Configuration: HyInv BF16 path (update 288, equal-diversity read-out) (HyInv report) Protocol: BF16 path minus reference sampler in scTM at fixed diversity (HyInv report) Dataset: CATH4.3 validation (1,529 chains, development panel) −0.0001 [−0.0009, +0.0016] Sctm at fixed diversity difference (BF16 path minus the FP32 reference sampler at development diversity D = 0.30, four-point quadratic fit, paired by chain, with a 95% interval; ESMFold2, 2 designs per chain, draft temperature 0.6; checkpoint at update 288; the reference system is the FP32 reference sampler)unitless · higher
Uncertainty: 95% CI -0.0009 to +0.0016. 95% interval over CATH topology clusters. The report's own pass rule is whether the lower end is at or above the -0.002 margin fixed before the check; for this row the printed verdict is 'yes'. The report does not state how many resamples this interval comes from.
Coverage: 1529/1529 chains scored (91 CATH topology clusters.)
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons Appendix C states 'We do not claim zero sequence overlap between the training data and the evaluation panels', and Appendix G's homology screen is a coverage rule under which partial hits remain (Table 10: 28 of 74 de novo backbones and 7 of 12 pocket backbones have partial hits, and all 4 named enzymes are homologous to training data by design). Results on these panels are therefore not clean held-out measurements and must not be compared with results from sources that do establish no overlap.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix C 'Data and pool', final bullet; Appendix G 'Homology of the panels' and Table 10
The ten baselines are single published settings at one temperature each, so their design-set diversity differs from HyInv's: on CATH4.3 test HyInv samples at diversity 0.331 against 0.093 for PiFold, 0.104 for CMLM, 0.144 for LM-Design 650M and 0.000 for DPLM-2 3B, which decodes by argmax. Appendix D.2 states that these four differences are not at matched diversity, and Appendix A states that even for the closest baselines the comparisons are 'close to, but not exactly at, equal diversity', the largest offset of 0.016 being worth at most 0.0011 scTM at the development slope. The scTM and scTM difference results of the protocols named here must not be read as a matched-diversity ranking. The results read at a stated diversity level (Tables 8 and 9 and Appendix D.3) are not covered by this entry, because they are the ones the report reads on a fitted curve at a fixed diversity; nor are the recovery, diversity, pLDDT, scRMSD and lDDT results, whose protocols carry the same single-setting caveat in their own limitations and whose metric qualifiers say the value is at each method's own setting.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix D.1 and D.2; Table 7; Figure 14 diversity block; Section 5.1 'Baselines' and 'HyInv'
The inference-time results were measured with the supervised-stage weights and a rewrite temperature of 0.23, not with the released checkpoint and its default temperatures, and the quality read-outs of the two numerical paths used the RL checkpoint at update 288, not update 640. Speed and quality in this report are therefore measured on different weights, and the 57x figure compares HyInv's optimised BF16 path with ProteinMPNN's released implementation rather than with an equally optimised baseline.HyInv Technical Report (Hyra Team, 2026-10-14) · Table 1 footnote; Appendix F 'Timing protocol' and 'Baseline timings'
Methods, coverage and source BF16 path minus the reference sampler on Whole development panel at fixed diversity (HyInv report)
hyinv-20261106-protocol-sctm-fixed-diversity-bf16-minus-reference
Aggregation: Not reported
HyInv Technical Report (Hyra Team, 2026-10-14) · Table 9, row 'Whole development panel', column 'S@D difference' (page 20 of 21) Configuration: HyInv plain sampling (update 640) (HyInv report) Protocol: Gain in scTM at fixed diversity from reinforcement learning (HyInv report) Dataset: CATH4.3 validation (1,529 chains, development panel) 0.0133 Sctm at fixed diversity difference (Update-640 checkpoint minus the supervised checkpoint at development diversity D = 0.30, plain single-temperature sampling, ESMFold2, development panel; the reference system is the supervised checkpoint Both values are read from least-squares quadratic fits of scTM against design-set diversity over a rewrite-temperature grid, so the level is interpolated from a fit and is not a measurement at a matched sampling setting.)unitless · higher
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons Appendix C states 'We do not claim zero sequence overlap between the training data and the evaluation panels', and Appendix G's homology screen is a coverage rule under which partial hits remain (Table 10: 28 of 74 de novo backbones and 7 of 12 pocket backbones have partial hits, and all 4 named enzymes are homologous to training data by design). Results on these panels are therefore not clean held-out measurements and must not be compared with results from sources that do establish no overlap.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix C 'Data and pool', final bullet; Appendix G 'Homology of the panels' and Table 10
The ten baselines are single published settings at one temperature each, so their design-set diversity differs from HyInv's: on CATH4.3 test HyInv samples at diversity 0.331 against 0.093 for PiFold, 0.104 for CMLM, 0.144 for LM-Design 650M and 0.000 for DPLM-2 3B, which decodes by argmax. Appendix D.2 states that these four differences are not at matched diversity, and Appendix A states that even for the closest baselines the comparisons are 'close to, but not exactly at, equal diversity', the largest offset of 0.016 being worth at most 0.0011 scTM at the development slope. The scTM and scTM difference results of the protocols named here must not be read as a matched-diversity ranking. The results read at a stated diversity level (Tables 8 and 9 and Appendix D.3) are not covered by this entry, because they are the ones the report reads on a fitted curve at a fixed diversity; nor are the recovery, diversity, pLDDT, scRMSD and lDDT results, whose protocols carry the same single-setting caveat in their own limitations and whose metric qualifiers say the value is at each method's own setting.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix D.1 and D.2; Table 7; Figure 14 diversity block; Section 5.1 'Baselines' and 'HyInv'
The inference-time results were measured with the supervised-stage weights and a rewrite temperature of 0.23, not with the released checkpoint and its default temperatures, and the quality read-outs of the two numerical paths used the RL checkpoint at update 288, not update 640. Speed and quality in this report are therefore measured on different weights, and the 57x figure compares HyInv's optimised BF16 path with ProteinMPNN's released implementation rather than with an equally optimised baseline.HyInv Technical Report (Hyra Team, 2026-10-14) · Table 1 footnote; Appendix F 'Timing protocol' and 'Baseline timings'
Methods, coverage and source HyInv update 640 over the supervised checkpoint at development diversity 0.30 (HyInv report)
hyinv-20261106-protocol-sctm-fixed-diversity-gain-vs-supervised-checkpoint
Aggregation: Not reported
HyInv Technical Report (Hyra Team, 2026-10-14) · Section 5.4 'Reinforcement learning', parenthesis '(+0.0133, Figure 8a)' (page 7 of 21) Configuration: HyInv decoupled draft Td = 0.3 (update 288) (HyInv report) Protocol: Gain in scTM at fixed diversity over the unmodified sampler (HyInv report) Dataset: CATH4.3 validation (1,529 chains, development panel) +0.0054 [+0.0022, +0.0067] Sctm at fixed diversity difference (Setting minus the unmodified single-temperature sampler of the same checkpoint on the same panel, both read at fixed diversity (D = 0.30 on the development panel, 0.32 on CATH4.3 test, 0.30 on RecentPDB), paired, with a 95% interval; ESMFold2; checkpoint at update 288; the reference system is the unmodified sampler; read at development diversity D = 0.30 Both values are read from least-squares quadratic fits of scTM against design-set diversity over a rewrite-temperature grid, so the level is interpolated from a fit and is not a measurement at a matched sampling setting.)unitless · higher
Uncertainty: 95% CI +0.0022 to +0.0067. 95% interval over CATH topology clusters, over chains on RecentPDB; Appendix A states the interval comes from 1,000 draws in which all points of all curves are recomputed from one cluster resample.
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons Appendix C states 'We do not claim zero sequence overlap between the training data and the evaluation panels', and Appendix G's homology screen is a coverage rule under which partial hits remain (Table 10: 28 of 74 de novo backbones and 7 of 12 pocket backbones have partial hits, and all 4 named enzymes are homologous to training data by design). Results on these panels are therefore not clean held-out measurements and must not be compared with results from sources that do establish no overlap.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix C 'Data and pool', final bullet; Appendix G 'Homology of the panels' and Table 10
The ten baselines are single published settings at one temperature each, so their design-set diversity differs from HyInv's: on CATH4.3 test HyInv samples at diversity 0.331 against 0.093 for PiFold, 0.104 for CMLM, 0.144 for LM-Design 650M and 0.000 for DPLM-2 3B, which decodes by argmax. Appendix D.2 states that these four differences are not at matched diversity, and Appendix A states that even for the closest baselines the comparisons are 'close to, but not exactly at, equal diversity', the largest offset of 0.016 being worth at most 0.0011 scTM at the development slope. The scTM and scTM difference results of the protocols named here must not be read as a matched-diversity ranking. The results read at a stated diversity level (Tables 8 and 9 and Appendix D.3) are not covered by this entry, because they are the ones the report reads on a fitted curve at a fixed diversity; nor are the recovery, diversity, pLDDT, scRMSD and lDDT results, whose protocols carry the same single-setting caveat in their own limitations and whose metric qualifiers say the value is at each method's own setting.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix D.1 and D.2; Table 7; Figure 14 diversity block; Section 5.1 'Baselines' and 'HyInv'
The inference-time results were measured with the supervised-stage weights and a rewrite temperature of 0.23, not with the released checkpoint and its default temperatures, and the quality read-outs of the two numerical paths used the RL checkpoint at update 288, not update 640. Speed and quality in this report are therefore measured on different weights, and the 57x figure compares HyInv's optimised BF16 path with ProteinMPNN's released implementation rather than with an equally optimised baseline.HyInv Technical Report (Hyra Team, 2026-10-14) · Table 1 footnote; Appendix F 'Timing protocol' and 'Baseline timings'
Methods, coverage and source Decoupled draft, Td = 0.3 over the unmodified sampler on Development at equal diversity (HyInv report)
hyinv-20261106-protocol-sctm-fixed-diversity-gain-vs-unmodified-sampler
Aggregation: Not reported
HyInv Technical Report (Hyra Team, 2026-10-14) · Table 8, row 'Decoupled draft, Td = 0.3', column 'Development' (page 19 of 21) Configuration: HyInv decoupled draft Td = 0.6 (update 288) (HyInv report) Protocol: Gain in scTM at fixed diversity over the unmodified sampler (HyInv report) Dataset: CATH4.3 validation (1,529 chains, development panel) +0.0072 [+0.0062, +0.0118] Sctm at fixed diversity difference (Setting minus the unmodified single-temperature sampler of the same checkpoint on the same panel, both read at fixed diversity (D = 0.30 on the development panel, 0.32 on CATH4.3 test, 0.30 on RecentPDB), paired, with a 95% interval; ESMFold2; checkpoint at update 288; the reference system is the unmodified sampler; read at development diversity D = 0.30 Both values are read from least-squares quadratic fits of scTM against design-set diversity over a rewrite-temperature grid, so the level is interpolated from a fit and is not a measurement at a matched sampling setting.)unitless · higher
Uncertainty: 95% CI +0.0062 to +0.0118. 95% interval over CATH topology clusters, over chains on RecentPDB; Appendix A states the interval comes from 1,000 draws in which all points of all curves are recomputed from one cluster resample.
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons Appendix C states 'We do not claim zero sequence overlap between the training data and the evaluation panels', and Appendix G's homology screen is a coverage rule under which partial hits remain (Table 10: 28 of 74 de novo backbones and 7 of 12 pocket backbones have partial hits, and all 4 named enzymes are homologous to training data by design). Results on these panels are therefore not clean held-out measurements and must not be compared with results from sources that do establish no overlap.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix C 'Data and pool', final bullet; Appendix G 'Homology of the panels' and Table 10
The ten baselines are single published settings at one temperature each, so their design-set diversity differs from HyInv's: on CATH4.3 test HyInv samples at diversity 0.331 against 0.093 for PiFold, 0.104 for CMLM, 0.144 for LM-Design 650M and 0.000 for DPLM-2 3B, which decodes by argmax. Appendix D.2 states that these four differences are not at matched diversity, and Appendix A states that even for the closest baselines the comparisons are 'close to, but not exactly at, equal diversity', the largest offset of 0.016 being worth at most 0.0011 scTM at the development slope. The scTM and scTM difference results of the protocols named here must not be read as a matched-diversity ranking. The results read at a stated diversity level (Tables 8 and 9 and Appendix D.3) are not covered by this entry, because they are the ones the report reads on a fitted curve at a fixed diversity; nor are the recovery, diversity, pLDDT, scRMSD and lDDT results, whose protocols carry the same single-setting caveat in their own limitations and whose metric qualifiers say the value is at each method's own setting.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix D.1 and D.2; Table 7; Figure 14 diversity block; Section 5.1 'Baselines' and 'HyInv'
The inference-time results were measured with the supervised-stage weights and a rewrite temperature of 0.23, not with the released checkpoint and its default temperatures, and the quality read-outs of the two numerical paths used the RL checkpoint at update 288, not update 640. Speed and quality in this report are therefore measured on different weights, and the 57x figure compares HyInv's optimised BF16 path with ProteinMPNN's released implementation rather than with an equally optimised baseline.HyInv Technical Report (Hyra Team, 2026-10-14) · Table 1 footnote; Appendix F 'Timing protocol' and 'Baseline timings'
Methods, coverage and source Decoupled draft, Td = 0.6 (HyInv default) over the unmodified sampler on Development at equal diversity (HyInv report)
hyinv-20261106-protocol-sctm-fixed-diversity-gain-vs-unmodified-sampler
Aggregation: Not reported
HyInv Technical Report (Hyra Team, 2026-10-14) · Table 8, row 'Decoupled draft, Td = 0.6 (HyInv default)', column 'Development' (page 19 of 21) Configuration: HyInv min-p 0.1 truncation (update 288) (HyInv report) Protocol: Gain in scTM at fixed diversity over the unmodified sampler (HyInv report) Dataset: CATH4.3 validation (1,529 chains, development panel) +0.0033 [+0.0023, +0.0076] Sctm at fixed diversity difference (Setting minus the unmodified single-temperature sampler of the same checkpoint on the same panel, both read at fixed diversity (D = 0.30 on the development panel, 0.32 on CATH4.3 test, 0.30 on RecentPDB), paired, with a 95% interval; ESMFold2; checkpoint at update 288; the reference system is the unmodified sampler; read at development diversity D = 0.30 Both values are read from least-squares quadratic fits of scTM against design-set diversity over a rewrite-temperature grid, so the level is interpolated from a fit and is not a measurement at a matched sampling setting.)unitless · higher
Uncertainty: 95% CI +0.0023 to +0.0076. 95% interval over CATH topology clusters, over chains on RecentPDB; Appendix A states the interval comes from 1,000 draws in which all points of all curves are recomputed from one cluster resample.
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons Appendix C states 'We do not claim zero sequence overlap between the training data and the evaluation panels', and Appendix G's homology screen is a coverage rule under which partial hits remain (Table 10: 28 of 74 de novo backbones and 7 of 12 pocket backbones have partial hits, and all 4 named enzymes are homologous to training data by design). Results on these panels are therefore not clean held-out measurements and must not be compared with results from sources that do establish no overlap.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix C 'Data and pool', final bullet; Appendix G 'Homology of the panels' and Table 10
The ten baselines are single published settings at one temperature each, so their design-set diversity differs from HyInv's: on CATH4.3 test HyInv samples at diversity 0.331 against 0.093 for PiFold, 0.104 for CMLM, 0.144 for LM-Design 650M and 0.000 for DPLM-2 3B, which decodes by argmax. Appendix D.2 states that these four differences are not at matched diversity, and Appendix A states that even for the closest baselines the comparisons are 'close to, but not exactly at, equal diversity', the largest offset of 0.016 being worth at most 0.0011 scTM at the development slope. The scTM and scTM difference results of the protocols named here must not be read as a matched-diversity ranking. The results read at a stated diversity level (Tables 8 and 9 and Appendix D.3) are not covered by this entry, because they are the ones the report reads on a fitted curve at a fixed diversity; nor are the recovery, diversity, pLDDT, scRMSD and lDDT results, whose protocols carry the same single-setting caveat in their own limitations and whose metric qualifiers say the value is at each method's own setting.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix D.1 and D.2; Table 7; Figure 14 diversity block; Section 5.1 'Baselines' and 'HyInv'
The inference-time results were measured with the supervised-stage weights and a rewrite temperature of 0.23, not with the released checkpoint and its default temperatures, and the quality read-outs of the two numerical paths used the RL checkpoint at update 288, not update 640. Speed and quality in this report are therefore measured on different weights, and the 57x figure compares HyInv's optimised BF16 path with ProteinMPNN's released implementation rather than with an equally optimised baseline.HyInv Technical Report (Hyra Team, 2026-10-14) · Table 1 footnote; Appendix F 'Timing protocol' and 'Baseline timings'
Methods, coverage and source Min-p 0.1 over the unmodified sampler on Development at equal diversity (HyInv report)
hyinv-20261106-protocol-sctm-fixed-diversity-gain-vs-unmodified-sampler
Aggregation: Not reported
HyInv Technical Report (Hyra Team, 2026-10-14) · Table 8, row 'Min-p 0.1', column 'Development' (page 19 of 21) Configuration: HyInv top-p 0.9 truncation (update 288) (HyInv report) Protocol: Gain in scTM at fixed diversity over the unmodified sampler (HyInv report) Dataset: CATH4.3 validation (1,529 chains, development panel) +0.0044 [+0.0028, +0.0063] Sctm at fixed diversity difference (Setting minus the unmodified single-temperature sampler of the same checkpoint on the same panel, both read at fixed diversity (D = 0.30 on the development panel, 0.32 on CATH4.3 test, 0.30 on RecentPDB), paired, with a 95% interval; ESMFold2; checkpoint at update 288; the reference system is the unmodified sampler; read at development diversity D = 0.30 Both values are read from least-squares quadratic fits of scTM against design-set diversity over a rewrite-temperature grid, so the level is interpolated from a fit and is not a measurement at a matched sampling setting.)unitless · higher
Uncertainty: 95% CI +0.0028 to +0.0063. 95% interval over CATH topology clusters, over chains on RecentPDB; Appendix A states the interval comes from 1,000 draws in which all points of all curves are recomputed from one cluster resample.
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons Appendix C states 'We do not claim zero sequence overlap between the training data and the evaluation panels', and Appendix G's homology screen is a coverage rule under which partial hits remain (Table 10: 28 of 74 de novo backbones and 7 of 12 pocket backbones have partial hits, and all 4 named enzymes are homologous to training data by design). Results on these panels are therefore not clean held-out measurements and must not be compared with results from sources that do establish no overlap.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix C 'Data and pool', final bullet; Appendix G 'Homology of the panels' and Table 10
The ten baselines are single published settings at one temperature each, so their design-set diversity differs from HyInv's: on CATH4.3 test HyInv samples at diversity 0.331 against 0.093 for PiFold, 0.104 for CMLM, 0.144 for LM-Design 650M and 0.000 for DPLM-2 3B, which decodes by argmax. Appendix D.2 states that these four differences are not at matched diversity, and Appendix A states that even for the closest baselines the comparisons are 'close to, but not exactly at, equal diversity', the largest offset of 0.016 being worth at most 0.0011 scTM at the development slope. The scTM and scTM difference results of the protocols named here must not be read as a matched-diversity ranking. The results read at a stated diversity level (Tables 8 and 9 and Appendix D.3) are not covered by this entry, because they are the ones the report reads on a fitted curve at a fixed diversity; nor are the recovery, diversity, pLDDT, scRMSD and lDDT results, whose protocols carry the same single-setting caveat in their own limitations and whose metric qualifiers say the value is at each method's own setting.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix D.1 and D.2; Table 7; Figure 14 diversity block; Section 5.1 'Baselines' and 'HyInv'
The inference-time results were measured with the supervised-stage weights and a rewrite temperature of 0.23, not with the released checkpoint and its default temperatures, and the quality read-outs of the two numerical paths used the RL checkpoint at update 288, not update 640. Speed and quality in this report are therefore measured on different weights, and the 57x figure compares HyInv's optimised BF16 path with ProteinMPNN's released implementation rather than with an equally optimised baseline.HyInv Technical Report (Hyra Team, 2026-10-14) · Table 1 footnote; Appendix F 'Timing protocol' and 'Baseline timings'
Methods, coverage and source Top-p 0.9 over the unmodified sampler on Development at equal diversity (HyInv report)
hyinv-20261106-protocol-sctm-fixed-diversity-gain-vs-unmodified-sampler
Aggregation: Not reported
HyInv Technical Report (Hyra Team, 2026-10-14) · Table 8, row 'Top-p 0.9', column 'Development' (page 19 of 21) Configuration: HyInv supervised checkpoint, plain sampling (HyInv report) Protocol: scTM at fixed sequence diversity on the development panel (HyInv report) Dataset: CATH4.3 validation (1,529 chains, development panel) 0.839 Sctm at fixed diversity (scTM at development diversity D = 0.30, read from a least-squares quadratic fit of scTM against design-set diversity over a rewrite-temperature grid, plain single-temperature sampling, ESMFold2, development panel (CATH4.3 validation); the checkpoint is the supervised checkpoint before reinforcement learning)unitless · higher
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons Appendix C states 'We do not claim zero sequence overlap between the training data and the evaluation panels', and Appendix G's homology screen is a coverage rule under which partial hits remain (Table 10: 28 of 74 de novo backbones and 7 of 12 pocket backbones have partial hits, and all 4 named enzymes are homologous to training data by design). Results on these panels are therefore not clean held-out measurements and must not be compared with results from sources that do establish no overlap.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix C 'Data and pool', final bullet; Appendix G 'Homology of the panels' and Table 10
The ten baselines are single published settings at one temperature each, so their design-set diversity differs from HyInv's: on CATH4.3 test HyInv samples at diversity 0.331 against 0.093 for PiFold, 0.104 for CMLM, 0.144 for LM-Design 650M and 0.000 for DPLM-2 3B, which decodes by argmax. Appendix D.2 states that these four differences are not at matched diversity, and Appendix A states that even for the closest baselines the comparisons are 'close to, but not exactly at, equal diversity', the largest offset of 0.016 being worth at most 0.0011 scTM at the development slope. The scTM and scTM difference results of the protocols named here must not be read as a matched-diversity ranking. The results read at a stated diversity level (Tables 8 and 9 and Appendix D.3) are not covered by this entry, because they are the ones the report reads on a fitted curve at a fixed diversity; nor are the recovery, diversity, pLDDT, scRMSD and lDDT results, whose protocols carry the same single-setting caveat in their own limitations and whose metric qualifiers say the value is at each method's own setting.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix D.1 and D.2; Table 7; Figure 14 diversity block; Section 5.1 'Baselines' and 'HyInv'
The inference-time results were measured with the supervised-stage weights and a rewrite temperature of 0.23, not with the released checkpoint and its default temperatures, and the quality read-outs of the two numerical paths used the RL checkpoint at update 288, not update 640. Speed and quality in this report are therefore measured on different weights, and the 57x figure compares HyInv's optimised BF16 path with ProteinMPNN's released implementation rather than with an equally optimised baseline.HyInv Technical Report (Hyra Team, 2026-10-14) · Table 1 footnote; Appendix F 'Timing protocol' and 'Baseline timings'
Methods, coverage and source HyInv the supervised checkpoint before reinforcement learning at development diversity 0.30 (HyInv report)
hyinv-20261106-protocol-sctm-fixed-diversity-esmfold2
Aggregation: Not reported
HyInv Technical Report (Hyra Team, 2026-10-14) · Section 5.4 'Reinforcement learning', sentence 'RL raises scTM at fixed diversity from 0.839 to 0.852 with plain sampling' (page 7 of 21) Configuration: HyInv plain sampling (update 640) (HyInv report) Protocol: scTM at fixed sequence diversity on the development panel (HyInv report) Dataset: CATH4.3 validation (1,529 chains, development panel) 0.852 Sctm at fixed diversity (scTM at development diversity D = 0.30, read from a least-squares quadratic fit of scTM against design-set diversity over a rewrite-temperature grid, plain single-temperature sampling, ESMFold2, development panel (CATH4.3 validation); the checkpoint is the released update-640 checkpoint)unitless · higher
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons Appendix C states 'We do not claim zero sequence overlap between the training data and the evaluation panels', and Appendix G's homology screen is a coverage rule under which partial hits remain (Table 10: 28 of 74 de novo backbones and 7 of 12 pocket backbones have partial hits, and all 4 named enzymes are homologous to training data by design). Results on these panels are therefore not clean held-out measurements and must not be compared with results from sources that do establish no overlap.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix C 'Data and pool', final bullet; Appendix G 'Homology of the panels' and Table 10
The ten baselines are single published settings at one temperature each, so their design-set diversity differs from HyInv's: on CATH4.3 test HyInv samples at diversity 0.331 against 0.093 for PiFold, 0.104 for CMLM, 0.144 for LM-Design 650M and 0.000 for DPLM-2 3B, which decodes by argmax. Appendix D.2 states that these four differences are not at matched diversity, and Appendix A states that even for the closest baselines the comparisons are 'close to, but not exactly at, equal diversity', the largest offset of 0.016 being worth at most 0.0011 scTM at the development slope. The scTM and scTM difference results of the protocols named here must not be read as a matched-diversity ranking. The results read at a stated diversity level (Tables 8 and 9 and Appendix D.3) are not covered by this entry, because they are the ones the report reads on a fitted curve at a fixed diversity; nor are the recovery, diversity, pLDDT, scRMSD and lDDT results, whose protocols carry the same single-setting caveat in their own limitations and whose metric qualifiers say the value is at each method's own setting.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix D.1 and D.2; Table 7; Figure 14 diversity block; Section 5.1 'Baselines' and 'HyInv'
The inference-time results were measured with the supervised-stage weights and a rewrite temperature of 0.23, not with the released checkpoint and its default temperatures, and the quality read-outs of the two numerical paths used the RL checkpoint at update 288, not update 640. Speed and quality in this report are therefore measured on different weights, and the 57x figure compares HyInv's optimised BF16 path with ProteinMPNN's released implementation rather than with an equally optimised baseline.HyInv Technical Report (Hyra Team, 2026-10-14) · Table 1 footnote; Appendix F 'Timing protocol' and 'Baseline timings'
Methods, coverage and source HyInv the released update-640 checkpoint at development diversity 0.30 (HyInv report)
hyinv-20261106-protocol-sctm-fixed-diversity-esmfold2
Aggregation: Not reported
HyInv Technical Report (Hyra Team, 2026-10-14) · Section 5.4 'Reinforcement learning', sentence 'RL raises scTM at fixed diversity from 0.839 to 0.852 with plain sampling' (page 7 of 21)