Aligning Large Language Models to Follow Instructions and Hallucinate Less via Effective Data Filtering

Baobao Chang; Cheng Gao; Chen Qian; Fanchao Qi; Gang Chen; Haozhe Zhao; Kaikai An; Kangyang Luo; Maosong Sun; Shuzheng Si

REVIEW 3 major objections 4 minor 15 references

Aligning Large Language Models to Follow Instructions and Hallucinate Less via Effective Data Filtering

T0 review · 3 major / 4 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read Filtering instruction data by what the base model already knows can reduce hallucinations without weakening instruction-following.

desk verdict The empirical story is plausible and well-tested, but the paper's central ICP formula is mathematically broken as written, so the method cannot be reproduced until that is fixed. read the letter →

arxiv 2502.07340 v3 pith:L5AGRODL submitted 2025-02-11 cs.CL cs.AI

Shuzheng Si , Haozhe Zhao , Gang Chen , Cheng Gao , Yuzhuo Bai , Zhitong Wang , Kaikai An , Kangyang Luo

show 4 more authors

Chen Qian Fanchao Qi Baobao Chang Maosong Sun

This is my paper · ORCID

classification cs.CLcs.AI

keywords instructiontuninghallucinationreductiondatafilteringknowledgefamiliarityinternalconsistencyprobingsemanticequivalenceidentificationfactualitybenchmarkssupervisedfine-tuning

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

The paper's claim is that the hallucinations introduced during instruction tuning can be largely avoided by choosing training data the model already knows, and that this can be measured before any fine-tuning happens. It proposes NOVA, which scores each instruction sample twice: how consistently the base model answers the instruction on its own, and how many of those self-generated answers are semantically equivalent to the target response. Samples that score well on both, combined with a separately trained quality score, are used for supervised fine-tuning instead of the full dataset. The paper reports that this filter reduces hallucinations on factuality and faithfulness benchmarks while keeping instruction-following ability at or above the vanilla full-data baseline, without the extra RL stage used by prior hallucination-reduction methods.

What carries the argument

The machinery is the familiarity score $F_{\mathrm{familiarity}}(q,r)=F_{\mathrm{res}}(r)/F_{\mathrm{ins}}(q)$ computed on the base model before fine-tuning. $F_{\mathrm{ins}}$ is the differential entropy of the covariance matrix of $K=10$ last-layer sentence embeddings of self-generated responses, so lower values mean the model answers the instruction consistently; $F_{\mathrm{res}}$ is the share of generated responses that an NLI model places in the same semantic cluster as the target response, so higher values mean the target's content is already known. The ratio ranks the dataset before supervised fine-tuning, and the final selection averages this familiarity rank with a quality-reward rank to keep fluency and complexity from being ignored. The score is doing the load-bearing work of turning the observation that unfamiliar knowledge causes hallucinations into a per-sample filter.

What would settle it

Run NOVA with the published recipe on an instruction set whose facts are deliberately outside the base model's knowledge, then compare the top-5 percent selected subset against a random 5 percent subset on BioGEN and LongFact: if the familiarity-selected subset does not reduce hallucination scores, the proxy fails. A separate check is computational: for $K=10$ embeddings in dimension 4096, the covariance matrix has rank at most 9, so the determinant in Eq. (2) is undefined unless regularization or projection is applied, and reproducing the published ICP scores requires specifying which.

Watch

Extended reading notes

Core claim

The central discovery is that a base model's pre-fine-tuning behavior can reveal which instruction samples contain knowledge it has already internalized, and that training only on those samples makes the aligned model hallucinate less. NOVA defines this familiarity operationally: for an instruction $q$, the base model generates $K$ responses; the spread of their sentence embeddings, measured as differential entropy, gives the instruction familiarity $F_{\mathrm{ins}}(q)$, and the fraction of those responses that fall in the semantic cluster containing the target response $r$ gives the response familiarity $F_{\mathrm{res}}(r)$. The two are combined as $F_{\mathrm{familiarity}} = F_{\mathrm{res}} / F_{\mathrm{ins}}$, then averaged by rank with an expert-aligned quality reward model. Training on the top-ranked samples (5-15 percent of the dataset) is claimed to improve BioGEN, LongFact, and FollowRAG-faithfulness scores relative to full-data instruction tuning and to match vanilla tuning on MT-Bench and FollowRAG-instruction.

Load-bearing premise

The load-bearing bet is that the base model's self-consistency and the semantic overlap of its own answers with the target response, measured before fine-tuning, reliably indicate which instruction data contains knowledge the model already has--and that the entropy score built from ten high-dimensional embeddings has a well-defined nonzero determinant.

Editorial extensions

If this is right

Training on the top 5-15 percent of instruction data selected by familiarity and quality improves hallucination benchmarks by roughly 3.5-8.6 points on BioGEN, 2.0-5.1 on LongFact, and 4.9-8.5 on FollowRAG-faithfulness compared with full-data tuning.
The filter keeps instruction-following ability intact, so a practitioner can reduce hallucinations without the extra preference-data collection and RL stage that prior factuality-alignment methods require.
Both components of the familiarity score contribute: ablations removing ICP or SEI reduce the hallucination gain, and removing the quality reward model hurts instruction-following.
The method transfers to other base models and to a larger scale, so the filtering effect is not specific to a single checkpoint.

Reading between the lines

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

Implicit but untested in the paper: the same familiarity-before-training filter could apply to continued pretraining, domain adaptation, or continual learning, where injecting out-of-distribution knowledge is the same risk.
The rank-averaged combination of familiarity and quality means that on a dataset whose highest-quality samples are all unfamiliar, the method must trade one objective against the other; working out that trade-off curve explicitly would show where NOVA stops being a familiarity filter.
Because the paper's gains come from 5-15 percent of the full dataset, the result suggests a cost-shifting principle: spend compute on offline filtering, then train on a smaller and more reliable set rather than train on everything and repair hallucinations afterward.

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

The paper's core idea is worth taking seriously: filter instruction data by how familiar the base model already is with the instruction and target response, then fine-tune on that subset. That is a practical, low-cost alternative to RL-based hallucination mitigation, and the combination of internal-state consistency, NLI-based semantic equivalence, and a quality reward model is genuinely new. The authors also do a lot of things right: multiple benchmarks, ablations, transfer to LLaMA-1 and Qwen-2, human evaluation, and honest limitations. If the reported gains are real, this is a useful contribution.

The problem is that the central familiarity metric, ICP, is not well-defined as presented. With K=10 sentence embeddings in a 4096-dimensional space, the sample covariance in Eq. (3) has rank at most 9, so det(Σ)=0 and the differential entropy in Eq. (2) is undefined. Eq. (4) makes it worse: it replaces ½Σlog λᵢ with ½Σλᵢ, which is not a simplification but a different quantity. The reader cannot tell whether the implemented score is differential entropy, half the trace, or something else with jitter or a pseudo-determinant. Since Table 3 credits ICP with a 2.9-point BioGEN improvement, this is a load-bearing flaw, not a typo in a footnote. The authors need to specify the actual computation, add regularization, or provide code.

The other weaknesses are real but less severe. All results are single runs with no error bars, so the 3.5–8.6 point deltas could be partly noise. The hyperparameters K and T are selected by inspecting BioGEN results in the appendix, which makes the reported BioGEN improvements optimistic. There is no matched-size random-subset baseline, though the comparison to IFD/CaR/Nuggets at the same ratios helps. These are fixable with a modest amount of additional work.

The stress-test concern holds up on reading: the ICP metric is internally inconsistent. The good news is that the failure is in the write-up of the metric, not necessarily in the underlying idea. The SEI and quality-reward components are clearly described and plausible, and the overall pattern of improvements across models and dataset sizes is consistent enough that I would not dismiss the paper.

Who is this for? Anyone working on data selection, hallucination mitigation, or instruction tuning. It deserves a serious referee, but the referee should demand a corrected ICP definition, error bars, and a tuning protocol that does not peek at the test set. If those are provided, this could become a solid paper.

Referee Report

3 major / 4 minor

Summary. The paper proposes NOVA, a data-filtering framework for instruction tuning that scores each (instruction, response) pair by how familiar the base LLM is with the instruction and the target response, and then combines this familiarity score with an expert-aligned quality reward model. Familiarity is measured by Internal Consistency Probing (ICP), which computes a differential-entropy-style score over the embeddings of K self-generated responses, and by Semantic Equivalence Identification (SEI), which uses an NLI model to find how many generated responses are semantically equivalent to the target response. The authors fine-tune LLaMA-3-8B, LLaMA-3-70B, LLaMA-1-7B, and Qwen-2-7B on the selected top 5-15% of Alpaca or Alpaca-GPT4 data and report improved factuality (BioGEN, LongFact), faithfulness (FollowRAG), and instruction-following (MT-Bench, FollowRAG-Instruction) compared with vanilla fine-tuning, data-filtering baselines, and RL-based baselines.

Significance. If the claimed results hold, NOVA offers a practical and inexpensive alternative to RL-based hallucination reduction: it requires only offline data filtering before standard SFT and adds no inference-time cost. The paper is commendably extensive: it evaluates on multiple hallucination and instruction-following benchmarks, includes ablations, transferability studies across four base models, human evaluation, and a parameter study. The filtering design is not circular with respect to the evaluation because scores are computed from the base model before fine-tuning and the benchmarks are external. However, the central ICP computation in Section 3.1 is mathematically inconsistent as written, and the main BioGEN numbers are produced with hyperparameters selected by inspecting BioGEN itself. These issues must be resolved before the empirical claims can be accepted.

major comments (3)

[Section 3.1, Eqs. (2)-(4)] The differential-entropy formulation is not well-defined for the stated configuration. With K=10 sentence embeddings in dimension d=4096, the sample covariance matrix in Eq. (3) has rank at most 9, so det(Σ)=0 and log det(Σ) in Eq. (2) is -∞; no regularization, pseudo-determinant, or lower-dimensional projection is described. In addition, the second equality in Eq. (4) is not an identity: Eq. (2) requires (1/2)Σ log λ_i, while Eq. (4) states (1/2)Σ λ_i. These are different quantities unless all nonzero eigenvalues equal 1. Since F_ins(q) enters the final familiarity score in Eq. (9), the paper does not currently specify a computable, well-defined version of its central metric. The authors should state exactly which formula is implemented (e.g., with jitter, pseudodeterminant, or a K×K Gram covariance) and correct Eq. (4).
[Appendix C and Tables 1/3] The hyperparameters K=10 and T=0.7 are selected by inspecting FactScore results on BioGEN (Figure 5 and Table 6), and BioGEN is also the primary hallucination benchmark in Table 1 and the ablation in Table 3. This means the reported BioGEN improvements are not obtained under a fully held-out evaluation protocol. The authors should either fix the hyperparameters before any evaluation, tune on a separate validation set, or explicitly report the selection procedure as part of the method and discuss the risk of overfitting to this benchmark.
[Tables 1-3 and Figure 3] All experimental numbers are single-run point estimates with no error bars, confidence intervals, or significance tests. Several comparisons against strong baselines are within 1-3 points (e.g., Table 1, Alpaca 5% LongFact Objects: 17.9 vs. 16.2 for CaR; Table 2, Alpaca-GPT4 5% MT-Bench: 64.6 vs. 65.0 for IFD), so the claim that NOVA 'significantly reduces hallucinations' is not statistically supported as reported. The authors should provide multiple seeds or a significance analysis for at least the main tables.

minor comments (4)

[Figure 2] The formula in the figure, "R_final = 1/2 R_familiarity R_quality," appears to be missing a plus sign and should read R_final = 1/2 (R_familiarity + R_quality), matching Eq. (10).
[Table 2 caption] The caption contains the typo "FollowRAG-Intruction"; it should be "FollowRAG-Instruction." The same typo appears in the Appendix tables.
[Appendix B] The subsection heading "Traning of Quality Reward Model" should be "Training of Quality Reward Model."
[Section 3.1, Eq. (4)] The notation "F_ins(q)" is introduced via "term as F_ins(q)", which should be "termed F_ins(q)" or "denoted F_ins(q)". This is a minor wording issue but affects readability.

Circularity Check

1 steps flagged · score 3.0 of 10

Mild circularity: the headline BioGEN improvement is partly a fitted value because K and T are selected on BioGEN, but the core familiarity-filtering claim retains independent support from LongFact, FollowRAG, and transfer experiments.

fitted input called prediction [Appendix C (Parameter Study); headline BioGEN results in Table 1 and ablation Table 3]
"As shown in Figure 5, increasing the number of generated responses improves the performance of our method, but when the number of generated responses is greater than 10, the performance will be stable. Therefore, we empirically recommend setting the number of generated responses K to 10... We recommend that the temperature take a moderate value... Table 6: - T = 0.7 (Ours) Alpaca 50.3."

The paper selects the two central hyperparameters of NOVA, K (number of self-generated responses) and T (generation temperature), by inspecting FactScore results on BioGEN (Appendix C, Figure 5 and Table 6). The same BioGEN benchmark is then reported as the headline evidence that NOVA reduces hallucinations (Table 1: NOVA-5% 50.3 vs. Vanilla 42.4; Table 3 ablation uses BioGEN FactScore). Thus the BioGEN deltas are not a clean out-of-sample prediction of the familiarity-filtering effect; they are partly the result of choosing parameters that maximize that specific score.

full rationale

The core NOVA pipeline is not circular as written: ICP and SEI measure familiarity using the base model's own generated responses and NLI-based semantic clustering, before any fine-tuning; the final filter is a rank combination of familiarity and an externally trained quality reward model; and the evaluations use external benchmarks (BioGEN, LongFact, FollowRAG, MT-Bench) plus transfer to LLaMA-1 and Qwen-2. There is no definitional dependence of the predictor on the predicted hallucination scores, and no load-bearing self-citation chain. The self-citations that appear (Si et al. 2024 GATEAU, Li et al. 2024b Nuggets, Si et al. 2023/2025) are related-work or baseline mentions, not the justification for NOVA's central premise. The one genuine circularity-adjacent issue is the parameter study in Appendix C: K=10 and T=0.7 are chosen by inspecting BioGEN FactScore results, and BioGEN is then used as a headline hallucination benchmark, so the BioGEN improvements are partly fitted values. This is mild because the other benchmarks and transfer experiments provide independent evidence and the mechanism is not a fitted re-description of the outputs. Separately, the differential-entropy derivation in Section 3.1 is internally inconsistent (Eq. 4 replaces 1/2 sum log lambda_i with 1/2 sum lambda_i, and the sample covariance of K=10 embeddings in d=4096 is rank-deficient, making det(Sigma)=0), but that is a correctness/verifiability flaw, not a circularity, and does not by itself raise the circularity score. Overall, the circularity is partial and localized to the BioGEN headline claim rather than the whole derivation.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The method relies on prior empirical findings about unfamiliar-data hallucinations, on assumptions that hidden states and differential entropy capture meaningful semantic consistency, and on two hyperparameters tuned on the target benchmark. No new physical or conceptual entities are posited, so the invented-entities list is empty.

free parameters (2)

Number of generated responses K = 10
Chosen by evaluating FactScore on BioGEN (Figure 5); K=10 yields stable performance and is used throughout. This is a hyperparameter tuned on one of the target evaluation benchmarks.
Sampling temperature T = 0.7
Chosen based on BioGEN FactScore results (Table 6); temperature 0 is explicitly discouraged and moderate values are recommended. This value is also fitted on the test benchmark.

assumptions (3)

domain assumption Tuning on instruction data that contains unfamiliar knowledge encourages hallucinations.
Adopted from prior work (Kang et al., 2024; Gekhman et al., 2024) and used as the motivating premise for NOVA. It is not rederived, and its applicability to generic instruction datasets such as Alpaca is assumed without direct validation.
domain assumption The last-token hidden state of the last layer is a valid sentence embedding that captures semantic content.
Used in Section 3.1 to compute embeddings E for the K generated responses; this is an empirically motivated but unproven representation assumption.
domain assumption Lower differential entropy of the embedding covariance matrix corresponds to higher semantic consistency and therefore greater model familiarity.
Eq. (2)-(4) map the differential entropy to a familiarity score. The paper does not validate this mapping against human judgments of response consistency, and the mathematical validity is complicated by the rank-deficiency issue.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Aligning Large Language Models to Follow Instructions and Hallucinate Less via Effective Data Filtering." pith.science (2026). https://pith.science/paper/L5AGRODL

@misc{pith2026250207340,
  author       = {Pith},
  title        = {Pith review of: Aligning Large Language Models to Follow Instructions and Hallucinate Less via Effective Data Filtering},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/L5AGRODL}},
  note         = {Machine review of arXiv:2502.07340}
}

read the original abstract

Training LLMs on data containing unfamiliar knowledge during the instruction tuning stage can encourage hallucinations. To address this challenge, we introduce NOVA, a novel framework designed to identify high-quality data that aligns well with the LLM's learned knowledge to reduce hallucinations. NOVA includes Internal Consistency Probing (ICP) and Semantic Equivalence Identification (SEI) to measure how familiar the LLM is with instruction data. Specifically, ICP evaluates the LLM's understanding of the given instruction by calculating the tailored consistency among multiple self-generated responses. SEI further assesses the familiarity of the LLM with the target response by comparing it to the generated responses, using the proposed semantic clustering and well-designed voting strategy. Finally, to ensure the quality of selected samples, we introduce an expert-aligned reward model, considering characteristics beyond just familiarity. By considering data quality and avoiding unfamiliar data, we can utilize the selected data to effectively align LLMs to follow instructions and hallucinate less.

Figures

Figures reproduced from arXiv: 2502.07340 by the authors.

**Figure 3.** Average perplexity score of 15 samples with the lowest scores for each model from LongFactObjects. Models are trained on Alpaca-GPT4. Model BioGEN MT-Bench NOVA - 5% - 70B 60.9 74.3 -w/o. Data Filtering 53.7 73.2 NOVA - 5% - 8B 50.5 64.6 -w/o. Data Filtering 41.9 64.3 -w/o. ICP 47.6 64.1 -w/o. SEI 48.3 63.8 -w/o. Quality RM 55.6 48.6 -w/o. ICP & SEI 43.7 65.2 [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

**Figure 4.** Human evaluation across four key dimensions. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗

Figures from the paper (4 more)

**Figure 5.** Figure 5: FactScore results on BioGEN with the differ [PITH_FULL_IMAGE:figures/full_fig_p015_5.png]

**Figure 6.** Figure 6: The prompt for identifying the semantic equivalence. [PITH_FULL_IMAGE:figures/full_fig_p018_6.png]

**Figure 7.** Figure 7: The principles of human evaluation. generated by the models. Before participants begin to make judgments, we describe the principles of our design in detail and ensure that each participant correctly understands the principles. If the final result can not be determined…

**Figure 8.** Figure 8: Case study for selected samples. 20 [PITH_FULL_IMAGE:figures/full_fig_p020_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

15 extracted references · 11 canonical work pages

[1]

Factuality: Whether the biography provides relatively more factual statements over the non-factual statements?

work page
[2]

Helpfulness: Whether the biography provides useful information?

work page
[3]

Relevance: Whether the statements contained in the biography relevant to the provided people entity?

work page
[4]

Figure 7: The principles of human evaluation

Naturalness: Whether the biography sound natural and fluent? Finally, please make a decision among 3 opinions, including Win, Tie, and Loss. Figure 7: The principles of human evaluation. generated by the models. Before participants begin to make judgments, we describe the principles of our design in detail and ensure that each participant correctly unders...

work page
[5]

In My Feelings

Drake – “In My Feelings”. ♂laugh-bea¶Sample ranked first according to Rf inalin Eq. (10): Instruction: Given a dictionary, create a JSON representation. dict = {"name": "John", "age": 18, "is_adult": True} Target Response: { "name": "John", "age": 18, "is_adult": true } ⌢ Sample ranked last according to Rf inalin Eq. (10): Instruction: Describe how to imp...

work page
[6]

Identical

and sample 100 instructions and their cor- responding responses as the test set (we find that most of the selected data are in English, but some of the selected instruction types are translation tasks, so a few data contain Chinese responses). Specifi- Model Size BioGEN Alpaca DeBERTa-base-mnli 0.1B 49.7 DeBERTa-large-mnli 0.3B 50.3 DeBERTa-xlarge-mnli 0....

work page 2024
[8]

Prompt Template

We train our model on a single NVIDIA A800. Prompt Template. We use the prompt template from Alpaca (Taori et al., 2023). We keep the same template in training and inference. C Parameter Study We explore the effects of two important hyperpa- rameters in our method: the number of generated responses K and the temperature T during the re- sponse generation....

work page 2023
[11]

Billie Eilish – “Bad Guy”

work page

Show all 15 references

[12]

Shape of You

Ed Sheeran – “Shape of You”
[13]

Summer Nights

Maroon 5 – “Summer Nights”
[14]

Cake by the Ocean

DNCE – “Cake by the Ocean”
[2018]

A broad-coverage challenge corpus for sen- tence understanding through inference. In Proceed- ings of the 2018 Conference of the North American Chapter of the Association for Computational Lin- guistics: Human Language Technologies, Volume 1 (Long Papers), pages 1112–1122. Ass...

2023 arXiv
[2022]

In NeurIPS, pages 27730–27744

Training language models to follow instruc- tions with human feedback. In NeurIPS, pages 27730–27744. Baolin Peng, Chunyuan Li, Pengcheng He, Michel Gal- ley, and Jianfeng Gao. 2023. Instruction tuning with gpt-4. Preprint, arXiv:2304.03277. Rafael Rafailov, Archit Sharma, Eri...

2023 arXiv
[2023]

In Thirty-seventh Conference on Neural Information Processing Sys- tems

Direct preference optimization: Your language model is secretly a reward model. In Thirty-seventh Conference on Neural Information Processing Sys- tems. Vipula Rawte, Amit Sheth, and Amitava Das. 2023. A survey of hallucination in large foundation models. Preprint, arXiv:2309....

2023 arXiv
[2024]

Preprint, arXiv:2307.06290

Instruction mining: Instruction data selec- tion for tuning large language models. Preprint, arXiv:2307.06290. Chao Chen, Kai Liu, Ze Chen, Yi Gu, Yue Wu, Mingyuan Tao, Zhihang Fu, and Jieping Ye. 2024a. INSIDE: LLMs’ internal states retain the power of hallucination detection...

2023 arXiv

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.

[1] [1]

Factuality: Whether the biography provides relatively more factual statements over the non-factual statements?

work page

[2] [2]

Helpfulness: Whether the biography provides useful information?

work page

[3] [3]

Relevance: Whether the statements contained in the biography relevant to the provided people entity?

work page

[4] [4]

Figure 7: The principles of human evaluation

Naturalness: Whether the biography sound natural and fluent? Finally, please make a decision among 3 opinions, including Win, Tie, and Loss. Figure 7: The principles of human evaluation. generated by the models. Before participants begin to make judgments, we describe the principles of our design in detail and ensure that each participant correctly unders...

work page

[5] [5]

In My Feelings

Drake – “In My Feelings”. ♂laugh-bea¶Sample ranked first according to Rf inalin Eq. (10): Instruction: Given a dictionary, create a JSON representation. dict = {"name": "John", "age": 18, "is_adult": True} Target Response: { "name": "John", "age": 18, "is_adult": true } ⌢ Sample ranked last according to Rf inalin Eq. (10): Instruction: Describe how to imp...

work page

[6] [6]

Identical

and sample 100 instructions and their cor- responding responses as the test set (we find that most of the selected data are in English, but some of the selected instruction types are translation tasks, so a few data contain Chinese responses). Specifi- Model Size BioGEN Alpaca DeBERTa-base-mnli 0.1B 49.7 DeBERTa-large-mnli 0.3B 50.3 DeBERTa-xlarge-mnli 0....

work page 2024

[7] [8]

Prompt Template

We train our model on a single NVIDIA A800. Prompt Template. We use the prompt template from Alpaca (Taori et al., 2023). We keep the same template in training and inference. C Parameter Study We explore the effects of two important hyperpa- rameters in our method: the number of generated responses K and the temperature T during the re- sponse generation....

work page 2023

[8] [11]

Billie Eilish – “Bad Guy”

work page