Research map / Learning theory and algorithms / Combinatorial and search optimization
A Pseudo-Semantic Loss for Autoregressive Models with Logical Constraints
Kareem Ahmed, Kai-Wei Chang, Guy Van den Broeck
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add pseudo-semantic loss to CNN-LSTM | CNN-LSTM · Warcraft | exact accuracy (%): 62.00 -> 66.00 | Improved |
| add pseudo-semantic loss to CNN-LSTM | CNN-LSTM · Warcraft | consistent accuracy (%): 76.60 -> 79.00 | Improved |
| use autoregressive LSTM instead of ResNet-18 | ResNet-18 · Warcraft | exact accuracy (%): 55.00 -> 62.00 | Improved |
| add semantic loss to ResNet-18 | ResNet-18 · Warcraft | exact accuracy (%): 55.00 -> 59.40 | Improved |
| add pseudo-semantic loss to RNN | RNN · Sudoku | exact accuracy (%): 22.40 -> 28.20 | Improved |
| add semantic loss to ConvNet | ConvNet · Sudoku | exact accuracy (%): 16.80 -> 22.10 | Improved |
| use autoregressive RNN instead of ConvNet with semantic loss | ConvNet + SL · Sudoku | exact accuracy (%): 22.10 -> 22.40 | No clear effect |
| add pseudo-semantic loss to SGEAT | SGEAT · RealToxicityPrompts full | expected maximum toxicity: 0.32 -> 0.29 | Improved |
| add pseudo-semantic loss to SGEAT | SGEAT · RealToxicityPrompts full | toxicity probability (%): 14.05 -> 9.80 | Improved |
| add pseudo-semantic loss to SGEAT | SGEAT · WebText validation | perplexity: 28.72 -> 28.14 | Improved |
| add Word Banning at decoding to PseudoSL (ablation) | PseudoSL · RealToxicityPrompts full | toxicity probability (%): 9.80 -> 9.20 | Improved |
| add pseudo-semantic loss to SGEAT with Word Banning | SGEAT + Word Banning · RealToxicityPrompts nontoxic | toxicity probability (%): 6.17 -> 6.55 | Hurt |
Closest papers by meaning
- Controllable Generation via Locally Constrained Resampling (ICLR 2025 · similarity 0.88)
- Constraints-Guided Diffusion Reasoner for Neuro-Symbolic Learning (AAAI 2026 · similarity 0.87)
- Zero-Shot Conditioning of Score-Based Diffusion Models by Neuro-Symbolic Constraints (AAAI 2025 · similarity 0.86)
- Logically Consistent Language Models via Neuro-Symbolic Integration (ICLR 2025 · similarity 0.86)
- Autoregressive Language Models are Secretly Energy-Based Models: Insights into the Lookahead Capabilities of Next-Token Prediction (ICML 2026 · similarity 0.86)
- Relational Neurosymbolic Markov Models (AAAI 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
