Concept Tokens: Learning Behavioral Embeddings Through Concept Definitions

Aiala Ros\'a; Ignacio Sastre

REVIEW 4 major objections 4 minor 11 references

Concept Tokens: Learning Behavioral Embeddings Through Concept Definitions

T0 review · 4 major / 4 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read A single special token, trained only on text definitions of a concept, can steer a frozen large language model to suppress or amplify that concept's behavior.

desk verdict Directional steering mostly works, but the hallucination evidence is underpowered; the recasting experiment is the convincing part. read the letter →

arxiv 2601.04465 v1 pith:QHMYZJXO submitted 2026-01-08 cs.CL cs.AIcs.LG

Ignacio Sastre , Aiala Rosá This is my paper

classification cs.CLcs.AIcs.LG

keywords concepttokensbehavioralembeddingslanguagemodelsteeringhallucinationrecastingfrozendefinitionalcorpusinstructionfollowing

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper proposes Concept Tokens: new special tokens whose embeddings are learned from natural-language definitions of a target concept while the language model is kept frozen. The central claim is that such a token becomes a compact, directional control signal: asserting the token in a prompt amplifies the associated behavior, negating it suppresses the behavior. Evidence comes from two behavioral tasks: hallucination in closed-book question answering (negation cuts hallucinated answers from 28.7% to 21.9%, mostly by increasing abstentions; assertion raises them to 31.2%) and recasting in second-language teaching (assertion raises recasting from 23.3% to 62.3% on erroneous answers). A qualitative study suggests the token acts as a semantic attractor rather than a reliable store of novel facts. If the claim is right, concept tokens offer a lightweight way to add concepts and steer behavior without fine-tuning weights.

What carries the argument

The key mechanism is the concept token embedding itself: a new row in the model's input embedding matrix, trained by backpropagating the next-token cross-entropy loss on an 'instantiated definitional corpus'—definitions where every mention of the concept is replaced by the token. Because the model is frozen, all learning must fit into the single embedding vector, and the diverse definitions act as competing constraints that force a generalized representation. The same token is then used directionally in prompts: asserted (as an instruction to generate the concept) or negated (as an instruction not to generate it).

What would settle it

Train two concept tokens from the same prompt template—one on definitions of 'hallucinations' and one on definitions of an unrelated concept (e.g., 'uncertainty')—and run the same closed-book question-answering evaluation. If both tokens produce identical shifts in hallucination and abstention rates, the effect is not concept-specific. Alternatively, probe the model's internal activations at the token position with a linear probe: if the embedding does not move the representation along a direction associated with 'hallucination' (as identified by dictionary learning), the semantic-reading assu

Watch

Extended reading notes

Core claim

The paper's central discovery is that a single newly added input embedding, optimized with the standard language-modeling objective on a set of definitions of a concept (with mentions replaced by the token), can act as a learned behavioral switch in a frozen pretrained model. The embedding is trained to satisfy multiple, sometimes conflicting, definitional constraints, so it must generalize rather than memorize. At inference, inserting the token in an instruction ('Generate <token>') increases the target behavior, while negating it ('Do not generate <token>') suppresses it. This directional effect appears in two independent tasks—hallucination avoidance and recasting of learner errors—and th

Load-bearing premise

The load-bearing premise is that a single optimized embedding is read by the frozen model as a meaningful semantic operator—so that 'do not generate <token>' actually suppresses the concept rather than merely producing a surface pattern—and no activation-level evidence is provided to confirm this.

Editorial extensions

If this is right

Negating a concept token can reduce a target behavior like hallucination, but the paper shows this works mainly by increasing abstention, not by improving the precision of attempted answers.
Asserting a concept token can induce a complex behavior (recasting an erroneous learner answer) with a single token, outperforming the word 'recasting' in the prompt and nearly matching a full in-context definition.
Concept tokens preserve instruction-following better than long in-context definitions: the recasting token kept follow-up questions at 98.04% versus 63.07% for the definitional baseline.
Concept tokens can add a novel concept to a frozen model, but the resulting representation is a semantic theme rather than a reliable fact table; factual details remain prone to hallucination.
The approach works on a quantized, instruction-tuned model, suggesting it is feasible for smaller on-premise deployments.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If the directional effect is mediated by the embedding's position in the model's semantic space, then combining several concept tokens in one prompt might steer behavior along multiple independent dimensions (the paper lists this as future work; we note it would be the natural next test).
The paper's next-token observation in the towers experiment—geographically proximate cities competing for the token's location—suggests that a single vector is low-capacity; pairing concept tokens with a fact-retrieval mechanism could fix the factual gaps while keeping the behavioral steering.
The result that negation mainly increases abstentions implies that 'hallucination' as learned from definitions is a behavior to avoid rather than a calibration signal; one could test whether training the token on 'uncertainty' definitions produces more selective abstention.
A cautious extension: the same method could be used to inject concepts like 'safety' or 'toxicity' as tokens, but the paper gives no evidence about whether negating such a token would actually reduce unsafe outputs or just shut down generation.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

I read the full preprint. The central claim—a single embedding trained only on definitions can directionally steer a frozen model—holds up directionally, but the strength of evidence varies a lot across experiments. The recasting result is solid: asserting the token raises recasting from 23 to 62 percent on error-containing answers, and the token strategy preserves follow-up questions at 98 percent versus 63 for the in-context definition baseline. That trade-off is the most useful finding in the paper. The hallucination experiment is weaker: negating the token cuts hallucinations from 28.7 to 21.9 percent, mostly by increasing abstentions, and precision is flat. The asserted token's increase to 31.2 percent versus the 28.7 baseline is within sampling noise; treat that sub-claim as exploratory until error bars appear.

What's new: this extends the memory-token line from compressing a single sequence to learning a multi-definition concept embedding, and the directional prompting trick is a nice practical idea. The comparison against in-context definitions, especially the follow-up question metric, is a useful contribution. The paper is well-written, ships full prompts, annotation guidelines, and code/data, and the limitations section is honest.

Soft spots: all behavior results come from one 4-bit Llama 3.1 8B with greedy decoding and single runs—no confidence intervals or significance tests. The 1000-instance HotpotQA subset's selection is unspecified. Hyperparameters were tuned on a small behavior-probing development set; that's a mild risk but the eval sets are separate. The towers experiment is qualitative but appropriately cautious: the token activates known concepts well but doesn't store novel facts reliably. The 'semantically compositional operator' assumption is untested, but the directional results are real even if the mechanism is unclear.

Who this is for: people working on cheap steering for small on-prem models, and the recasting+follow-up trade-off is relevant for tutoring systems. It deserves a serious referee; send it to review as a solid empirical paper, likely with major revisions for statistical rigor. I agree with the conditional verdict.

Referee Report

4 major / 4 minor

Summary. The paper introduces Concept Tokens, a method for adding a new special token to a frozen pretrained LLM and learning only its input embedding from a corpus of natural-language definitions of a target concept. At inference, the token is inserted into prompts and can be asserted or negated to steer behavior. The authors evaluate the method in three settings: (1) reducing hallucinations in closed-book QA on a 1000-instance HotpotQA subset, where negating the hallucination token lowers hallucinated answers (28.7% to 21.9%) but mainly by increasing abstentions, while asserting it raises hallucinations to 31.2% and lowers precision; (2) inducing recasting in a second-language teaching dialog, where asserting the recasting token raises recasting from 23.26% to 62.33% on erroneous student answers and preserves follow-up questions (98.04%) better than the in-context definition baseline (63.07%); and (3) a qualitative study with the Eiffel Tower and a fictional 'Austral Tower', showing the embeddings activate a coherent semantic theme but do not reliably store novel factual details. The paper argues that concept tokens provide a compact, definition-driven control signal for frozen LLMs.

Significance. If validated, Concept Tokens would be a lightweight and interpretable alternative to in-context prompting for behavior steering, with potential applications in low-resource or privacy-sensitive settings. The method is simple, does not require labeled behavioral examples, and the authors provide code and data. The recasting experiment is particularly suggestive: the concept token achieves comparable or superior compliance with secondary instructions compared to a long in-context definition, a result that is practically relevant. The paper is also unusually candid about limitations, including the abstention-driven hallucination reduction and the under-determination of factual details in the towers study. However, the central claims rest on evidence that is currently under-powered (single run, one model, one quantization) and lacks a crucial control for the mere presence of a newly added token. The significance would be considerably strengthened by a more rigorous statistical treatment and by controls that isolate the contribution of the learned embedding.

major comments (4)

[§4.1, Table 1] The key directional claims for hallucination steering rest on single-run percentages from a 1000-instance subset. The asserted-token vs. no-instruction gap (31.2% vs. 28.7%) is small; with n=1000 the standard error is about 1.4 points, so a 2.5-point difference is not statistically meaningful. The negated-token gap (21.9% vs. 28.7%) is larger, but the paper itself attributes it primarily to increased abstention. Please provide confidence intervals or significance tests (e.g., bootstrap, multiple seeds, or multiple randomly selected subsets), and specify how the 1000 instances were selected (random, stratified, first N?). Without this, the hallucination-steering claim is not adequately supported.
[§4.1.1 / §4.2.1] There is no control for the effect of introducing a new special token into the vocabulary. The observed behavioral changes could in principle be caused by the model's reaction to any novel token (or to an out-of-vocabulary string), rather than by the semantics learned from the definitional corpus. Please include a control condition with (a) an untrained/randomly initialized embedding for the same token, or (b) a token trained on a definition of an unrelated concept (e.g., 'gravity') using the same protocol and prompt frames. This control is necessary to attribute the directional effects to the learned concept representation.
[§3, central hypothesis] The central hypothesis—'the best possible embedding ... is the one that better captures the original concept'—is not formally defined or independently tested. The paper later shows (towers experiment) that the embedding does not reliably capture factual details, and the §4.1 results show that the hallucination token mainly affects abstention behavior. Please operationalize what 'captures the concept' means in a testable way (e.g., a predicted ordering of behavioral outcomes across asserted, negated, and baseline conditions) and specify which observations would falsify it. As stated, the hypothesis is close to a definition and is not informative about the mechanism.
[Appendix C / §4.1.1] The evaluation pipeline has several free parameters that are not systematically varied or reported with sensitivity analysis: training hyperparameters (learning rate, epochs) are chosen on a small hand-selected development set (10 prompts) without reporting the chosen values for each concept; the definitional corpora are generated by GPT-5 with no human verification of their quality or diversity; and the HotpotQA subset is described only as '1000 instances from the validation set' without selection criteria. Since the method's core claim is that definition-only supervision can steer behavior, the results should demonstrate robustness across at least a few corpus designs and subset choices, or explicitly report the selection and hyperparameter values so readers can assess the degrees of freedom.

minor comments (4)

[§4.1.1, HotpotQA subset] Please state whether the 1000 instances were randomly sampled and whether any filtering was applied (e.g., by question type or answer length). This affects the generalizability of the reported percentages.
[§4.2.2, Table 3] The follow-up-question rate is reported for only two methods (importantly, the two that differ most in prompt length). It would be useful to also report the follow-up rate for 'No instruction' and 'Recasting mention' to understand the baseline tendency to ask follow-ups.
[§3, notation] The notation for the concept token embedding alternates between 'e_c' and 'ec' in the text; unify to avoid confusion.
[Appendix E.2] In the recasting prompts, the concept-token condition and the 'recasting mention' condition differ only by replacing the token with the word 'recasting'. However, the system prompt already contains the word 'technique' and the word 'correct mistakes', which may prime the model. Consider a control that uses an unrelated word in the same position.

Circularity Check

1 steps flagged · score 2.0 of 10

Mostly non-circular: the behavioral tests are independent of the embedding training, though the §3 'captures the concept' hypothesis is a near-tautology and the dev-selected hyperparameters add mild self-confirmation.

self definitional [Section 3, 'Concept Tokens' hypothesis paragraph]
"Therefore, we propose the following hypothesis: given a definitional corpus Dc with multiple definitions, the best possible embedding ec to be learned in the process described before, is the one that better captures the original concept c."

The embedding ec is defined as the optimizer of the language-modeling loss on the instantiated definitional corpus Dc(tc). If 'captures the concept c' is operationalized by that same corpus, the hypothesis is true by construction rather than an independent empirical claim. The paper's behavioral experiments are not logically entailed by this tautology and are tested on separate data, so the circularity is limited to the framing; the directional steering results remain empirical findings.

full rationale

The central behavioral claims are not forced by the training objective. In §4.1 and §4.2, the embeddings are trained only on definitional corpora with the LLM frozen; the HotpotQA subset and the 306 QA pairs are not used to train or select the embeddings, and the assert/negate direction is tested on held-out prompts. The reported directional effects (hallucination 21.9 vs 28.7 and 31.2; recasting 62.33 vs 23.26) could plausibly have failed and are therefore empirical. The paper also reports divergence between the concept-token condition and the in-context baseline (κ = 0.35), showing the results are not just a restatement of the definitional input. The main mild issue is the §3 hypothesis, which is close to defining 'captures the concept' as performance on the definitional corpus; the stronger behavioral interpretation is validated experimentally rather than derived from that definition. The self-citations to Sastre and Rosá (2025) are motivational (memory tokens as a special case) and used for noting computational cost; they are not load-bearing for any result. Absence of confidence intervals and selection of hyperparameters on development prompts targeting the behavior are genuine statistical weaknesses, but they are concerns about robustness, not circularity.

Assumptions & free parameters 3 free parameters · 5 assumptions · 2 invented entities

The method assumes pretrained frozen LLM competence, faithful synthetic definitions, judge validity, and that optimizing one embedding on definitions yields a generalizable concept. Hyperparameters and corpus design are tuned by hand; no independent evidence for the token's semantics beyond the paper's own tests.

free parameters (3)

Training hyperparameters (LR, epochs) = Hallucinations: LR 2e-4, 200 epochs + 200 concatenated epochs; Recasting: LR 2e-4, 200 epochs; Towers: LR 2e-2, 3 epochs
Manually tuned on ~10 hand-designed development prompts per behavior (Appendix C); not derived.
Definitional corpus design = Hallucinations: 20 paragraphs/102 occurrences; Recasting: 8 paragraphs/64 occurrences; Towers: chunked Wikipedia-style a
Chosen by the authors; no systematic study of how corpus size, diversity, or occurrence placement affects behavior.
HotpotQA evaluation subset = 1000 of 7405 validation instances
Selection procedure and random seed are not reported (§4.1.1).

assumptions (5)

domain assumption Pretrained Llama 3.1 8B Instruct with 4-bit quantization is a suitable frozen base model for all three tasks.
All experiments use this single model; the generality of the method across model families is not tested (Limitations).
domain assumption Synthetic definitional corpora generated by GPT-5 / Claude are faithful definitions of the target concepts.
The hallucination and recasting corpora were generated by GPT-5, and the Austral Tower article by Claude 4 Sonnet; no human verification of definition quality is reported beyond examples.
domain assumption LLM-as-a-judge labels correlate with human truth.
Validated on 100 HotpotQA instances (Cohen's κ=0.88), but the judge prompt includes the gold answer and the judge model is not independently audited.
domain assumption Cross-entropy LM loss on definitions is a sufficient training signal to induce a generalizable concept representation.
This is the paper's central hypothesis, assumed by the method (Section 3); it is tested behaviorally but not mechanistically.
domain assumption The directional prompting protocol (assert/negate) is semantically transparent to the frozen model.
The model must interpret 'Generate <tc>' and 'Do not generate <tc>' as concept-affirmation and concept-suppression; the paper does not analyze internal activations.

invented entities (2)

Concept token (tc)
purpose: New special token whose embedding is trained on definitions to steer frozen LLM behavior
The token is a new artifact introduced by the method; its only evidence is the paper's own behavioral experiments, with no external validation of the embedding itself.
Austral Tower (fictional landmark)
purpose: Test concept for whether a novel concept can be internalized from a synthetic Wikipedia-style article
Deliberately fictional; no external factual grounding (Appendix B.3), used to probe generalization and factual storage limits.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Concept Tokens: Learning Behavioral Embeddings Through Concept Definitions." pith.science (2026). https://pith.science/paper/QHMYZJXO

@misc{pith2026260104465,
  author       = {Pith},
  title        = {Pith review of: Concept Tokens: Learning Behavioral Embeddings Through Concept Definitions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QHMYZJXO}},
  note         = {Machine review of arXiv:2601.04465}
}

read the original abstract

We propose Concept Tokens, a lightweight method that adds a new special token to a pretrained LLM and learns only its embedding from multiple natural language definitions of a target concept, where occurrences of the concept are replaced by the new token. The LLM is kept frozen and the embedding is optimized with the standard language-modeling objective. We evaluate Concept Tokens in three settings. First, we study hallucinations in closed-book question answering on HotpotQA and find a directional effect: negating the hallucination token reduces hallucinated answers mainly by increasing abstentions, whereas asserting it increases hallucinations and lowers precision. Second, we induce recasting, a pedagogical feedback strategy for second language teaching, and observe the same directional effect. Moreover, compared to providing the full definitional corpus in-context, concept tokens better preserve compliance with other instructions (e.g., asking follow-up questions). Finally, we include a qualitative study with the Eiffel Tower and a fictional "Austral Tower" to illustrate what information the learned embeddings capture and where their limitations emerge. Overall, Concept Tokens provide a compact control signal learned from definitions that can steer behavior in frozen LLMs.

Figures

Figures reproduced from arXiv: 2601.04465 by the authors.

**Figure 3.** Qualitative examples comparing concept to [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

**Figure 4.** Category proportions (Recasting or Explicit correction) for three conditions: concept token negated, no instruction, and concept token asserted. 4.2.2 Results [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

Figures from the paper (1 more)

**Figure 5.** Figure 5: Qualitative examples comparing concept to [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

11 extracted references · 5 linked inside Pith

[2]

InProceedings of the 2025 Conference on Empiri- cal Methods in Natural Language Processing, pages 28949–28961, Suzhou, China

Toward ef- ficient sparse autoencoder-guided steering for im- proved in-context learning in large language models. InProceedings of the 2025 Conference on Empiri- cal Methods in Natural Language Processing, pages 28949–28961, Suzhou, China. Association for Com- putational Linguistics. Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen ...

2025
[3]

Preprint, arXiv:2507.06261

Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. Preprint, arXiv:2507.06261. Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al- Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, and 1 others

arXiv
[5]

Yuri Kuratov, Mikhail Arkhipov, Aydar Bulatov, and Mikhail Burtsev

Why language models hallucinate.Preprint, arXiv:2509.04664. Yuri Kuratov, Mikhail Arkhipov, Aydar Bulatov, and Mikhail Burtsev

arXiv
[7]

Prefix-tuning: Optimizing continuous prompts for generation. In Proceedings of the 59th Annual Meeting of the Asso- ciation for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 4582– 4597, Online. Association for Computational Lin- guistics. Zongqian Li, Yinhong Liu, Yixuan...

2025
[9]

Adly Templeton, Tom Conerly, Jonathan Marcus, Jack Lindsey, Trenton Bricken, and 1 others

Large concept models: Language mod- eling in a sentence representation space.Preprint, arXiv:2412.08821. Adly Templeton, Tom Conerly, Jonathan Marcus, Jack Lindsey, Trenton Bricken, and 1 others

arXiv
[2018]

InProceedings of the 2018 Conference on Empiri- cal Methods in Natural Language Processing, pages 2369–2380, Brussels, Belgium

HotpotQA: A dataset for diverse, explainable multi-hop question answering. InProceedings of the 2018 Conference on Empiri- cal Methods in Natural Language Processing, pages 2369–2380, Brussels, Belgium. Association for Com- putational Linguistics. Hanyu Zhang, Xiting Wang, Chengao Li, Xiang Ao, and Qing He

2018
[2021]

InProceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 3045–3059, Online and Punta Cana, Domini- can Republic

The power of scale for parameter-efficient prompt tuning. InProceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 3045–3059, Online and Punta Cana, Domini- can Republic. Association for Computational Lin- guistics. Xiang Lisa Li and Percy Liang

2021
[2022]

InFindings of the Asso- ciation for Computational Linguistics: ACL 2022, pages 566–581, Dublin, Ireland

Extracting latent steering vectors from pretrained language models. InFindings of the Asso- ciation for Computational Linguistics: ACL 2022, pages 566–581, Dublin, Ireland. Association for Computational Linguistics. LCM team, Loïc Barrault, Paul-Ambroise Duquenne, Maha Elbayad, Artyom Kozhevnikov, Belen Alas- truey, Pierre Andrews, Mariano Coria, Guillaum...

2022

Show all 11 references

[2023]

Eiffel Tower

Judging LLM-as-a-judge with MT-bench and chatbot arena. InThirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track. A Towers experiment: setup and results A.1 Setup We divided the experiment into two similar entities: (1) the Eiffel Towe...

1925
[2024]

Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, and Ting Liu

The llama 3 herd of models.Preprint, arXiv:2407.21783. Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, and Ting Liu

arXiv
[2025]

Preprint, arXiv:2507.21509

Persona vectors: Monitoring and controlling character traits in language models. Preprint, arXiv:2507.21509. Ikhyun Cho and Julia Hockenmaier

arXiv

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.

[1] [2]

InProceedings of the 2025 Conference on Empiri- cal Methods in Natural Language Processing, pages 28949–28961, Suzhou, China

Toward ef- ficient sparse autoencoder-guided steering for im- proved in-context learning in large language models. InProceedings of the 2025 Conference on Empiri- cal Methods in Natural Language Processing, pages 28949–28961, Suzhou, China. Association for Com- putational Linguistics. Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen ...

2025

[2] [3]

Preprint, arXiv:2507.06261

Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. Preprint, arXiv:2507.06261. Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al- Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, and 1 others

arXiv

[3] [5]

Yuri Kuratov, Mikhail Arkhipov, Aydar Bulatov, and Mikhail Burtsev

Why language models hallucinate.Preprint, arXiv:2509.04664. Yuri Kuratov, Mikhail Arkhipov, Aydar Bulatov, and Mikhail Burtsev

arXiv

[4] [7]

Prefix-tuning: Optimizing continuous prompts for generation. In Proceedings of the 59th Annual Meeting of the Asso- ciation for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 4582– 4597, Online. Association for Computational Lin- guistics. Zongqian Li, Yinhong Liu, Yixuan...

2025

[5] [9]

Adly Templeton, Tom Conerly, Jonathan Marcus, Jack Lindsey, Trenton Bricken, and 1 others

Large concept models: Language mod- eling in a sentence representation space.Preprint, arXiv:2412.08821. Adly Templeton, Tom Conerly, Jonathan Marcus, Jack Lindsey, Trenton Bricken, and 1 others

arXiv

[6] [2018]

InProceedings of the 2018 Conference on Empiri- cal Methods in Natural Language Processing, pages 2369–2380, Brussels, Belgium

HotpotQA: A dataset for diverse, explainable multi-hop question answering. InProceedings of the 2018 Conference on Empiri- cal Methods in Natural Language Processing, pages 2369–2380, Brussels, Belgium. Association for Com- putational Linguistics. Hanyu Zhang, Xiting Wang, Chengao Li, Xiang Ao, and Qing He

2018

[7] [2021]

InProceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 3045–3059, Online and Punta Cana, Domini- can Republic

The power of scale for parameter-efficient prompt tuning. InProceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 3045–3059, Online and Punta Cana, Domini- can Republic. Association for Computational Lin- guistics. Xiang Lisa Li and Percy Liang

2021

[8] [2022]

InFindings of the Asso- ciation for Computational Linguistics: ACL 2022, pages 566–581, Dublin, Ireland

Extracting latent steering vectors from pretrained language models. InFindings of the Asso- ciation for Computational Linguistics: ACL 2022, pages 566–581, Dublin, Ireland. Association for Computational Linguistics. LCM team, Loïc Barrault, Paul-Ambroise Duquenne, Maha Elbayad, Artyom Kozhevnikov, Belen Alas- truey, Pierre Andrews, Mariano Coria, Guillaum...

2022