# Natural Language Understanding on GLUE (Efficiency Metrics)

Source: [Natural Language Understanding on GLUE (Efficiency Metrics)](<https://www.sota2.com/research/sota/natural-language-understanding-on-glue-efficiency-metrics>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2026-02-26T06:03:31.749Z

Evaluation of model efficiency, specifically measuring speedup and performance drop across the constituent tasks of the GLUE benchmark suite (SST-2, CoLA, MRPC, QNLI, QQP, RTE, MNLI).

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

| Scope | Recorded value |
| --- | --- |
| Task | Natural Language Understanding |
| Dataset (including any stated split) | GLUE |
| Metric | SST-2 Speedup |
| Unit | SCALAR |
| Direction | Higher is better |
| Results loaded | 32 of 32 |
| Coverage | All results currently returned for this benchmark |

Evaluation settings may differ between results. The original papers and recorded configurations are needed to establish comparability. A shared metric name alone does not establish an identical protocol.

## Metrics

- [SST-2 Speedup](<https://www.sota2.com/research/sota/natural-language-understanding-on-glue-efficiency-metrics.md?metric=SST-2%20Speedup>): Higher is better; unit: SCALAR.

- [SST-2 Performance Drop](<https://www.sota2.com/research/sota/natural-language-understanding-on-glue-efficiency-metrics.md?metric=SST-2%20Performance%20Drop>): Lower is better; unit: SCALAR.

- [CoLA Speedup](<https://www.sota2.com/research/sota/natural-language-understanding-on-glue-efficiency-metrics.md?metric=CoLA%20Speedup>): Higher is better; unit: SCALAR.

- [CoLA Performance Drop](<https://www.sota2.com/research/sota/natural-language-understanding-on-glue-efficiency-metrics.md?metric=CoLA%20Performance%20Drop>): Lower is better; unit: SCALAR.

- [MRPC Speedup](<https://www.sota2.com/research/sota/natural-language-understanding-on-glue-efficiency-metrics.md?metric=MRPC%20Speedup>): Higher is better; unit: SCALAR.

- [MRPC Performance Drop](<https://www.sota2.com/research/sota/natural-language-understanding-on-glue-efficiency-metrics.md?metric=MRPC%20Performance%20Drop>): Lower is better; unit: SCALAR.

- [QNLI Speedup](<https://www.sota2.com/research/sota/natural-language-understanding-on-glue-efficiency-metrics.md?metric=QNLI%20Speedup>): Higher is better; unit: SCALAR.

- [QNLI Performance Drop](<https://www.sota2.com/research/sota/natural-language-understanding-on-glue-efficiency-metrics.md?metric=QNLI%20Performance%20Drop>): Lower is better; unit: SCALAR.

- [QQP Speedup](<https://www.sota2.com/research/sota/natural-language-understanding-on-glue-efficiency-metrics.md?metric=QQP%20Speedup>): Higher is better; unit: SCALAR.

- [QQP Performance Drop](<https://www.sota2.com/research/sota/natural-language-understanding-on-glue-efficiency-metrics.md?metric=QQP%20Performance%20Drop>): Lower is better; unit: SCALAR.

- [RTE Speedup](<https://www.sota2.com/research/sota/natural-language-understanding-on-glue-efficiency-metrics.md?metric=RTE%20Speedup>): Higher is better; unit: SCALAR.

- [RTE Performance Drop](<https://www.sota2.com/research/sota/natural-language-understanding-on-glue-efficiency-metrics.md?metric=RTE%20Performance%20Drop>): Lower is better; unit: SCALAR.

- [MNLI Speedup](<https://www.sota2.com/research/sota/natural-language-understanding-on-glue-efficiency-metrics.md?metric=MNLI%20Speedup>): Higher is better; unit: SCALAR.

- [MNLI Performance Drop](<https://www.sota2.com/research/sota/natural-language-understanding-on-glue-efficiency-metrics.md?metric=MNLI%20Performance%20Drop>): Lower is better; unit: SCALAR.

- [MNLI-mm Speedup](<https://www.sota2.com/research/sota/natural-language-understanding-on-glue-efficiency-metrics.md?metric=MNLI-mm%20Speedup>): Higher is better; unit: SCALAR.

- [MNLI-mm Performance Drop](<https://www.sota2.com/research/sota/natural-language-understanding-on-glue-efficiency-metrics.md?metric=MNLI-mm%20Performance%20Drop>): Lower is better; unit: SCALAR.

## Recorded results

| Method | SST-2 Speedup | Result date | Report paper |
| --- | --- | --- | --- |
| DeeBERT-base | 3.06 | Not recorded | 68fa6f741e209131cea45b28 |
| A2T | 3.06 | Not recorded | 68fa6f741e209131cea45b28 |
| HotFlip | 2.77 | Not recorded | 68fa6f741e209131cea45b28 |
| DeeRoBERTa-base | 2.74 | Not recorded | 68fa6f741e209131cea45b28 |
| A2T | 2.7 | Not recorded | 68fa6f741e209131cea45b28 |
| PWWS | 2.44 | Not recorded | 68fa6f741e209131cea45b28 |
| DeeBERT-base | 2.4 | Not recorded | 68fa6f741e209131cea45b28 |
| PWWS | 2.37 | Not recorded | 68fa6f741e209131cea45b28 |
| A2T | 2.35 | Not recorded | 68fa6f741e209131cea45b28 |
| TextFooler | 2.35 | Not recorded | 68fa6f741e209131cea45b28 |
| A2T | 2.34 | Not recorded | 68fa6f741e209131cea45b28 |
| DeeRoBERTa-base | 2.31 | Not recorded | 68fa6f741e209131cea45b28 |
| TextFooler | 2.3 | Not recorded | 68fa6f741e209131cea45b28 |
| TextBugger | 2.29 | Not recorded | 68fa6f741e209131cea45b28 |
| HotFlip | 2.28 | Not recorded | 68fa6f741e209131cea45b28 |
| TextBugger | 2.25 | Not recorded | 68fa6f741e209131cea45b28 |
| HotFlip | 2.21 | Not recorded | 68fa6f741e209131cea45b28 |
| PWWS | 2.13 | Not recorded | 68fa6f741e209131cea45b28 |
| TextFooler | 2.09 | Not recorded | 68fa6f741e209131cea45b28 |
| TextBugger | 2.07 | Not recorded | 68fa6f741e209131cea45b28 |
| PWWS | 1.94 | Not recorded | 68fa6f741e209131cea45b28 |
| TextBugger | 1.83 | Not recorded | 68fa6f741e209131cea45b28 |
| TextFooler | 1.83 | Not recorded | 68fa6f741e209131cea45b28 |
| SAME-Word | 1.79 | Not recorded | 68fa6f741e209131cea45b28 |
| HotFlip | 1.78 | Not recorded | 68fa6f741e209131cea45b28 |
| SAME-Word | 1.63 | Not recorded | 68fa6f741e209131cea45b28 |
| SAME-Char | 1.61 | Not recorded | 68fa6f741e209131cea45b28 |
| SAME-Char | 1.53 | Not recorded | 68fa6f741e209131cea45b28 |
| SAME-Char | 1.21 | Not recorded | 68fa6f741e209131cea45b28 |
| SAME-Word | 1.19 | Not recorded | 68fa6f741e209131cea45b28 |
| SAME-Char | 1.13 | Not recorded | 68fa6f741e209131cea45b28 |
| SAME-Word | 1.1 | Not recorded | 68fa6f741e209131cea45b28 |

### DeeBERT-base

SST-2 Speedup: 3.06. Result ID: 6990d2e63a6c314849039133.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6f741e209131cea45b28.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa676c1e209131cea16147.md>)

Recorded evaluation configuration: {"Backbone":"DeeBERT-base","Performance Drop Threshold":"\< 4%"}. This record does not assert independent reproduction.

### A2T

SST-2 Speedup: 3.06. Result ID: 6990d2e63a6c31484903913d.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6f741e209131cea45b28.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6ab31e209131cea29dcb.md>)

Recorded evaluation configuration: {"Backbone":"DeeBERT-base","Performance Drop Threshold":"\< 4%"}. This record does not assert independent reproduction.

### HotFlip

SST-2 Speedup: 2.77. Result ID: 6990d2e63a6c314849039145.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6f741e209131cea45b28.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa64611e209131cea042e3.md>)

Recorded evaluation configuration: {"Backbone":"DeeRoBERTa-base","Performance Drop Threshold":"\< 4%"}. This record does not assert independent reproduction.

### DeeRoBERTa-base

SST-2 Speedup: 2.74. Result ID: 6990d2e63a6c314849039143.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6f741e209131cea45b28.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa676c1e209131cea16147.md>)

Recorded evaluation configuration: {"Backbone":"DeeRoBERTa-base","Performance Drop Threshold":"\< 4%"}. This record does not assert independent reproduction.

### A2T

SST-2 Speedup: 2.7. Result ID: 6990d2e63a6c31484903914d.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6f741e209131cea45b28.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6ab31e209131cea29dcb.md>)

Recorded evaluation configuration: {"Backbone":"DeeRoBERTa-base","Performance Drop Threshold":"\< 4%"}. This record does not assert independent reproduction.

### PWWS

SST-2 Speedup: 2.44. Result ID: 6990d2e63a6c314849039137.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6f741e209131cea45b28.md>)

Recorded evaluation configuration: {"Backbone":"DeeBERT-base","Performance Drop Threshold":"\< 4%"}. This record does not assert independent reproduction.

### DeeBERT-base

SST-2 Speedup: 2.4. Result ID: 6990d2e63a6c314849039132.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6f741e209131cea45b28.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa676c1e209131cea16147.md>)

Recorded evaluation configuration: {"Backbone":"DeeBERT-base","Performance Drop Threshold":"\< 2%"}. This record does not assert independent reproduction.

### PWWS

SST-2 Speedup: 2.37. Result ID: 6990d2e63a6c314849039147.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6f741e209131cea45b28.md>)

Recorded evaluation configuration: {"Backbone":"DeeRoBERTa-base","Performance Drop Threshold":"\< 4%"}. This record does not assert independent reproduction.

### A2T

SST-2 Speedup: 2.35. Result ID: 6990d2e63a6c31484903913c.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6f741e209131cea45b28.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6ab31e209131cea29dcb.md>)

Recorded evaluation configuration: {"Backbone":"DeeBERT-base","Performance Drop Threshold":"\< 2%"}. This record does not assert independent reproduction.

### TextFooler

SST-2 Speedup: 2.35. Result ID: 6990d2e63a6c31484903914b.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6f741e209131cea45b28.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa67531e209131cea157a0.md>)

Recorded evaluation configuration: {"Backbone":"DeeRoBERTa-base","Performance Drop Threshold":"\< 4%"}. This record does not assert independent reproduction.

### A2T

SST-2 Speedup: 2.34. Result ID: 6990d2e63a6c31484903914c.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6f741e209131cea45b28.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6ab31e209131cea29dcb.md>)

Recorded evaluation configuration: {"Backbone":"DeeRoBERTa-base","Performance Drop Threshold":"\< 2%"}. This record does not assert independent reproduction.

### DeeRoBERTa-base

SST-2 Speedup: 2.31. Result ID: 6990d2e63a6c314849039142.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6f741e209131cea45b28.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa676c1e209131cea16147.md>)

Recorded evaluation configuration: {"Backbone":"DeeRoBERTa-base","Performance Drop Threshold":"\< 2%"}. This record does not assert independent reproduction.

### TextFooler

SST-2 Speedup: 2.3. Result ID: 6990d2e63a6c31484903913b.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6f741e209131cea45b28.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa67531e209131cea157a0.md>)

Recorded evaluation configuration: {"Backbone":"DeeBERT-base","Performance Drop Threshold":"\< 4%"}. This record does not assert independent reproduction.

### TextBugger

SST-2 Speedup: 2.29. Result ID: 6990d2e63a6c314849039149.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6f741e209131cea45b28.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa65351e209131cea08f5c.md>)

Recorded evaluation configuration: {"Backbone":"DeeRoBERTa-base","Performance Drop Threshold":"\< 4%"}. This record does not assert independent reproduction.

### HotFlip

SST-2 Speedup: 2.28. Result ID: 6990d2e63a6c314849039144.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6f741e209131cea45b28.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa64611e209131cea042e3.md>)

Recorded evaluation configuration: {"Backbone":"DeeRoBERTa-base","Performance Drop Threshold":"\< 2%"}. This record does not assert independent reproduction.

### TextBugger

SST-2 Speedup: 2.25. Result ID: 6990d2e63a6c314849039139.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6f741e209131cea45b28.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa65351e209131cea08f5c.md>)

Recorded evaluation configuration: {"Backbone":"DeeBERT-base","Performance Drop Threshold":"\< 4%"}. This record does not assert independent reproduction.

### HotFlip

SST-2 Speedup: 2.21. Result ID: 6990d2e63a6c314849039135.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6f741e209131cea45b28.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa64611e209131cea042e3.md>)

Recorded evaluation configuration: {"Backbone":"DeeBERT-base","Performance Drop Threshold":"\< 4%"}. This record does not assert independent reproduction.

### PWWS

SST-2 Speedup: 2.13. Result ID: 6990d2e63a6c314849039146.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6f741e209131cea45b28.md>)

Recorded evaluation configuration: {"Backbone":"DeeRoBERTa-base","Performance Drop Threshold":"\< 2%"}. This record does not assert independent reproduction.

### TextFooler

SST-2 Speedup: 2.09. Result ID: 6990d2e63a6c31484903914a.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6f741e209131cea45b28.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa67531e209131cea157a0.md>)

Recorded evaluation configuration: {"Backbone":"DeeRoBERTa-base","Performance Drop Threshold":"\< 2%"}. This record does not assert independent reproduction.

### TextBugger

SST-2 Speedup: 2.07. Result ID: 6990d2e63a6c314849039148.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6f741e209131cea45b28.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa65351e209131cea08f5c.md>)

Recorded evaluation configuration: {"Backbone":"DeeRoBERTa-base","Performance Drop Threshold":"\< 2%"}. This record does not assert independent reproduction.

### PWWS

SST-2 Speedup: 1.94. Result ID: 6990d2e63a6c314849039136.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6f741e209131cea45b28.md>)

Recorded evaluation configuration: {"Backbone":"DeeBERT-base","Performance Drop Threshold":"\< 2%"}. This record does not assert independent reproduction.

### TextBugger

SST-2 Speedup: 1.83. Result ID: 6990d2e63a6c314849039138.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6f741e209131cea45b28.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa65351e209131cea08f5c.md>)

Recorded evaluation configuration: {"Backbone":"DeeBERT-base","Performance Drop Threshold":"\< 2%"}. This record does not assert independent reproduction.

### TextFooler

SST-2 Speedup: 1.83. Result ID: 6990d2e63a6c31484903913a.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6f741e209131cea45b28.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa67531e209131cea157a0.md>)

Recorded evaluation configuration: {"Backbone":"DeeBERT-base","Performance Drop Threshold":"\< 2%"}. This record does not assert independent reproduction.

### SAME-Word

SST-2 Speedup: 1.79. Result ID: 6990d2e63a6c31484903914f.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6f741e209131cea45b28.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6f741e209131cea45b28.md>)

[Implementation](<https://github.com/MatthewCYM/SAME>)

Recorded evaluation configuration: {"Backbone":"DeeRoBERTa-base","Performance Drop Threshold":"\< 4%"}. This record does not assert independent reproduction.

### HotFlip

SST-2 Speedup: 1.78. Result ID: 6990d2e63a6c314849039134.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6f741e209131cea45b28.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa64611e209131cea042e3.md>)

Recorded evaluation configuration: {"Backbone":"DeeBERT-base","Performance Drop Threshold":"\< 2%"}. This record does not assert independent reproduction.

### SAME-Word

SST-2 Speedup: 1.63. Result ID: 6990d2e63a6c31484903914e.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6f741e209131cea45b28.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6f741e209131cea45b28.md>)

[Implementation](<https://github.com/MatthewCYM/SAME>)

Recorded evaluation configuration: {"Backbone":"DeeRoBERTa-base","Performance Drop Threshold":"\< 2%"}. This record does not assert independent reproduction.

### SAME-Char

SST-2 Speedup: 1.61. Result ID: 6990d2e63a6c314849039151.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6f741e209131cea45b28.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6f741e209131cea45b28.md>)

[Implementation](<https://github.com/MatthewCYM/SAME>)

Recorded evaluation configuration: {"Backbone":"DeeRoBERTa-base","Performance Drop Threshold":"\< 4%"}. This record does not assert independent reproduction.

### SAME-Char

SST-2 Speedup: 1.53. Result ID: 6990d2e63a6c314849039150.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6f741e209131cea45b28.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6f741e209131cea45b28.md>)

[Implementation](<https://github.com/MatthewCYM/SAME>)

Recorded evaluation configuration: {"Backbone":"DeeRoBERTa-base","Performance Drop Threshold":"\< 2%"}. This record does not assert independent reproduction.

### SAME-Char

SST-2 Speedup: 1.21. Result ID: 6990d2e63a6c314849039141.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6f741e209131cea45b28.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6f741e209131cea45b28.md>)

[Implementation](<https://github.com/MatthewCYM/SAME>)

Recorded evaluation configuration: {"Backbone":"DeeBERT-base","Performance Drop Threshold":"\< 4%"}. This record does not assert independent reproduction.

### SAME-Word

SST-2 Speedup: 1.19. Result ID: 6990d2e63a6c31484903913f.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6f741e209131cea45b28.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6f741e209131cea45b28.md>)

[Implementation](<https://github.com/MatthewCYM/SAME>)

Recorded evaluation configuration: {"Backbone":"DeeBERT-base","Performance Drop Threshold":"\< 4%"}. This record does not assert independent reproduction.

### SAME-Char

SST-2 Speedup: 1.13. Result ID: 6990d2e63a6c314849039140.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6f741e209131cea45b28.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6f741e209131cea45b28.md>)

[Implementation](<https://github.com/MatthewCYM/SAME>)

Recorded evaluation configuration: {"Backbone":"DeeBERT-base","Performance Drop Threshold":"\< 2%"}. This record does not assert independent reproduction.

### SAME-Word

SST-2 Speedup: 1.1. Result ID: 6990d2e63a6c31484903913e.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6f741e209131cea45b28.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6f741e209131cea45b28.md>)

[Implementation](<https://github.com/MatthewCYM/SAME>)

Recorded evaluation configuration: {"Backbone":"DeeBERT-base","Performance Drop Threshold":"\< 2%"}. This record does not assert independent reproduction.

Page 1 of 1. This is an index page, not the full collection.
