Research map / Architectures and efficiency / Attention and transformers
Language Models are Injective and Hence Invertible
Giorgos Nikolaou, Tommaso Mencattini, Donato Crisostomi, Andrea Santilli, Yannis Panagakis, Emanuele Rodolà
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 |
|---|---|---|---|
| use FP4 weight quantization instead of FP32 | Llama-3.1-8B · Collision search, final layer | min pairwise l2 distance: 1.274 -> 2.281 | Improved |
| use INT8 weight quantization instead of FP32 | Llama-3.1-8B · Collision search, final layer | min pairwise l2 distance: 1.274 -> 6.597 | Improved |
| use FP4 weight quantization instead of FP32 | Mistral-7B-v0.1 · Collision search, final layer | min pairwise l2 distance: 1.136 -> 1.748 | Improved |
| use INT8 weight quantization instead of FP32 | Phi-4-mini-instruct · Collision search, final layer | min pairwise l2 distance: 8.780 -> 20.956 | Improved |
| go from first layer to last layer | Llama-3.1-8B · Collision search | min pairwise l2 distance: 0.001 -> 0.620 | Improved |
| go from first layer to last layer | Llama-3.1-70B · Collision search | min pairwise l2 distance: 0.005 -> 3.975 | Improved |
| use SipIt instead of BruteForce | prompt inversion · Prompt inversion, 20 tokens | mean time (s): 3889.61 -> 28.01 | Improved |
| use SipIt instead of HardPrompts | prompt inversion · Prompt inversion, 20 tokens | accuracy: 0.00 -> 1.00 | Improved |
| increase vocabulary from 32000 to 128255 | SipIt with FP4 models · SipIt inversion, FP4 | vocab explored (%): 0.19 ± 0.08 % -> 0.21 ± 0.10 % | No clear effect |
| increase prompt length (GPT-2 Small) | last-token states · Collision search | min pairwise l2 distance: rises at short lengths then stabilizes | Mixed |
| deeper layer for inversion | SipIt · Inversion time, prompts 20-200 tokens | time (s): runtimes rise only slightly | No clear effect |
Closest papers by meaning
- Distilling Sequential Computation in Transformer Language Models (NeurIPS 2026 · similarity 0.89)
- Bigram Subnetworks: Mapping to Next Tokens in Transformer Language Models (NeurIPS 2025 · similarity 0.89)
- Characterizing the Expressivity of Fixed-Precision Transformer Language Models (NeurIPS 2025 · similarity 0.88)
- Transformer Block Coupling and its Correlation with Generalization in LLMs (ICLR 2025 · similarity 0.88)
- Transformers need glasses! Information over-squashing in language tasks (NeurIPS 2024 · similarity 0.87)
- Beyond Components: Singular Vector-Based Interpretability of Transformer Circuits (NeurIPS 2025 · similarity 0.87)
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.
