Conclusion Generation on EntailmentBank (test)
67BLEUT5 original (large)
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| T5 original (large)Architecture=seq2seqLM: encoder-decoder, Injection Method (INJ)=EP: encoder prefix2026.01 | 67 | 0.97 | 0.5 | |
| T5 original (small)Architecture=seq2seqLM: encoder-decoder, Injection Method (INJ)=EP: encoder prefix2026.01 | 65 | 0.97 | 0.45 | |
| Qwen2.5(0.5B)Architecture=CausalLM: decoder only, Injection Method (INJ)=DP: decoder prefix2026.01 | 65 | 0.97 | 0.48 | |
| T5 original (large)Architecture=seq2seqLM: encoder-decoder, Injection Method (INJ)=DP: decoder prefix2026.01 | 64 | 0.97 | 0.44 | |
| Qwen2.5(0.5B)Architecture=CausalLM: decoder only, Injection Method (INJ)=NO: no inference type2026.01 | 63 | 0.97 | 0.45 | |
| Llama3.2(1B)Architecture=CausalLM: decoder only, Injection Method (INJ)=DP: decoder prefix2026.01 | 63 | 0.97 | 0.46 | |
| T5 original (base)Architecture=seq2seqLM: encoder-decoder, Injection Method (INJ)=EP: encoder prefix2026.01 | 61 | 0.97 | 0.39 | |
| T5 original (large)Architecture=seq2seqLM: encoder-decoder, Injection Method (INJ)=DE: decoder end2026.01 | 60 | 0.97 | 0.46 | |
| Llama3.2(1B)Architecture=CausalLM: decoder only, Injection Method (INJ)=NO: no inference type2026.01 | 60 | 0.96 | 0.42 | |
| T5 original (small)Architecture=seq2seqLM: encoder-decoder, Injection Method (INJ)=DP: decoder prefix2026.01 | 59 | 0.96 | 0.34 | |
| T5 original (base)Architecture=seq2seqLM: encoder-decoder, Injection Method (INJ)=NO: no inference type2026.01 | 57 | 0.96 | 0.33 | |
| T5 original (large)Architecture=seq2seqLM: encoder-decoder, Injection Method (INJ)=NO: no inference type2026.01 | 57 | 0.96 | 0.31 | |
| T5 original (small)Architecture=seq2seqLM: encoder-decoder, Injection Method (INJ)=DE: decoder end2026.01 | 55 | 0.96 | 0.3 | |
| T5 original (small)Architecture=seq2seqLM: encoder-decoder, Injection Method (INJ)=NO: no inference type2026.01 | 54 | 0.96 | 0.22 | |
| T5 original (base)Architecture=seq2seqLM: encoder-decoder, Injection Method (INJ)=DP: decoder prefix2026.01 | 53 | 0.96 | 0.25 | |
| T5 original (base)Architecture=seq2seqLM: encoder-decoder, Injection Method (INJ)=DE: decoder end2026.01 | 46 | 0.96 | 0.23 | |
| T5 bottleneck (base)Architecture=seq2seqLM with sentence bottleneck, Injection Method (INJ)=EP: encoder prefix2026.01 | 42 | 0.92 | -0.07 | |
| T5 bottleneck (base)Architecture=seq2seqLM with sentence bottleneck, Injection Method (INJ)=DP: decoder prefix2026.01 | 39 | 0.91 | -0.13 | |
| T5 bottleneck (base)Architecture=seq2seqLM with sentence bottleneck, Injection Method (INJ)=DE: decoder end2026.01 | 35 | 0.91 | -0.15 | |
| T5 bottleneck (base)Architecture=seq2seqLM with sentence bottleneck, Injection Method (INJ)=NO: no inference type2026.01 | 35 | 0.91 | -0.2 | |
| GPT2(xl)Architecture=CausalLM: decoder only, Injection Method (INJ)=DP: decoder prefix2026.01 | 28 | 0.91 | -0.9 | |
| GPT2(xl)Architecture=CausalLM: decoder only, Injection Method (INJ)=NO: no inference type2026.01 | 27 | 0.9 | -0.97 | |
| Optimus (BERT-GPT2)Architecture=seq2seqLM with sentence bottleneck, Injection Method (INJ)=DE: decoder end2026.01 | 26 | 0.8 | -1.11 | |
| Optimus (BERT-GPT2)Architecture=seq2seqLM with sentence bottleneck, Injection Method (INJ)=DP: decoder prefix2026.01 | 25 | 0.79 | -1.14 | |
| Optimus (BERT-GPT2)Architecture=seq2seqLM with sentence bottleneck, Injection Method (INJ)=EP: encoder prefix2026.01 | 9 | 0.74 | -1.17 | |
| Optimus (BERT-GPT2)Architecture=seq2seqLM with sentence bottleneck, Injection Method (INJ)=NO: no inference type2026.01 | 7 | 0.74 | -1.2 |