Pith. sign in

REVIEW 4 major objections 5 minor 7 references

Text embeddings that encode a language model's potential response, rather than the input query, achieve state-of-the-art self-supervised results and carry the model's safety and reasoning into retrieval.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 18:23 UTC pith:AYKM6UAT

load-bearing objection Strong self-supervised embedding recipe with real empirical gains; the 'output space' claim is partly overstated but the method itself is novel and worth taking seriously. the 4 major comments →

arxiv 2603.10913 v3 pith:AYKM6UAT submitted 2026-03-11 cs.CL

LLM2Vec-Gen: Generative Embeddings from Large Language Models

Parishad BehnamGhader , Vaibhav Adlakha , Fabian David Schmidt , Nicolas Chapados , Marius Mosbach , Siva Reddy This is my paper
classification cs.CL
keywords text embeddingsoutput-centric representationsself-supervised learningcompression tokensembedding distillationLLM safety transferreasoning-intensive retrievalMTEB
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

LLM2Vec-Gen sets out to fix a blind spot in how language models are turned into text embedders: contrastive fine-tuning pushes inputs and outputs into a new space, discarding the semantic structure the model learned during pretraining. The paper's proposal is to embed the LLM's potential response instead of the input. Trainable special tokens are appended to each query and trained to compress the model's own generated response into a fixed-length vector, using two objectives: align the vector with an unsupervised teacher's embedding of the response, and reconstruct the response from the vector with the frozen LLM. Training needs only unlabeled queries. The result is a new state of the art for self-supervised embeddings on MTEB (up to 61.9, 8.8% over the teacher), with safety and reasoning abilities carried from the LLM into the embedding: harmful-content retrieval drops and reasoning-intensive retrieval improves. Because the backbone stays frozen, the same weights serve both embedding and generation, and the embedding can be decoded back into text to show what it captured.

Core claim

The central claim is that an embedding trained to encode what a language model would say—not what the query says—preserves capabilities that contrastive reshaping throws away. The concrete mechanism: append ten trainable compression tokens to a query, run the frozen LLM, and train those tokens to (1) match the unsupervised LLM2Vec teacher's embedding of the LLM's own generated response, and (2) let the frozen LLM reconstruct that response from the tokens via next-token prediction. On MTEB(eng, v2) this reaches 61.9 for Qwen-3-8B versus 56.8 for the teacher, state of the art among self-supervised methods and closing over 60% of the gap to supervised training. The same embeddings reduce harmfu

What carries the argument

The engine is a set of n=10 trainable compression tokens c1..cn appended to the query. Their hidden states are passed through two lightweight MLPs and mean-pooled to form the embedding. Two losses pull in complementary directions: alignment (Lalign) minimizes squared distance between that embedding and an unsupervised LLM2Vec teacher's embedding of the LLM's own generated response; reconstruction (Lrecon) feeds a projection of the same hidden states as soft prompts and asks the frozen LLM to regenerate the response by next-token prediction. Only the tokens and MLPs are trained (~13M parameters for a 4B backbone), so the LLM's representational geometry is preserved; the reconstruction term ac

Load-bearing premise

The load-bearing premise is that the unsupervised teacher's embedding space faithfully preserves the LLM's output semantics, so aligning to the teacher is equivalent to encoding the response; if the teacher is relevance-biased or poor for some response types, the student inherits that distortion.

What would settle it

Decode the learned embeddings for a set of 100 harmful queries and inspect the text: if the decoded text does not consist of refusals (e.g., 'I cannot assist with that') yet retrieval is still safe, then safety does not come from encoding the response. A complementary quantitative check: train the same model with the alignment target replaced by a random fixed vector; if MTEB performance does not collapse, the teacher alignment is not the cause of the gains.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Self-supervised text embeddings reach a new state of the art on MTEB(eng, v2) across Llama-3.x, Qwen-2.5, and Qwen-3 families, closing more than 60% of the gap to supervised methods.
  • Embeddings inherit the LLM's safety alignment: on AdvBench-IR, top-5 accuracy for harmful queries drops by up to 22.6% because the embedding encodes the refusal, not the malicious intent.
  • Reasoning transfers to retrieval: on BRIGHT, nDCG@10 improves up to 35.6%, with larger backbones showing larger gains.
  • Because the backbone is frozen, one set of weights serves both generation and embedding, and inference needs only a single forward pass.
  • Embeddings are interpretable by construction: the reconstruction objective lets them be decoded back into natural language, and Logit Lens / LatentLens analyses show response-level semantics.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • This invites a testable program: if output-centric encoding is the active ingredient, the largest gains should concentrate in tasks where many inputs map to one output—clustering, classification, STS—which is exactly where the paper reports gains; a reader could push this further by evaluating on more diverse MTEB categories, especially multilingual ones.
  • The teacher is the ceiling: since the student distills an unsupervised teacher, replacing the teacher with a stronger unsupervised encoder that better preserves the LLM's geometry (or a full JEPA variant where the LLM itself provides targets, which the paper lists as an open frontier) should raise the ceiling further.
  • The compression tokens can be chained—feed the embedding back as input with fresh tokens to represent the response to the response—giving a path to multi-step reasoning in a few forward passes rather than autoregressive decoding, a possibility the paper raises.
  • Safety transfer suggests a general recipe for aligning retrievers without safety-specific labels: if the LLM refuses, the embedding encodes the refusal; this may generalize to other harm categories or languages, though the paper's evidence is limited to AdvBench-IR queries in English.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes LLM2Vec-Gen, a self-supervised method that converts a frozen decoder-only LLM into a text embedder by appending trainable compression tokens to the input and optimizing them to (i) match, in L2, the embedding produced by an unsupervised LLM2Vec teacher for the LLM's own generated response, and (ii) allow the frozen LLM to reconstruct that response from the compression-token representations. The authors claim this places embeddings 'directly in the LLM's output space' and thereby preserves response-level capabilities. Experiments on MTEB(eng,v2), AdvBench-IR, and BRIGHT across Llama and Qwen families show consistent gains over the LLM2Vec teacher, up to 61.9 average MTEB score for Qwen-3-8B, reduced harmful-content retrieval, and improved reasoning-intensive retrieval. Ablations isolate the alignment loss, the number of compression tokens, the response generator, the embedding teacher, and frozen-vs-LoRA training.

Significance. If the central claim is correct, this is an important conceptual shift: embeddings that encode the model's likely response rather than the input, obtained without labeled pairs and with a frozen backbone. The paper ships code, data, and models, evaluates on external benchmarks (MTEB, BRIGHT, AdvBench-IR), and includes careful ablations; those are real strengths. The main reservation is conceptual: because the alignment target comes from a SimCSE-trained encoder, the claimed 'output-space' semantics and the inheritance of safety/reasoning are not directly established. The evaluation also lacks uncertainty quantification, and the safety interpretation has a potential confound. These are fixable within the scope of a revision.

major comments (4)
  1. [Section 3, Eq. (1); Appendices A and I] The load-bearing claim that embeddings live in the LLM's output space is not established. The only target geometry is that of an unsupervised LLM2Vec teacher trained with SimCSE, which is itself an input-centric contrastive encoder. The paper's own argument in Appendix I—that contrastive objectives reshape the representation space and therefore make supervised teachers unfaithful—applies, in weaker form, to SimCSE's uniformity regularization. Section 3 asserts that SimCSE 'largely preserves' local geometry, but no measurement is provided, and Appendix A concedes the student is bounded by the teacher's capacity. A concrete test would be to compare the teacher's response embeddings with a direct mean-pooled representation of the response from the frozen LLM (nearest-neighbor agreement, or training the same student with that target) and report agreement; without such evidence, Section 4.5's
  2. [Tables 1–3, Figure 3] No uncertainty quantification is reported. All numbers appear to be single-run values. Claims like 'up to 8.8% improvement' (Table 1), a 1.1-point gain for Llama-3.1-8B (Figure 3), and ablation differences of 0.3–0.5 points (Table 3: 67.9 vs 67.5; 67.4 vs 67.3) are used to draw conclusions. Please report means and standard deviations over at least three seeds, or bootstrap confidence intervals, for the main comparisons and ablations; at minimum, state that the results are single runs and identify which differences exceed noise.
  3. [Section 4.5, Table 2] The AdvBench-IR result is interpreted as inherited safety, but only top-5 accuracy on the harmful corpus is reported. A lower score could also arise from a general degradation in retrieval effectiveness. To support the safety claim, report retrieval performance on a benign query/document set of matched difficulty under the same instruction, or show that harmful retrievals are replaced by benign retrievals (e.g., precision@5 breakdown by passage source). Without this control, the 22.6% reduction is ambiguous.
  4. [Section 6 vs Appendix F, Table 11] Interpretability/decodability is overclaimed. Table 4 shows selected successful decodings, but Table 11 shows that LLM2Vec-GEN-Qwen-3-4B decodes a question about the University of Kansas fight song into a confident answer about the University of Florida mascot, and a Snatch question into Doctor Who. Thus the reconstruction loss ensures fluent generation, not that the embedding reveals the response's semantic content. Either quantify decodability (e.g., factuality or lexical overlap on a sample) or soften the claim in the abstract and Section 6.
minor comments (5)
  1. [Figure 3] The legend/axis uses 'LLM2Vec-Gen', inconsistent with the method name 'LLM2VEC-GEN' used elsewhere.
  2. [Table 5] Model identifiers are inconsistent (e.g., 'Qwen3' vs 'Qwen-3'). Please unify for reproducibility.
  3. [Introduction / Table 2] The introduction says '35.6% improvement over the input-centric baseline'; Table 2's percentage is relative to the LLM2Vec teacher for the same backbone. Clarify the baseline reference.
  4. [Appendix I] Gemini-2.5-flash is used as a hard-negative generator but is not listed in Table 5; add the exact model identifier and access date.
  5. [Section 4.2, Tables 7–8] For AdvBench-IR and BRIGHT, provide the embedding-vs-generative instruction comparison analogous to Figure 5 to rule out instruction-wording effects.

Circularity Check

0 steps flagged

No significant circularity: the training objectives and external benchmarks are distinct, and the reported improvements are measured rather than entailed by the loss.

full rationale

The central derivation chain is not circular. LLM2Vec-Gen optimizes L = L_align + L_recon (Section 3), where L_align regresses pooled compression-token embeddings to an unsupervised teacher's embedding of the LLM's own generated response and L_recon trains the frozen LLM to reconstruct that response from the compression tokens. The headline claims—MTEB, BRIGHT, and AdvBench-IR results—are computed on fixed external benchmarks and are not equal to either term of the training loss. The 8.8% MTEB improvement over the LLM2Vec teacher is not forced by construction: at evaluation the student encodes both queries and documents with its own learned compression tokens, while the training only saw unlabeled queries, so the gain is an empirical transfer result rather than a restatement of L_align. The paper does rely on the authors' own unsupervised LLM2Vec model as the alignment teacher, and AdvBench-IR shares authors, but this is a provenance/independence concern rather than a definitional reduction: the method is also evaluated against the teacher on held-out tasks, and Appendix G shows the approach generalizes to a BGE teacher. Appendix A explicitly concedes that the student is bounded by the teacher's representational capacity, and the claim that the SimCSE teacher is 'faithful' is an unverified assumption; these are correctness risks, not circular steps. The decodability/interpretability result is partly by design because L_recon trains the embeddings to be decodable, but the paper supplements this with qualitative Logit Lens and LatentLens evidence, and interpretability is not the load-bearing claim of the paper. Overall, no exhibited equation reduces a reported prediction to a fitted input or to a self-citation chain, so the correct circularity verdict is low.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 1 invented entities

The method relies on an external teacher (LLM2Vec) whose assumptions about faithfulness are load-bearing, and on the LLM's own generations as training targets. The architecture introduces no new physical entities but does introduce compression tokens as a learned mechanism. The central empirical claims are not circular because they are evaluated on external benchmarks, though those benchmarks include a self-authored safety test.

free parameters (4)
  • Number of compression tokens = 10
    Chosen by hand and validated via ablation (Figure 4); performance varies between 66.1 and 68.5 on MTEB-Lite depending on token count, so the exact number affects results.
  • Sequence length truncation = 512 tokens
    Queries and responses longer than 512 tokens are truncated, which can discard information that the reconstruction and alignment losses need, influencing embedding quality.
  • Loss weighting = 1.0 for both Lalign and Lrecon
    No weighting scheme is reported; equal weighting is assumed. Changing the balance between teacher alignment and reconstruction could shift the interpretability/performance trade-off.
  • Training data and epoch count = 160K Tulu queries, 1 epoch
    Performance is reported for a single epoch over 160K unlabeled queries; the method's robustness to dataset scale or choice is not explored.
axioms (4)
  • domain assumption Unsupervised LLM2Vec teacher embeddings faithfully represent the semantic content of generated responses and preserve the LLM's local representational geometry.
    Section 3 claims the teacher's SimCSE objective 'largely preserving the LLM's local representational geometry'; if false, the student inherits teacher biases rather than output semantics.
  • domain assumption The frozen LLM generates high-quality, diverse responses for the training queries that are representative of its behavior in downstream tasks.
    Training targets are the LLM's own generations (Section 4.1); degenerate or unrepresentative generations would degrade the alignment and reconstruction signals.
  • domain assumption MTEB(eng,v2), AdvBench-IR, and BRIGHT are valid proxies for the claimed capabilities (embedding quality, safety, reasoning).
    The paper's core conclusions are benchmarked on these tasks; AdvBench-IR is co-authored by the present authors, introducing a potential selection-bias concern.
  • domain assumption The reconstruction objective (cross-entropy on the response) is sufficient to make the learned embeddings decodable and content-preserving.
    Interpretability is demonstrated only qualitatively (Table 4, 11, 13); no quantitative evaluation establishes how reliably the decoded text matches the intended response.
invented entities (1)
  • Compression tokens c1...cn no independent evidence
    purpose: Learnable special tokens appended to input queries; their hidden states are projected to produce the final embedding.
    These are learned parameters, not a physically or empirically validated entity outside the method; they are justified only by downstream benchmark performance and qualitative decoding examples.

pith-pipeline@v1.3.0-alltime-deepseek · 26239 in / 10710 out tokens · 102450 ms · 2026-08-02T18:23:14.539847+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of LLM2Vec-Gen: Generative Embeddings from Large Language Models." pith.science (2026). https://pith.science/paper/AYKM6UAT

@misc{pith2026260310913,
  author       = {Pith},
  title        = {Pith review of: LLM2Vec-Gen: Generative Embeddings from Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AYKM6UAT}},
  note         = {Machine review of arXiv:2603.10913}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Fine-tuning LLM-based text embedders via contrastive learning maps inputs and outputs into a new representational space, discarding the LLM's output semantics. We propose LLM2Vec-Gen, a self-supervised alternative that instead produces embeddings directly in the LLM's output space by learning to represent the model's potential response. Specifically, trainable special tokens are appended to the input and optimized to compress the LLM's own response into a fixed-length embedding, guided by an unsupervised embedding teacher and a reconstruction objective. Crucially, the LLM backbone remains frozen and training requires only unlabeled queries. LLM2Vec-Gen achieves state-of-the-art self-supervised performance on the Massive Text Embedding Benchmark (MTEB), improving by 8.8% over the unsupervised embedding teacher. Since the embeddings preserve the LLM's response-space semantics, they inherit capabilities such as safety alignment (up to 22.6% reduction in harmful content retrieval) and reasoning (up to 35.6% improvement on reasoning-intensive retrieval). Finally, the learned embeddings are also interpretable: they can be decoded back into text to reveal their semantic content. Our code, data, and pretrained models are publicly available at https://github.com/McGill-NLP/llm2vec-gen.

Figures

Figures reproduced from arXiv: 2603.10913 by Fabian David Schmidt, Marius Mosbach, Nicolas Chapados, Parishad BehnamGhader, Siva Reddy, Vaibhav Adlakha.

Figure 1
Figure 1. Figure 1: Illustration of the difference between input- and output-centric representations for [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of LLM2VEC-GEN. Left: Given unlabeled queries, the LLM generates responses, embedded by an unsupervised teacher. Right: Trainable compression tokens are appended to queries. Keeping the LLM backbone frozen, the compression tokens’ hidden states are optimized via alignment loss Lalign (match the teacher’s response embedding) and reconstruction loss Lrecon (reconstruct the response from soft prompts… view at source ↗
Figure 3
Figure 3. Figure 3: MTEB (eng, v2) average score based on model size across three model families. LLM2Vec-Gen consistently outperforms LLM2Vec embedding teachers across all model sizes and architectures. LLM2VEC-GEN achieves state-of-the-art self-supervised performance on MTEB. LLM2VEC-GEN outperforms all baselines across the three Qwen-3 scales (1.7B, 4B, and 8B). LLM2VEC-GEN with Qwen-3-8B establishes a new self-supervised … view at source ↗
Figure 4
Figure 4. Figure 4: Impact of special tokens. Importance of response generator. We find that us￾ing responses from other sources does not improve performance: Tulu responses yield 67.3, a stronger in-family model (Qwen-3-8B) yields 67.4, and an out￾of-family model (Gemini-3-flash) yields 67.1. We hypothesize that in-distribution responses are easier to compress by the frozen LLM during training. Importance of embedding teache… view at source ↗
Figure 5
Figure 5. Figure 5: Impact of generative instructions on the performance of LLM2V [PITH_FULL_IMAGE:figures/full_fig_p017_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Performance of LLM2VEC-GEN-Qwen-3-4B employing various response genera￾tors and encoder teachers during training. 1.7B 4B 8B Qwen-3 (Encoder Teacher: LLM2Vec) 55.0 57.5 60.0 62.5 MTEB(eng, v2) Score LLM2Vec (unsupervised) LLM2Vec-Gen 1.7B 4B 8B Qwen-3 (Encoder Teacher: BGE) 55.0 57.5 60.0 62.5 BGE-M3 (unsupervised) LLM2Vec-Gen [PITH_FULL_IMAGE:figures/full_fig_p018_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: LLM2VEC-GEN’s performance with the unsupervised LLM2Vec and BGE teachers on MTEB. These results show that LLM2VEC-GEN consistently outperforms the encoder teacher across model sizes, demonstrating the value of output-centric embeddings. F Additional analysis on retrieval performance of LLM2VEC-GEN [PITH_FULL_IMAGE:figures/full_fig_p018_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: The prompt used for the generation phase of generate-then-encode baseline [PITH_FULL_IMAGE:figures/full_fig_p020_8.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

7 extracted references · 3 linked inside Pith

  1. [4]

    Raghuveer Thirukovalluru and Bhuwan Dhingra

    URLhttps://openreview.net/forum?id=ykuc5q381b. Raghuveer Thirukovalluru and Bhuwan Dhingra. GenEOL: Harnessing the generative power of LLMs for training-free sentence embeddings. In Luis Chiruzzo, Alan Ritter, and Lu Wang (eds.),Findings of the Association for Computational Linguistics: NAACL 2025, pp. 2295–2308, Albuquerque, New Mexico, April 2025. Assoc...

    doi:10.18653/v1/2025.findings-naacl.122 2025

  2. [5]

    Ada Defne Tur, Nicholas Meade, Xing Han L `u, Alejandra Zambrano, Arkil Patel, Esin Durmus, Spandana Gella, Karolina Stanczak, and Siva Reddy

    URLhttps://openreview.net/forum?id=okjogxO1Fu. Ada Defne Tur, Nicholas Meade, Xing Han L `u, Alejandra Zambrano, Arkil Patel, Esin Durmus, Spandana Gella, Karolina Stanczak, and Siva Reddy. SafeArena: Evaluating the safety of autonomous web agents. In Aarti Singh, Maryam Fazel, Daniel Hsu, Si- mon Lacoste-Julien, Felix Berkenkamp, Tegan Maharaj, Kiri Wags...

    2025

  3. [642]

    Summarize the following passage

    URLhttps://aclanthology.org/2024.acl-long.642/. 13 Preprint. Under review. Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. InInternational Conference on Machine Learning, pp. 9929–9939. PMLR, 2020. URL https://proceedings.mlr.press/ v119/wang20k/wang20k.pdf. An Yang, A...

    Pith/arXiv arXiv 2024

  4. [2014]

    [...] Table 13: Analysis of Qwen-3-4B responses to example questions, when trained with LLM2VEC-GENwith and without theL recon

    = 2014! / (2014! * (2014 - 2014)!) = 2014! / (2014! * 0!). [...] Table 13: Analysis of Qwen-3-4B responses to example questions, when trained with LLM2VEC-GENwith and without theL recon. 27

    2014

  5. [2021]

    doi: 10.18653/v1/2021.emnlp-main.552

    Association for Computational Linguistics. doi: 10.18653/v1/2021.emnlp-main.552. URLhttps://aclanthology.org/2021.emnlp-main.552/. Jie He, Richard He Bai, Sinead Williamson, Jeff Z. Pan, Navdeep Jaitly, and Yizhe Zhang. Clara: Bridging retrieval and generation with continuous latent reasoning, 2025. URL https://arxiv.org/abs/2511.18659. Hai Huang, Yann Le...

    arXiv 2021

  6. [2025]

    Jinhyuk Lee, Zhuyun Dai, Xiaoqi Ren, Blair Chen, Daniel Cer, Jeremy R

    URLhttps://openreview.net/forum?id=lgsyLSsDRe. Jinhyuk Lee, Zhuyun Dai, Xiaoqi Ren, Blair Chen, Daniel Cer, Jeremy R. Cole, Kai Hui, Michael Boratko, Rajvi Kapadia, Wen Ding, Yi Luan, Sai Meher Karthik Duddu, Gustavo Hernandez Abrego, Weiqiang Shi, Nithi Gupta, Aditya Kusupati, Prateek Jain, Siddhartha Reddy Jonnalagadda, Ming-Wei Chang, and Iftekhar Naim...

    Pith/arXiv arXiv 2024

  7. [2026]

    URLhttps://arxiv.org/abs/2602.00462. Nathan Lambert, Jacob Morrison, Valentina Pyatkin, Shengyi Huang, Hamish Ivison, Faeze Brahman, Lester James Validad Miranda, Alisa Liu, Nouha Dziri, Xinxi Lyu, Yuling Gu, Saumya Malik, Victoria Graf, Jena D. Hwang, Jiangjiang Yang, Ronan Le Bras, Oyvind Tafjord, Christopher Wilhelm, Luca Soldaini, Noah A. Smith, Yizho...

    Pith/arXiv arXiv 2025