, long text divided with
/
3. Lists use
- , quotes use
, code uses
, important sentences can use
4. Preserve all original information, no adding/removing/modifying content
5. No document-level tags like //
The text is a scientific/technical report about text embedders and sentiment decoding. Let me analyze the structure:
The text has these main sections:
1. Opening paragraph about the systematic dissociation finding (sentiment decodable from embeddings but absent from metric)
2. Details about eleven frozen encoders, the probe accuracy vs cosine similarity margin
3. Qwen3-Embedding-4B specifics
4. Impact on retrieval, clustering, deduplication, RAG systems
5. Task-conditioning control explanation
6. Sentiment-Injected Sentence Embeddings (SISE) description
7. Performance improvements
8. Trade-offs (specialization, STS-B losses)
9. Comparison of 0.6B encoder with head vs 4B encoder
10. Release of pipeline and diagnostic
This is a relatively short technical report. Let me think about how to structure it:
- It could be one main section with an intro and a few subsections
- The content is dense, so I should use for key findings/numbers
- There are enumerations that could be lists, but the text is written as prose. I need to be careful - rule 4 says preserve all information without adding/removing/modifying. So I shouldn't convert prose into lists unless the original has list structure.
Wait, let me re-read: "列表内容用- " - list content should use
- . The original text doesn't have explicit lists, so I shouldn't force lists. The text is all prose. Let me structure it as: - An introductory paragraph - Maybe an
for the main findings section - Subsections with
for specific aspects like SISE, trade-offs, etc. Actually, since this is a relatively short report, let me think about the natural flow: Paragraph 1: Opening finding - the dissociation (sentiment decodable from embeddings, absent from metric) Paragraph 2: The eleven encoders study - probe accuracy up to 91.3%, cosine similarity margin 0.003-0.064, Qwen3-Embedding-4B details Paragraph 3: Consequences - retrieval, clustering, deduplication, RAG blind to information Paragraph 4: Task-conditioning control - pretraining objective explanation Paragraph 5: SISE method description Paragraph 6: SISE improvements Paragraph 7: Trade-offs - specialization, STS-B losses Paragraph 8: 0.6B vs 4B comparison Paragraph 9: Release of pipeline and diagnostic Let me think about heading structure. The whole thing is one report. I could do:
The Dissociation
- covering paragraphs 1-4
Sentiment-Injected Sentence Embeddings
- covering paragraphs 5-9 Or maybe:
Systematic Dissociation in Modern Text Embedders
...
Evidence Across Encoders
... Let me draft:
Systematic Dissociation in Modern Text Embedders
We report a systematic dissociation in modern text embedders: sentiment is almost perfectly decodable from their embeddings, yet almost entirely absent from their metric.
Probe Accuracy vs. Metric Margin
Across eleven frozen encoders spanning 2021 to 2026, including 2026-generation models and the 4B-parameter LLM embedder Qwen3-Embedding-4B, a linear probe recovers sentence polarity at up to 91.3% accuracy while cosine similarity separates polarity by a margin of only 0.003-0.064.
Qwen3-Embedding-4B
Qwen3-Embedding-4B, the strongest probe in the panel, reaches 91.3% probe accuracy but only 64.3 retrieval MAP.
Hmm, but this might fragment the text too much. Let me keep it more cohesive. The original is essentially a few dense paragraphs. Let me preserve paragraph structure but add a couple of headings to divide the content. Let me count the natural paragraphs in the original: 1. "We report a systematic dissociation..." (opening) 2. "Across eleven frozen encoders..." (study details) 3. "Qwen3-Embedding-4B, the strongest probe..." (specific model) 4. "Because retrieval, clustering..." (consequences) 5. "A task-conditioning control suggests..." (control experiment) 6. "Sentiment-Injected Sentence Embeddings (SISE) train..." (method) 7. "The head improves every encoder..." (results) 8. "This is a specialization, not a free gain..." (trade-offs) 9. "With the head, a 0.6B encoder..." (comparison) 10. "We release a CPU-reproducible pipeline..." (release) Actually, paragraphs 2 and 3 could be merged as they're about the same study. But rule 4 says don't modify content. Merging paragraphs isn't really modifying content, just formatting. But to be safe, I'll keep the natural paragraph breaks. Actually, let me re-read the original more carefully to identify natural paragraph breaks: "We report a systematic dissociation in modern text embedders: sentiment is almost perfectly decodable from their embeddings, yet almost entirely absent from their metric." - paragraph 1 "Across eleven frozen encoders spanning 2021 to 2026, including 2026-generation models and the 4B-parameter LLM embedder Qwen3-Embedding-4B, a linear probe recovers sentence polarity at up to 91.3% accuracy while cosine similarity separates polarity by a margin of only 0.003-0.064. Qwen3-Embedding-4B, the strongest probe in the panel, reaches 91.3% probe accuracy but only 64.3 retrieval MAP." - paragraph 2 "Because retrieval, clustering, deduplication, and retrieval-augmented generation operate on the metric rather than on a trained probe, these systems are effectively blind to information their encoder demonstrably contains." - paragraph 3 "A task-conditioning control suggests this reflects the pretraining objective rather than a prompting artifact: conditioning Jina-v5 on its classification task rather than its retrieval task recovers only about half of the MAP gain that injection achieves on the same weights (63.2 to 73.8, versus 85.1), and leaves the alignment gap essentially unchanged." - paragraph 4 "Sentiment-Injected Sentence Embeddings (SISE) train a small (0.13M-0.69M-parameter) supervised-contrastive head on a frozen encoder in seconds on a CPU, using the labelled training set." - paragraph 5 "The head improves every encoder released since 2022 on full-data probe accuracy (up to +5.8 points), retrieval MAP (up to +26.0) and polarity margin (7-61x)." - paragraph 6 "This is a specialization, not a free gain: the compact 128-d head loses 38-41 Spearman points on STS-B, while a fused variant that concatenates it with the base embedding lose
, code uses, important sentences can use 4. Preserve all original information, no adding/removing/modifying content 5. No document-level tags like // The text is a scientific/technical report about text embedders and sentiment decoding. Let me analyze the structure: The text has these main sections: 1. Opening paragraph about the systematic dissociation finding (sentiment decodable from embeddings but absent from metric) 2. Details about eleven frozen encoders, the probe accuracy vs cosine similarity margin 3. Qwen3-Embedding-4B specifics 4. Impact on retrieval, clustering, deduplication, RAG systems 5. Task-conditioning control explanation 6. Sentiment-Injected Sentence Embeddings (SISE) description 7. Performance improvements 8. Trade-offs (specialization, STS-B losses) 9. Comparison of 0.6B encoder with head vs 4B encoder 10. Release of pipeline and diagnostic This is a relatively short technical report. Let me think about how to structure it: - It could be one main section with an intro and a few subsections - The content is dense, so I should use for key findings/numbers - There are enumerations that could be lists, but the text is written as prose. I need to be careful - rule 4 says preserve all information without adding/removing/modifying. So I shouldn't convert prose into lists unless the original has list structure. Wait, let me re-read: "列表内容用
- " - list content should use
- . The original text doesn't have explicit lists, so I shouldn't force lists. The text is all prose. Let me structure it as: - An introductory paragraph - Maybe an
for the main findings section - Subsections with
for specific aspects like SISE, trade-offs, etc. Actually, since this is a relatively short report, let me think about the natural flow: Paragraph 1: Opening finding - the dissociation (sentiment decodable from embeddings, absent from metric) Paragraph 2: The eleven encoders study - probe accuracy up to 91.3%, cosine similarity margin 0.003-0.064, Qwen3-Embedding-4B details Paragraph 3: Consequences - retrieval, clustering, deduplication, RAG blind to information Paragraph 4: Task-conditioning control - pretraining objective explanation Paragraph 5: SISE method description Paragraph 6: SISE improvements Paragraph 7: Trade-offs - specialization, STS-B losses Paragraph 8: 0.6B vs 4B comparison Paragraph 9: Release of pipeline and diagnostic Let me think about heading structure. The whole thing is one report. I could do:
The Dissociation
- covering paragraphs 1-4Sentiment-Injected Sentence Embeddings
- covering paragraphs 5-9 Or maybe:Systematic Dissociation in Modern Text Embedders
...
Evidence Across Encoders
... Let me draft:Systematic Dissociation in Modern Text Embedders
We report a systematic dissociation in modern text embedders: sentiment is almost perfectly decodable from their embeddings, yet almost entirely absent from their metric.
Probe Accuracy vs. Metric Margin
Across eleven frozen encoders spanning 2021 to 2026, including 2026-generation models and the 4B-parameter LLM embedder Qwen3-Embedding-4B, a linear probe recovers sentence polarity at up to 91.3% accuracy while cosine similarity separates polarity by a margin of only 0.003-0.064.
Qwen3-Embedding-4B
Qwen3-Embedding-4B, the strongest probe in the panel, reaches 91.3% probe accuracy but only 64.3 retrieval MAP.
Hmm, but this might fragment the text too much. Let me keep it more cohesive. The original is essentially a few dense paragraphs. Let me preserve paragraph structure but add a couple of headings to divide the content. Let me count the natural paragraphs in the original: 1. "We report a systematic dissociation..." (opening) 2. "Across eleven frozen encoders..." (study details) 3. "Qwen3-Embedding-4B, the strongest probe..." (specific model) 4. "Because retrieval, clustering..." (consequences) 5. "A task-conditioning control suggests..." (control experiment) 6. "Sentiment-Injected Sentence Embeddings (SISE) train..." (method) 7. "The head improves every encoder..." (results) 8. "This is a specialization, not a free gain..." (trade-offs) 9. "With the head, a 0.6B encoder..." (comparison) 10. "We release a CPU-reproducible pipeline..." (release) Actually, paragraphs 2 and 3 could be merged as they're about the same study. But rule 4 says don't modify content. Merging paragraphs isn't really modifying content, just formatting. But to be safe, I'll keep the natural paragraph breaks. Actually, let me re-read the original more carefully to identify natural paragraph breaks: "We report a systematic dissociation in modern text embedders: sentiment is almost perfectly decodable from their embeddings, yet almost entirely absent from their metric." - paragraph 1 "Across eleven frozen encoders spanning 2021 to 2026, including 2026-generation models and the 4B-parameter LLM embedder Qwen3-Embedding-4B, a linear probe recovers sentence polarity at up to 91.3% accuracy while cosine similarity separates polarity by a margin of only 0.003-0.064. Qwen3-Embedding-4B, the strongest probe in the panel, reaches 91.3% probe accuracy but only 64.3 retrieval MAP." - paragraph 2 "Because retrieval, clustering, deduplication, and retrieval-augmented generation operate on the metric rather than on a trained probe, these systems are effectively blind to information their encoder demonstrably contains." - paragraph 3 "A task-conditioning control suggests this reflects the pretraining objective rather than a prompting artifact: conditioning Jina-v5 on its classification task rather than its retrieval task recovers only about half of the MAP gain that injection achieves on the same weights (63.2 to 73.8, versus 85.1), and leaves the alignment gap essentially unchanged." - paragraph 4 "Sentiment-Injected Sentence Embeddings (SISE) train a small (0.13M-0.69M-parameter) supervised-contrastive head on a frozen encoder in seconds on a CPU, using the labelled training set." - paragraph 5 "The head improves every encoder released since 2022 on full-data probe accuracy (up to +5.8 points), retrieval MAP (up to +26.0) and polarity margin (7-61x)." - paragraph 6 "This is a specialization, not a free gain: the compact 128-d head loses 38-41 Spearman points on STS-B, while a fused variant that concatenates it with the base embedding lose