentheses; files are evaluated separately.
Matrices
Spins
Raw
Decoded
Main (50)
177–521
48/2/0
0/49/1
Enlarged (7)
1,021–2,446
7/0/0
1/6/0
Table 5: Physical CIM execution of antibody objectives. Entries are CIM wins/ties/losses versus FSA-10. Each solver contributes ten retained states; physical trial budgets differ. Raw comparison maximizes s⊤Js ; decoded comparison minimizes the molecular objective after strict restoration, repair, and common polishing.
Extension
Qualified
Counting unit
Noncanonical AA
8/8
Selected designs; one complex
Linear to cyclic
2/4
Source pools; one sequence–receptor pair
Linear tripeptide
19/32
Requests; fresh coordinates
Cyclic tetrapeptide
28/32
Requests; fresh coordinates
Table 6: Peptide capability outputs from QUBO re-decoding. Counts use the stated unit and each study’s chemical and geometric checks. Fresh coordinates use ETKDGv3/MMFF94s.
Appendix figures & tables19 assets
Supplementary material from the paper’s appendix.
Appendix
Study
Candidates and scores
Objective and constraints
Search, budget, and interpretation
FLOWR: best-of-4 to re-decoding
Complete outputs versus first-output terminal-score states
Same delivery target and checks; re-decoding adds composition guidance and an edit cap
Same four generated outputs; postprocessing differs. Tests the expanded decision procedure.
FLOWR: local to full-domain
Same states and terminal scores
Same objective, chemistry constraints, and edit cap
Same eight one-second solve caps; only local restricts updates to four groups. Tests search scope.
RNA: independent/native to canonical
Same score matrices; canonical filter changes eligible edges
Pairing restrictions and native decoding rules differ
Independent selection, native flow, or matching. Tests structural decoding, with unmatched solver costs.
RNA: canonical to stacking
Same scores and canonical pair rules; negative edges may matter with stacking
Degree constraints shared; adds a unit stacking reward
Matching is the zero-reward objective reference. Tied optima can yield different predictions.
RNA: local to full-domain stack
Same candidate edges and scores
Same degree constraints and stacking objective
Up to 500 local updates versus 60-second MILP with numerical optimality status. Tests search under distinct limits.
DiffAb: unary to pairwise
Same amino-acid/rotamer states and unary costs
Same context filtering; adds pair terms and hard-pair fallback
Per-site minimization versus twelve coordinate sweeps. Tests pairwise modeling plus its policy.
Appendix
Table S1: Detailed contrasts for the four primary evaluations. Each row identifies the changes in candidate support, objective, or search. Search limits and objective certificates are reported separately.
File
Records
Unique
Train names
Development
Full − native [95% interval]
Full − canonical [95% interval]
ArchiveII
393
383
0
0
3.313 [2.877, 3.739]
0.482 [0.141, 0.808]
TS0
1,305
1,305
714
0
0.677 [0.279, 1.073]
− 1.662 [ − 2.028, − 1.296]
TS1
60
60
0
0
− 0.470 [ − 1.570, 0.800]
0.992 [ − 0.119, 2.319]
TS2
38
38
0
21
0.816 [ − 1.129, 3.119]
2.622 [0.836, 4.852]
TS3
18
18
0
0
2.196 [ − 1.379, 6.826]
3.583 [0.443, 7.907]
bpnew
5,401
5,401
0
0
1.137 [0.978, 1.297]
− 0.983 [ − 1.134, − 0.833]
Appendix
Table S2: RNA provenance and paired macro-F1 differences for full-domain stacking versus native flow and canonical matching. Differences and pointwise 95% sequence-hash-cluster intervals are in percentage points. The training-name count uses the released filename list.
Stratum
n
Indep.
Native
Canon.
Local
Full
Full − native [95% interval]
TS0: training-name match
714
74.140
78.879
81.409
79.788
79.605
0.726 [0.154, 1.331]
TS0: no name match
591
62.327
67.903
70.010
68.697
68.520
0.617 [0.104, 1.143]
TS2: development
21
78.325
82.637
81.373
83.165
83.673
1.036 [ −1.536 , 4.319]
TS2: outside development
17
84.063
85.937
83.462
86.481
86.481
0.544 [ −2.189 , 3.579]
Appendix
Table S3: RNA performance stratified by documented exposure. All F1 values are percentages; differences and 95% sequence-hash-cluster intervals are percentage points. “No name match” refers only to the released training-name list, and “outside development” refers to the documented 21-record set.
File (records)
Candidate edges
Stacking terms
Solve time (s)
ArchiveII (393)
82 / 712.2 / 1583
47 / 401 / 977
0.0084 / 0.3798 / 1.8553
TS0 (1,305)
66 / 449.6 / 1583
41 / 273 / 977
0.0060 / 0.1659 / 1.9042
TS1 (60)
42.5 / 226.1 / 1065
23 / 134.2 / 749
0.0035 / 0.0467 / 1.0484
TS2 (38)
28.5 / 97.9 / 239
19 / 72.3 / 138
0.0026 / 0.0135 / 0.0474
TS3 (18)
74.5 / 289.2 / 398
49.5 / 177.4 / 242
0.0073 / 0.0755 / 0.1326
bpnew (5,401)
49 / 246 / 1348
32 / 153 / 878
0.0043 / 0.0536 / 1.4979
Appendix
Table S4: RNA problem dimensions and recorded MILP solve time: median / 95th percentile / maximum within each file. All 7,215 records are represented. Solve time excludes model inference and candidate compilation; these are recorded component times. Percentiles use linear interpolation.
Endpoint
Unary
Pairwise
Pairwise − unary [95% interval]
Normalized overlap, Å 2 /atom
0.327344
0.261613
− 0.065732 [ − 0.078456, − 0.054008]
Clashes per 1,000 atoms
320.732926
282.859400
− 37.873525 [ − 44.410927, − 31.643099]
Maximum overlap, Å
1.399863
1.236854
− 0.163009 [ − 0.195692, − 0.132723]
Minimum distance, Å
1.796311
1.932882
0.136571 [0.109991, 0.166171]
Appendix
Table S5: Full CDR geometry comparison. Differences and intervals retain the precision of the paired source analysis, so subtracting rounded displayed means can differ in the last digit.
Proxy identity
Complexes
Cases
Δ overlap, Å 2 /atom
Δ clashes/1,000 atoms
100% shared positions
4
288
−0.12564 [ −0.16231 , −0.09454 ]
−69.172 [ −87.405 , −55.251 ]
90% to below 100%
9
648
−0.06815 [ −0.08980 , −0.04760 ]
−39.620 [ −51.490 , −28.398 ]
Below 90%
37
2,664
−0.05867 [ −0.07285 , −0.04590 ]
−34.065 [ −41.599 , −27.224 ]
Appendix
Table S6: Primary DiffAb geometry by disjoint Fv-proxy strata. Differences are pairwise minus unary, with equal-complex means and 95% complex-bootstrap intervals. Every complex retains 72 cases. Identity uses coverage-qualified shared positions; these strata do not identify checkpoint training membership.
Quantity per CDR
Unary
Pairwise
Difference [95% interval]
Side-chain heavy atoms
29.54778
28.88889
−0.65889 [ −0.78389 , −0.54083 ]
Placed CDR heavy atoms
65.65444
64.99556
−0.65889 [ −0.78333 , −0.54167 ]
Base composite cost
18.31982
18.37356
+0.05374 [0.04399, 0.06442]
Full candidate cost
97.97434
69.01861
−28.95574 [ −33.73315 , −24.43085 ]
Appendix
Table S7: Composition and composite costs in the primary DiffAb comparison. Means average cases within complex, then weight the 50 complexes equally. Costs are archived composite energies, which include model and heuristic terms.
Fixed identities
Evaluable/planned
Δ overlap, Å 2 /atom
Δ clashes/1,000 atoms
Emitted DiffAb sequence
3,600/3,600
−0.066829 [ −0.075667 , −0.058381 ]
−34.370 [ −38.273 , −30.455 ]
Unary-selected sequence
2,795/3,600
−0.049793 [ −0.061750 , −0.038884 ]
−25.482 [ −31.126 , −20.334 ]
Appendix
Table S8: Archived exploratory fixed-identity DiffAb controls. Paired geometry uses equal-complex means and the original 95% complex-bootstrap intervals. All 50 complexes contribute. Evaluable pairs are shown against the full 3,600-case denominator; these controls are outside the primary Holm family.
Target
Original
Best-of-4
Local
Full
Full quality passes
Single
381
1,195
873
946
1,904
Dual
62
266
174
196
1,903
Appendix
Table S9: SemlaFlow target-success counts out of all 2,048 requests. The final column reports verified quality passes for the full-domain method.
Dataset/contract
Original
Best-of-8
Local
Full
Open-support rule
RICO/C1
1,358
1,388
1,406
1,406
1,517
RICO/C2
1,358
1,368
1,405
1,405
1,516
PubLayNet/C1
8,469
9,163
9,618
9,622
9,616
PubLayNet/C2
8,469
8,491
9,538
9,551
9,595
Appendix
Table S10: Layout deliveries satisfying the specified geometric contract. Denominators are 1,518 for RICO and 9,810 for PubLayNet.
Guidance λ
Dual success (%)
KL to pC (nat)
Joint TV to pC
Entropy (nat)
0
7.3509
0
0
7.570392
1
13.1567
0.044079
0.120264
7.695946
2
21.8102
0.190252
0.241685
7.788548
4
46.0854
0.777334
0.449763
7.784909
Appendix
Table S11: Exact dual-property target on the full feasible domain, equally weighted across six backbones. These are target-distribution quantities, not empirical MCMC estimates.
Backbone
Pool mass (%)
Recombination mass (%)
Reward gain
Strict-gain streams
1CRN
4.6796
77.7950
0.620710
4/4
1L2Y
3.9794
80.5786
0.487415
4/4
1LE0
11.2977
74.5258
0.039529
3/4
1UBQ
7.0123
75.2626
0.534213
4/4
1VII
10.6920
72.4432
0.515815
4/4
2GB1
50.6339
84.2110
0.138048
4/4
Appendix
Table S12: Exact support results at K=64 by backbone, averaging its four source streams. Reward gain compares recombination with whole-pool support at the common KL budget of 4 nat for the dual-target reward. Streams are nested replicates.
Budget
Proposal
Single MAE
Dual MAE
Joint TV
Unvisited mass
5 s
Native MH
3.793
2.525
0.7483
0.7012
5 s
Conditional product MH
1.258
0.820
0.4261
0.3485
5 s
Mixed-center MH
2.153
1.924
0.5142
0.4516
15 s
Native MH
2.284
1.975
0.6499
0.5810
15 s
Conditional product MH
0.900
0.705
0.3043
0.2164
15 s
Mixed-center MH
2.080
1.413
0.3558
0.2709
Appendix
Table S13: Complete proposal diagnostic. Each row contains 36 runs. Success errors are measured in percentage points relative to the exact target expectations.
Draw budget
Distinct labels
Expected binary score
Regret to best policy
8
7.585
0.883658
0.021957
16
14.320
0.887004
0.018611
32
25.885
0.888038
0.017578
64
44.250
0.893146
0.012469
128
71.825
0.896124
0.009491
256
111.745
0.900631
0.004984
Appendix
Table S14: Policy selection from public GB1 outcomes, averaged over 200 query streams from one assay.
Study
Candidates and scores
Objective and constraints
Search, budget, and interpretation
Linear-model cyclization
Supplied sequence; 32 sampled linear poses per pool
Cyclic graph, fixed anchors, closure and geometry checks
QUBO results over four pools, with 32 closure starts each. Tests geometric cyclization capability.
Noncanonical substitutions
External chemistry/rotamers and backbone-only model scores
Exactly two substitutions; physical and family-preference terms
Exact atom coverage and peptide topology; soft editing costs
Per-template mapping QUBOs; chemistry and two separate coordinate routes checked. Tests specification delivery.
Thioether poses
Fixed Vina pose fragments; source-score and geometric costs
One state per block; cross-block bond and clash restrictions
All 20,736 assignments enumerated; full geometry checks follow. Tests fragment reconstruction.
Appendix
Table S15: Peptide capability settings. These studies verify different chemical or geometric outputs under their own candidate domains and budgets.
Rank
Substitutions
Physical (REU)
Family cost (nat)
Objective (REU)
Qualified
1
B142:HLU, B144:NVL
-6.8190
4.1059
-2.7131
Pass
2
B142:HLU, B152:NVL
-7.2086
4.9253
-2.2833
Pass
3
B141:NVL, B142:HLU
-8.1620
7.2067
-0.9554
Pass
4
B142:NVL, B144:NVL
-5.1826
6.2826
1.1000
Pass
5
B142:NVL, B152:NVL
-5.5722
7.1021
1.5299
Pass
6
B141:NVL, B142:NVL
-6.5256
9.3834
2.8579
Pass
Appendix
Table S16: Eight distinct noncanonical designs selected by QUBO re-decoding on the deposited PUMA–MCL1 backbone. Each design has exactly two new substitutions and one selected conformer. Physical values use the finite-library strain reference; the QUBO soft objective adds family cost at one REU/nat. Qualification requires all declared chemical and geometric checks.
New source pool
Closure starts
Geometry-qualified
Qualified RMSD ≤2 Å
1
32
0
0
2
32
2
0
3
32
1
0
4
32
0
0
Total
128
3
0
Appendix
Table S17: QUBO cyclization results for four new source pools of one supplied 9HGD/GABARAP sequence. Qualified counts require all declared chemical and geometric checks after closure and refinement. Denominators are closure starts. The final column additionally requires receptor-aligned heavy-atom RMSD ≤2 Å.
Specification
Size eligible
Fresh delivery
Source delivery
Unique delivered
QUBO bits
Linear tripeptide
19/32
19/32
2/32
17
169–484
Cyclic tetrapeptide
28/32
28/32
3/32
22
256–729
Appendix
Table S18: QUBO peptide-grammar capability on the same 32 cached FlowMol3 endpoints. Requests of incompatible size remain in the denominator. Fresh coordinates are generated by ETKDGv3/MMFF94s; the source route relaxes original model coordinates. Each route reports any qualified result among at most four checked graphs per request. Unique delivered graphs count the first qualified fresh-route result per request.
Output
Rank
Source seeds for the four blocks
Full RMSD (Å)
Bridge RMSD (Å)
Accepted delivery
2
510019, 510019, 510019, 510019
10.775
10.768
Qualified recombination
12
510020, 510020, 510026, 510026
11.990
12.887
Qualified recombination
14
510004, 510004, 510004, 510019
10.951
12.363
Appendix
Table S19: Verified 5KEZ outputs in QUBO energy order. The accepted delivery retains one complete source; two later qualified outputs combine sources. All rows pass the declared chemical and geometric checks. RMSD uses the fixed receptor frame and graph-valid symmetry mappings, without separately aligning the ligand.
Constrained decoding steers language model outputs to satisfy formal or task-specific rules while balancing validity, quality, and faithful probabilities.
Large language models (LLMs) are increasingly used to generate executable outputs, JSON objects, and API calls, where a single syntax error can make the output unusable. Constrained decoding enforces validity token-by-token via masking and renormalization, but it can distort generation when the model assigns low probability mass to valid continuations, pushing decoding toward locally valid yet semantically incorrect trajectories. We propose Draft-Conditioned Constrained Decoding (DCCD), a simple two-step, training-free inference procedure that decouples semantic planning from structural enforcement: an unconstrained draft is generated first, and constrained decoding is then applied, conditioned on this draft, to guarantee validity. We analyze DCCD through a KL-projection view, showing that draft conditioning increases feasible mass and reduces the cumulative "projection tax" induced by hard constraints, with an optional best-of-K draft selection. Across structured reasoning benchmarks, DCCD improves strict structured accuracy by up to +24 percentage points over standard constrained decoding (e.g., 15.2% to 39.0% on GSM8K with a 1B model), and enables smaller model pairs to match or exceed much larger constrained baselines, yielding substantial gains in parameter efficiency.
Avinash Reddy, Thayne T. Walker, James S. Ide +1
Department of Computer Science, University of Central Florida, FL, US · Lockheed Martin AI Center, CT, US
Inference-time decoding methods improve LLM reasoning by exploring multiple candidate trajectories, yet treat each trajectory as atomic: either retaining it whole or discarding it irreversibly. This wastes computation on partially promising candidates whose high-quality prefixes are abandoned alongside degraded suffixes. We introduce Selective Regenerative Decoding (SRD), which routes each candidate to discard, keep, or refine only the degraded portion of the suffix while preserving the useful prefix of borderline candidates, without requiring a larger target model. Under mild assumptions, SRD achieves a provable 1.28-to-1.36-fold gain in sample efficiency over rejection sampling with strictly higher expected trajectory quality, with the gain growing as the candidate pool grows. Across MATH500, GPQA Diamond, HotpotQA, and AlpacaEval with multiple generation-reward model pairs, SRD matches Best-of-N accuracy with substantially fewer generated tokens and outperforms speculative rejection in low-compute regimes. By enabling segment-level intervention rather than whole-trajectory selection, SRD opens a previously underexplored region of the accuracy-compute tradeoff for inference-time reasoning.
Sophia Xiao Pu, Yumo Xu, Sailik Sengupta +5
University of California, Santa Barbara · Netflix · Amazon Science +1
Speculative decoding accelerates large language model inference by employing a compact draft model to propose candidate tokens that are verified in parallel by a target model. Conventional training relies on soft-label distillation, which forces the draft model to mimic the target's full output distribution, including high-entropy regions that degrade acceptance rates and increase early rejection risk. We propose RCSD (Reinforcement and Continual Learning for Speculative Decoding), a training framework that aligns draft model optimization with the system-level objective of maximizing acceptance length. RCSD leverages naturally occurring accepted and rejected sequences during speculative inference to construct implicit preference pairs, with masked loss to avoid penalizing target-aligned tokens. Evaluated across multiple benchmarks, RCSD improves average acceptance length by 35% over HASS and achieves 1.35x faster end-to-end inference than HASS on practical workloads, without compromising generation quality.