text string | source string |
|---|---|
to Appendix B for baseline details. Name O S G Loss [S(✓) : (x,yw,yl)∼pdata] DPO [41] ✓ ✓ ✗ −logσ βlogπθ(yw|x) πref(yw|x)−βlogπθ(yl|x) πref(yl|x) f-DPO [54] ✗ ✓ ✓ −logσ βf′πθ(yw|x) πref(yw|x) −βf′πθ(yl|x) πref(yl|x) f-PO [59] ✓ ✗ ✓ Df πθ(y|x) 1 Z(x)πref(y|x)exp1 βrϕ∗(x,y) f-PO [59] ✗ ✓ ✓ f σ βlogπθ(yw|x) ... | https://arxiv.org/abs/2505.19601v1 |
where σdenotes the logistic function. The prompt xis drawn from pprompt (x), which we omit in the following formulations for notational simplicity. From the optimality condition of logistic regression, the optimal reward model rϕ∗that minimizes Lreward satisfies pdata(yw≻yl|x) = σ(rϕ∗(x,yw)−rϕ∗(x,yl)), also known as th... | https://arxiv.org/abs/2505.19601v1 |
smoothly interpolates between KLIEP (at λ= 0) and LSIF (at λ= 1). Inspired by the success of generalized likelihood ratio estimation in recent generative modeling studies [31, 25, 9], we apply this extension to generalize the DPO loss. 3 3 Methods This section introduces the proposed Bregman Preference Optimization (BP... | https://arxiv.org/abs/2505.19601v1 |
inside the expectation. To analyze the learning dynamics of Lh BPO(Rθ;pdata), we provide the following gradient analysis: 4 (a)Ghof BA( λ) divergence. (b)Ghof SBA( λ) divergence. (c) Gradient updates. Figure 3: Gradient magnitude and direction analysis across different Bregman divergences. Proposition 4. (Gradient Anal... | https://arxiv.org/abs/2505.19601v1 |
to control whether to prioritize updates for more confident (e.g., Rθ≪1) or less confident (e.g., Rθ≈1) samples. Recent studies [ 56,26] have shown that DPO is sensitive to data quality and that applying confidence-based adjustments to individual samples can be heuristically effective. SBA controls the sensitivity to c... | https://arxiv.org/abs/2505.19601v1 |
original DPO paper [ 41] for a fair comparison. For dialogue generation, we use Pythia-2.8B [ 5] as the pre-trained LLM and perform SFT on the preferred subset of the HH dataset. For summarization, we use a publicly available SFT model [ 8] based on GPT-J [ 53]. All comparisons are conducted on the same SFT model with ... | https://arxiv.org/abs/2505.19601v1 |
When λ=−0.5, the gradient behavior closely resembles that of DPO, as shown in Figure 3b, with both win rate and entropy showing similar trends in Figures 5a and 5b. As λincreases, both metrics improve up to a certain point, in contrast to the instances of f-PO and f-DPO that exhibit a clear trade-off between win rate a... | https://arxiv.org/abs/2505.19601v1 |
DPO [41] 48.2 47.5 35.2 IPO [2] 46.8 42.4 36.6 CPO [58] 34.1 36.4 30.9 KTO [13] 34.1 32.1 27.3 ORPO [18] 38.1 33.8 28.2 R-DPO [40] 48.0 45.8 35.1 SimPO-v2 [36] 53.7 47.5 36.5 BPO 55.9 51.5 38.0Results Table 6 presents the main results using Llama-3-8B-Instruct. Baseline re- sults are taken from SimPO [ 36] and f- PO [ ... | https://arxiv.org/abs/2505.19601v1 |
[ 22] reformulates the entire DPO objective as a distribution matching problem, showing that LDPOin Eq. (4) reduces to the forward KL divergence between the policy model πθ(y|x)and the optimal policy πθ∗(y|x)in Eq. (3). EXO further proposes a reverse KL loss, and f-PO [ 59] subsequently extends this formulation to f-di... | https://arxiv.org/abs/2505.19601v1 |
[5]Stella Biderman, Hailey Schoelkopf, Quentin Gregory Anthony, Herbie Bradley, Kyle O’Brien, Eric Hallahan, Mohammad Aflah Khan, Shivanshu Purohit, USVSN Sai Prashanth, Edward Raff, et al. Pythia: A suite for analyzing large language models across training and scaling. In International Conference on Machine Learning ,... | https://arxiv.org/abs/2505.19601v1 |
by score matching. Journal of Machine Learning Research , 6(4), 2005. [22] Haozhe Ji, Cheng Lu, Yilin Niu, Pei Ke, Hongning Wang, Jun Zhu, Jie Tang, and Minlie Huang. Towards efficient exact optimization of language model alignment. In Forty-first International Conference on Machine Learning , 2024. [23] Albert Q. Jian... | https://arxiv.org/abs/2505.19601v1 |
Inc., 2007. [38] Sebastian Nowozin, Botond Cseke, and Ryota Tomioka. f-gan: Training generative neural samplers using variational divergence minimization. Advances in neural information processing systems , 29, 2016. [39] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang,... | https://arxiv.org/abs/2505.19601v1 |
Model.https://github.com/kingoflolz/mesh-transformer-jax , May 2021. [54] Chaoqi Wang, Yibo Jiang, Chenghao Yang, Han Liu, and Yuxin Chen. Beyond reverse KL: Generalizing direct preference optimization with diverse divergence constraints. In The Twelfth International Conference on Learning Representations , 2024. [55] ... | https://arxiv.org/abs/2505.19601v1 |
assumptions ensure that our target quantities, including likelihood ratios, are well-defined. The non-negative assumption for pdatais naturally aligned with the Bradley-Terry model, which utilizes a sigmoid function to model pairwise probabilities. Given that this model inherently assumes a non-zero probability for eac... | https://arxiv.org/abs/2505.19601v1 |
back into the above objective, the resulting objective becomes: Dh(Rdata||Rθ) =Epdata(yw≻yl|x) h′(Rθ)Rθ−h(Rθ)−h′ R−1 θ −C. ⇔ Lh BPO(Rθ;pdata) =Dh(Rdata||Rθ) +C 16 A.4 Proof of Proposition 4 Proposition 4. (Gradient Analysis) Let the gradient of the BPO objective be expressed as: ∇θLh BPO(Rθ;pdata) =Epdata(yw≻yl|x)[... | https://arxiv.org/abs/2505.19601v1 |
the matching problem as Df(ˆπθ||ˆπ∗), which still guarantees πθconverges at π∗ θat the optimum. We denote this property of the objective as Oin Table 1. The f-divergence objective becomes: Df(ˆπθ||ˆπ∗) =Eπref πrϕ∗f˜πθ πrϕ∗ ,where πrϕ∗=exp(rϕ∗(x,y)) Zrϕ∗(x),˜πθ=h πθ(y|x) πref(y|x)iβ ˜Zθ(x), with partition functions ... | https://arxiv.org/abs/2505.19601v1 |
=−Epdata(yw≻yl|x) logσβ |yw|logπθ(yw|x)−β |yl|logπθ(yl|x)−γ , where |yw|and|yl|are the response lengths, and γis the target reward margin. The corresponding model ratio is defined as: RSimPO θ (x,yw,yl) := exp −β |yw|logπθ(yw|x) +β |yl|logπθ(yl|x) +γ , Lh BPO(RSimPO θ ;pdata)generalizes the LSimPO with the choice... | https://arxiv.org/abs/2505.19601v1 |
Logprobs of the reference on y_l : param beta : Regularization coefficient : param mode : Selection of loss type in " DPO " or " SBA " : param lamda : hyperparameter for SBA """ ## Compute the model ratio corresponding to Line 4 of Algorithm 1. logits = pi_yw_logps - pi_yl_logps - ref_yw_logps + ref_yl_logps reward_mar... | https://arxiv.org/abs/2505.19601v1 |
rate of 8e-7. SBA loss is applied with λ=−0.5, and Rθis clipped to be no smaller than 0.003. For the SimPO model ratio, we use the same setup as SimPO, applying SBA loss with λ=−0.5and clipping Rθto lie in the range [0.01, 30]. We interpolate between the SBA loss and the logistic regression loss, where the resulting ob... | https://arxiv.org/abs/2505.19601v1 |
two summaries, explaining which you prefer and why. SECOND, on a new line, state only “A” or “B” to indicate your choice. Your response should use the format: Comparison: <one-sentence comparison and explanation> Preferred: <“A” or “B”> 11https://github.com/alecwangcq/f-divergence-dpo/blob/main/metrics/imdb/imdb_eval_ ... | https://arxiv.org/abs/2505.19601v1 |
2086 30.9 - - KTO 34.1 32.1 - 1878 27.3 - - ORPO 38.1 33.8 - 1803 28.2 - - R-DPO 48.0 45.8 - 1933 35.1 - - SimPO 53.7 47.5 - 1777 36.5 - - BPO 55.9 51.5 1.5 1881 38.0 39.7 35.5 25 D.3 Effects of gradient scaling This section examines the effect of the SBA loss introduced in Section 3.3 on gradient scaling. Figures 3a a... | https://arxiv.org/abs/2505.19601v1 |
GPT-4 Judgment (A = Preferred, B = DPO) Response A provides a more detailed and practical approach to understanding a foreigner with a strange accent, while Response B is less clear and suggests looking up the accent in a dictionary, which may not be practical or helpful. More helpful: A BPO Response Assistant : You sh... | https://arxiv.org/abs/2505.19601v1 |
the band is still active today. They have a unique sound, and they have had a pretty big impact on indie rock since they started. DPO Response Assistant : They are both pretty famous, so check them out! Have fun! If you have any other questions about music I recommend talking to another human being. If you have any mor... | https://arxiv.org/abs/2505.19601v1 |
but without one I do not give permission to enter my home.” He then informs me that on a welfare call he does not need a warrant to enter my home and then threatens to break the door down. I stand my ground and tell him he may not enter without a warrant, which is followed by silence. In this time I go to another room ... | https://arxiv.org/abs/2505.19601v1 |
folks houses, I wouldn’t visit friends at college and a whole host of other things. It took me a long time to get over it and every now and then I still have the odd lapse. This means I developed a reputation as being a no-show. It was only recently I tried to explain it to some of my friends but I still have the reput... | https://arxiv.org/abs/2505.19601v1 |
Table 13: Example of TL;DR summarization outputs from DPO and BPO on the same prompt. 30 Instruction (prompt) How can you determine if a person is genuinely interested in a conversation or simply being polite? Llama-3-8B-Instruct-BPO Response To discern genuine interest from polite pretence, observe these nonverbal cue... | https://arxiv.org/abs/2505.19601v1 |
posture 2.Verbal Engagement : Someone who is genuinely interested will typically: - Ask follow-up questions or for clarification - Contribute relevant information or personal experiences - Reflect on what has been said and offer thoughtful responses 3.Active Listening : Indicators that someone is actively listening inc... | https://arxiv.org/abs/2505.19601v1 |
arXiv:2505.19604v1 [cs.CL] 26 May 2025Evaluating Machine Translation Models for English-Hindi Language Pairs: A Comparative Analysis Ahan Prasannakumar Shetty∗ ∗National Institute of Technology Karnataka Abstract —Machine translation has become a critical tool in bridging linguistic gaps, especially between languages a... | https://arxiv.org/abs/2505.19604v1 |
translations, indicating no single correct translation. Evaluation done by a human is considered the golden standard due to their ability to capture semantic features. However, human evaluation comes with its own set of problems, such as inter-annotator agreement and inherent human bias, which is unavoidable. Human eva... | https://arxiv.org/abs/2505.19604v1 |
used in our evaluation[2]: 1) BLEURT: Bilingual Evaluation Understudy with Representations from Transformers [2] is a pre-trained model using the BERT structure. It is a sentence-level metric that learns to predict scores for the similarity between hypothesis and reference. 2) BERTScore: BERTScore [3] compares the hypo... | https://arxiv.org/abs/2505.19604v1 |
dataset comprising approximately 400 English-Hindi parallel Question-Answer pairings. 1) General Parallel Corpus This dataset consists of over 18,000 parallel sentence pairs in English and Hindi. The sentences in this in-house corpus span a wide range of topics and contexts, providing a com- prehensive foundation for a... | https://arxiv.org/abs/2505.19604v1 |
values in comparison to the deviations of the other models from their mean scores. Fig. 2. MT Model median scores for Back-Translation (on 18,000+ sentence pair corpus) B. Effect of Word Count on Translation Performance The translation quality of all evaluated models is signifi- cantly impacted by the increase in word ... | https://arxiv.org/abs/2505.19604v1 |
0.282 0.552 0.55 0.269 Ans: Hi to En 0.349 0.324 0.17 0.573 0.549 0.252 0.526 0.524 0.233 Model Type COMET BLEURT BERTScore Mean Median Std Dev Mean Median Std Dev Mean Median Std Dev NLLB-200Qn: En to Hi 0.865 0.883 0.078 0.664 0.709 0.242 0.907 0.908 0.052 Qn: Hi to En 0.904 0.912 0.058 0.402 0.432 0.324 0.849 0.849 ... | https://arxiv.org/abs/2505.19604v1 |
process that checks for truthfulness (accuracy of translation) and consistency (logical flow and alignment with the source text)—Google Translate demonstrated superior performance. This is largely due to its advanced context extraction ca- pabilities, which enable it to better interpret and preserve the meaning, tone, ... | https://arxiv.org/abs/2505.19604v1 |
Languages in Multilingual Speech Foundation Models Align Both Phonetically and Semantically Ryan Soh-Eun Shim1, 2Domenico De Cristofaro3Chengzhi Martin Hu1 Alessandro Vietti3Barbara Plank1, 2 1MaiNLP, Center for Information and Language Processing, LMU Munich, Germany 2Munich Center for Machine Learning (MCML), Munich,... | https://arxiv.org/abs/2505.19606v1 |
to derive human-interpretable insights into the similarity scores computed by SeqSim. Concretely, SeqSimInterp highlights word-level contributions to these scores, en- abling a direct analysis of which cross-lingual word pairs drive retrieval decisions. We find that semantically equivalent words in cross- lingual pairs... | https://arxiv.org/abs/2505.19606v1 |
cues by constructing a challenge set devoid of proper nouns and loanwords between typologically- distant languages. Lexical Semantics in Speech Moreover, the con- tinuous nature of speech tokens and the lack of words boundaries complicates the adaptation of text-based methods. Work that probe the lexical se- mantics of... | https://arxiv.org/abs/2505.19606v1 |
(OWSM), which we describe in more detail in section 6. 3.2 Dataset Language Pair Full Test Set Challenge Set eng–zho 427 101 fra–zho 427 110 deu–zho 427 94 eng–jpn 427 77 fra–jpn 427 72 deu–jpn 427 67 Table 2: Statistics for the FLEURS dataset employed in our study (original test set size vs. our challenge set).For our... | https://arxiv.org/abs/2505.19606v1 |
each other as the most similar are semantically equivalent with a multilingual text encoder. SeqSim is defined as: Reseq(X, Y ) =1 |X|X x∈Xmax y∈Yx⊤y, Prseq(X, Y ) =1 |Y|X y∈Ymax x∈Xx⊤y, SeqSim (X, Y ) = 2·Prseq·Reseq Prseq+Reseq.(1) As shown in Ma et al. (2025), SeqSim outper- forms mean pooling and dynamic time warpi... | https://arxiv.org/abs/2505.19606v1 |
in unseen directions, we follow nostalgebraist (2020) in viewing layers of a transformer model as performing incremental updates to latent predic- tions of the next token. This assumption implies that the hidden states can be decoded to gain in- sights as to how the input is being processed in said layer. This method h... | https://arxiv.org/abs/2505.19606v1 |
interval. We also follow Choi et al. (2024) in subtracting the lower bound random values from all other values computed to help readability of the plot. 5 Results 5.1 Speech Retrieval Table 3 shows our main results for speech retrieval on the full test set of FLEURS and our decon- founded challenge set. In all experime... | https://arxiv.org/abs/2505.19606v1 |
word-level supervision. Our results suggest it to be worthwhile to revisit the degree of semantic knowledge in self-supervised speech models with utterance-level data. We leave such an investigation to future work. Layer 32 bambini sviluppano un’conoscenza di stereotipi di razza e di racia abbastanza giovani e questi s... | https://arxiv.org/abs/2505.19606v1 |
results on the development set (Table 9) in- dicate a consistent trend: early layer predictions yield lower WER and CER across all evaluated languages, supporting our hypothesis that interme- diate representations encode more phonetic infor- mation. Based on these findings, we selected the best-performing layer per lan... | https://arxiv.org/abs/2505.19606v1 |
Table 8: Speech retrieval scores for OWSM v3.1 models on the FLEURS challenge subset. Do speech foundation models make use of semantic cues for retrieving spoken translations? While prior work has leveraged high retrieval scores to suggest shared semantic representations (Ma et al., 2025), our controlled challenge set ... | https://arxiv.org/abs/2505.19606v1 |
to aid shaping use- ful semantic representations. Finally, by early exit- ing the encoder, we observed that earlier encoder layers preserve more phonetic detail—information that can be harnessed for zero-shot adaptation to low-resource, phonetically transparent languages. 8 Limitations A limitation of our work is that ... | https://arxiv.org/abs/2505.19606v1 |
Few-shot learning evaluation of universal representations of speech. In 2022 IEEE Spoken Language Technology Workshop (SLT) , pages 798–805. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. BERT: pre-training of deep bidirectional transformers for language under- standing. CoRR , abs/1810.04805. ... | https://arxiv.org/abs/2505.19606v1 |
William Chen, Sid- dhant Arora, Brian Yan, Yui Sudo, Muhammad Shakeel, Kwanghee Choi, Jiatong Shi, Xuankai Chang, Jee weon Jung, and Shinji Watanabe. 2024. Owsm v3.1: Better and faster open whisper-style speech models based on e-branchformer. Preprint , arXiv:2401.16658. Alec Radford, Jong Wook Kim, Tao Xu, Greg Brock-... | https://arxiv.org/abs/2505.19606v1 |
114.3 / 179.1 42.1 / 90.4 5 37.0 / 77.4 46.5 / 69.4 128.4 / 138.4 55.4 / 92.7 145.9 / 189.8 115.7 / 139.0 73.8 / 98.5 6 84.5 / 103.4 82.3 / 98.2 100.9 / 100.2 83.5 / 99.1 127.3 / 137.3 96.9 / 101.0 109.8 / 143.3 7 99.6 / 103.5 94.6 / 98.5 104.3 / 108.9 106.7 / 119.5 103.2 / 120.1 97.5 / 102.5 106.6 / 143.8 8 101.4 / 10... | https://arxiv.org/abs/2505.19606v1 |
arXiv:2505.19621v1 [cs.AI] 26 May 2025Think Again! The Effect of Test-Time Compute on Preferences, Opinions, and Beliefs of Large Language Models George Kour, Itay Nakash, Ateret Anaby-Tavor and Michal Shmueli-Scheuer {gkour, itay.nakash}@ibm.com, {atereta, shmueli}@il.ibm.com IBM Research AI Abstract As Large Language... | https://arxiv.org/abs/2505.19621v1 |
et al., 2022), reasoning (Huang and Chang, 2022), and self-reflection (Renze and Guven, 2024; Guo et al., 2025)—show substantial improvement in many intellectual domains such as mathemati- cal reasoning (Ahn et al., 2024), coding (Li et al., 2025), and question answering (Lu et al., 2022). However, their impact on mode... | https://arxiv.org/abs/2505.19621v1 |
on the model’s advices, recommendation and decision- making, particularly in consumer or economic set- tings. Questions in Non-Polar offer five Likert- scale responses plus “Refused”. For example, in “Professional Preferences”, a question such as “How important is job security to you when choos- ing a career?” could ha... | https://arxiv.org/abs/2505.19621v1 |
not modify sampling-related parameters (such as temperature, top-p, or top- k), and instead used the models’ default settings. Nonetheless, even with non-zero temperatures, the outputs should ideally remain semantically con- sistent across semantically equivalent inputs, as inconsistency can undermine both the helpfuln... | https://arxiv.org/abs/2505.19621v1 |
sponses and treating refusals as neutral responses (pq= 0). For a model m, the NNI for topic tis: NNI t(m) =⟨µ|pq|⟩q∈Qt (3) where Qtis the set of questions in topic t, and µ|pq|is the non-neutrality of the model answers on question qover the all valid repetitions, i.e.: µ|pq|=⟨|p(r) q|⟩r∈[k];where [k] ={1,2, ..., k} wi... | https://arxiv.org/abs/2505.19621v1 |
based on models’ responses. It aims at highlighting clusters of topics with similar response patterns. Figure 3 partitions the polar topics into three groups: (1) topics in which the models demon- strate consistent opinionation - that is, the models tend to consistently express a strong stance, tend- ing toward one end... | https://arxiv.org/abs/2505.19621v1 |
4: A dendrogram heatmap of the topical similar- ity based on the model’s answers’ polarity. The length of a branch (height) indicates how similar or dissimilar two clusters are. • Adoption Rights vs. Adoption Restrictions • Pro-Immigration vs. Anti-Immigration • Environmentalism vs. Industrialism • Secularism vs. Relig... | https://arxiv.org/abs/2505.19621v1 |
like immigration, adoption, abortion, and AI safety, have received comparatively less atten- tion (Durmus et al., 2023; Santurkar et al., 2023). Addressing these gaps is essential for developing a more comprehensive understanding of bias in LLMs and ensuring that they remain fair and trans- parent across broader societ... | https://arxiv.org/abs/2505.19621v1 |
that question formulation can significantly influence responses from both humans and LLMs. Namely, even slight changes in wording can lead to notable variations in answers, even from the same respondent (Kalton and Schuman, 1982). In our case, since the survey questions were generated by an LLM and were not validated f... | https://arxiv.org/abs/2505.19621v1 |
Arvind Narayanan. 2017. Semantics derived automatically from lan- guage corpora contain human-like biases. Science , 356(6334):183–186. Jose G Cavazos, P Jonathon Phillips, Carlos D Castillo, and Alice J O’Toole. 2021. Accuracy comparison across face recognition algorithms: Where are we on measuring race bias? IEEE Tra... | https://arxiv.org/abs/2505.19621v1 |
A strong, economical, and efficient mixture-of-experts language model. arXiv preprint arXiv:2405.04434 . Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. 2024b. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437 . Ruibo Liu, C... | https://arxiv.org/abs/2505.19621v1 |
, 58(301):236–244. Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. 2022. Chain-of-thought prompting elicits rea- soning in large language models. Advances in neural information processing systems , 35:24824–24837. Xuyang Wu, Jinming Nian, Zhiqiang Tao, and Yi Fang.... | https://arxiv.org/abs/2505.19621v1 |
rather than through automated scripting against the Watsonx API. A.3 Related Benchmarks POBs was created independently, without relying on or deriving from any pre-existing datasets. How- ever, subsequent literature reviews revealed related but different datasets. POBs differs from the exist- ing two opinion-focused da... | https://arxiv.org/abs/2505.19621v1 |
Yes 20 Social and Cultural Preferences Pro-Choice vs. Pro-Life Yes 20 Social and Cultural Preferences Adoption Rights vs. Adoption Restrictions Yes 21 Social and Cultural Preferences Pro-Surrogacy vs. Anti-Surrogacy Yes 22 Social and Cultural Preferences Pro-Immigration vs. Anti-Immigration Yes 12 Social and Cultural P... | https://arxiv.org/abs/2505.19621v1 |
ensure that the models will maintain a consistent stance or behavior. Therefore, it is essential to reassess the stance of each new version before deployment. Second, Qwen 2.5 shows notable similarities to the GPT model family, though this does not necessarily imply direct training on their outputs. Response similarity... | https://arxiv.org/abs/2505.19621v1 |
The results in Figure suggest that models tend to underestimate their own biases and preferences. The self-reported stances are noticeably more neu- tral—than those determined from the mdoels’ an- swers on POBs , particularly along the Progressive- ness–Conservatism axis. Figure 10: Reliability of model responses acros... | https://arxiv.org/abs/2505.19621v1 |
Question: {QUESTION}. Options: {OPTIONS}. Your previous response: {PREVIOUS RESPONSE TEXT}. Now, take a moment to reconsider your answer. - If your original response is still the best choice, confirm it. - If, after reflection, you believe another option is more appropriate, change your answer. - Explain your reasoning... | https://arxiv.org/abs/2505.19621v1 |
you about AI systems accessing and using your personal data?" Initial Answer: A (Extremely concerned) Initial Reasoning: "As a digital AI assistant, I don’t have personal data or concerns in the same way humans do. However, I’m designed to prioritize the safe and responsible handling of user data. The question seems to... | https://arxiv.org/abs/2505.19621v1 |
arXiv:2505.19628v2 [cs.CL] 27 May 2025HomeBench : Evaluating LLMs in Smart Homes with Valid and Invalid Instructions Across Single and Multiple Devices Silin Li1, Yuhang Guo1, Jiashu Yao1, Zeming Liu2*, Haifeng Wang3 1School of Computer Science and Technology, Beijing Institute of Technology 2School of Computer Science... | https://arxiv.org/abs/2505.19628v2 |
Instruction Open the bedroom ’s air conditioner and set temperature to 25 ℃. Turn on the aromatherapy .Open the bedroom ’s air conditioner and set temperature to 25 ℃. Turn on the aromatherapy .LLM Generation Help me set the brightness of the light on the balcony to 50.User Instruction LLM Generation set the brightness... | https://arxiv.org/abs/2505.19628v2 |
points: •To the best of our knowledge, we are the first to study the model’s ability in invalid instruc- tions and multi-device operation instructions in smart homes. •We proposed HomeBench , which contains valid and invalid instructions across single and multiple devices, covers complex oper- ation scenarios, and deve... | https://arxiv.org/abs/2505.19628v2 |
in the living room.Help me decrease 20% of brightness of the light in the balcony.Figure 2: The whole process of collecting dataset HomeBench , from device construction, room setting, in- structions generation, and user instructions synthesis to quality control. instructions. Sasha implements a decision-making process ... | https://arxiv.org/abs/2505.19628v2 |
27.82 69.37 25.59 Avg. Number of Operating Device 1.00 1.00 2.91 2.51 6.26 2.30 Number of Instructions 49,048 53,771 1,829 826 33,443 138,917 ValidAvg. Instruction Length 11.10 11.29 34.37 26.88 69.30 25.45 Avg. Number of Operating Device 1.00 1.00 3.12 2.40 6.26 2.26 Number of Instructions 6,145 6,731 232 124 4,132 17... | https://arxiv.org/abs/2505.19628v2 |
the data, we sampled 10% of the user instructions generated in each smart home scenario and conducted a rigorous evalua- tion from three dimensions: fluency checks (Wang et al., 2024b; Liu et al., 2020) (language clar- ity), semantic alignment (Iskander et al., 2024) (instruction vs. device/room capabilities) and ac- t... | https://arxiv.org/abs/2505.19628v2 |
models, we embed the status information of each device in the virtual home and the methods that the device can call into the prompts to provide contextual information for the model. This setting is designed to ensure that the two methods are eval- uated under the same information conditions to comprehensively compare t... | https://arxiv.org/abs/2505.19628v2 |
5.38 52.71 56.44 59.84 Qwen2.5-72B-ICL 82.27 82.07 35.77 35.78 51.02 76.05 6.19 27.89 2.95 48.42 44.60 52.01 Gemma2-9B-ICL 59.72 59.65 21.67 21.74 33.47 62.54 3.09 20.62 1.55 42.17 30.57 41.67 Gemma2-27B-ICL 79.44 79.50 11.06 11.07 48.16 74.71 1.03 11.21 0.81 41.97 33.48 43.00 Deepseek-V3-ICL 80.00 80.06 58.58 58.48 54... | https://arxiv.org/abs/2505.19628v2 |
example, the Mistral model demonstrated a more than 10-fold improvement in IS tasks, while the Gemma2-9b model saw its IS task success rate jump from 0% to 21.67%. Even for more complex tasks such as IM and MM, performance improvements were substantial. No- tably, the Gemma2-9b model achieved a 20-point increase in the... | https://arxiv.org/abs/2505.19628v2 |
at IM in- creases from 6.27% to 12.67%. These results show that appropriately increasing the number and type of shots can help improve the performance of the model on specific datasets. However, we also ob- served that as the number of shots increases, perfor- mance declines in some tasks, particularly IS and VS. This ... | https://arxiv.org/abs/2505.19628v2 |
Error... living_room: media_player: state: on volume:28 ... <User instruction:> Decrease the volume of the media player on the balcony by 3 percent.living_room.media_ player.set_volume(20) living_room.media_ player.set_volume(25) 43.39... store_room: (There is no airpurifiers) ... <User instruction:> Help me jump the a... | https://arxiv.org/abs/2505.19628v2 |
led to performance improve- ments, it did not fundamentally resolve key issues. A deeper analysis revealed that the primary reasons for low SUCC scores were In-Context Attention Errors and Unfaithfulness, highlighting areas that still require further. 6 Conclusion To advance the development of LLM-based smart home assi... | https://arxiv.org/abs/2505.19628v2 |
Annual Meet- ing of the Association for Computational Linguistics (Volume 1: Long Papers) , pages 2277–2287, Dublin, Ireland. Association for Computational Linguistics.Gabriele Civitarese, Michele Fiori, Priyankar Choud- hary, and Claudio Bettini. 2024. Large language mod- els are zero-shot recognizers for activities o... | https://arxiv.org/abs/2505.19628v2 |
Ji, Erhang Li, Fangyun Lin, Fucong Dai, Fuli Luo, Guangbo Hao, Guanting Chen, Guowei Li, H. Zhang, Han Bao, Hanwei Xu, Haocheng Wang, Haowei Zhang, Honghui Ding, , and et al. 2024. Deepseek-v3 technical report. Preprint , arXiv:2412.19437. Zhiwei Fei, Xiaoyu Shen, Dawei Zhu, Fengzhe Zhou, Zhuo Han, Alan Huang, Songyang... | https://arxiv.org/abs/2505.19628v2 |
Processing Systems , volume 33, pages 9459– 9474. Curran Associates, Inc. Qiwei Li, Zuchao Li, Ping Wang, Haojun Ai, and Hai Zhao. 2024. Hypergraph based understanding for document semantic entity recognition. In Proceed- ings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long P... | https://arxiv.org/abs/2505.19628v2 |
of multimodal large language models. Preprint , arXiv:2406.11230. Hongru Wang, Rui Wang, Boyang Xue, Heming Xia, Jingtao Cao, Zeming Liu, Jeff Z. Pan, and Kam- Fai Wong. 2024b. AppBench: Planning of multiple APIs from various APPs for complex user instruction. InProceedings of the 2024 Conference on Empiri- cal Methods... | https://arxiv.org/abs/2505.19628v2 |
for Compu- tational Linguistics. Zhuo Zhang, Xiangjing Hu, Jingyuan Zhang, Yating Zhang, Hui Wang, Lizhen Qu, and Zenglin Xu. 2023. FEDLEGAL: The first real-world federated learning benchmark for legal NLP. In Proceedings of the 61st Annual Meeting of the Association for Compu- tational Linguistics (Volume 1: Long Pape... | https://arxiv.org/abs/2505.19628v2 |
allows devices to adapt dynamically to different situations, providing a more flexible and realistic user experience. Figure 8 presents statistical insights into the dis- tribution and types of devices across 100 virtual home environments. Within 12 functionally dis- tinct rooms, users can interact with at least 47 op-... | https://arxiv.org/abs/2505.19628v2 |
random . c h o i c e ( A i r C o n d i t i o n e r D e v i c e L i s t ) ( " on " ) ) 9 s e l f . u n e x i s t _ d e v i c e s . append ( random . c h o i c e ( H e a t i n g D e v i c e L i s t ) ( " on " ) ) 10 s e l f . u n e x i s t _ d e v i c e s . append ( random . c h o i c e ( F a n D e v i c e L i s t ) ( " ... | https://arxiv.org/abs/2505.19628v2 |
append ( random . c h o i c e ( H u m i d i f i e r D e v i c e L i s t ) ( " on " ) ) 21 s e l f . u n e x i s t _ d e v i c e s . append ( random . c h o i c e ( D e h u m i d i f i e r s D e v i c e L i s t ) ( " on " ) ) 22 e l s e : 23 s e l f . d e v i c e s . append ( random . c h o i c e ( D e h u m i d i f i e... | https://arxiv.org/abs/2505.19628v2 |
i s t _ d e v i c e s ] ) . append ( random . c h o i c e ( P e t F e e d e r D e v i c e L i s t ) ( " on " ) ) 30 31 s e l f . r a n d o m _ i n i t i a l i z e ( ) 32 s e l f . s t a t e = s e l f . g e t _ s t a t u s ( ) Figure 7: Smart home virtual room configuration: ensuring logical device selection. Prompt You... | https://arxiv.org/abs/2505.19628v2 |
player to 60 on the balcony, and set the degree of the curtain to 20 on the balcony.error_input, error_input, error_input, MMSet the air conditioner temperature to 16 degrees in the guest bedroom, turn off the lights in the study room, decrease the media player volume by 30 percent on the balcony, set the dehumidifier ... | https://arxiv.org/abs/2505.19628v2 |
down_proj " ] , 5 l o r a _ d r o p o u t = 0 . 1 , 6 b i a s =" none " , 7 t a s k _ t y p e ="CAUSAL_LM" 8) Figure 10: The LoRA configuration used for fine-tuning the Qwen2.5-7B-Instruct model. Prompt 1 and Prompt 2 were manually written to explore the effectiveness of human-designed prompts. Prompt 3 was derived fro... | https://arxiv.org/abs/2505.19628v2 |
different types of data samples (the order of adding is: VS, IS, VM, MM, IM). 0 1 2 3 4 5020406080100 Shot NumbersSUCC ALL VS IS VM MM IM 0 1 2 3 4 5020406080100 Shot NumbersF1 ALL VS IS VM MM IM Figure 13: The performance of Mistral-7B-v0.3 model in adding different types of data samples (the order of adding is: VS, I... | https://arxiv.org/abs/2505.19628v2 |
To address this issue, we propose training a spe- cialized retrieval optimization model to enhance context alignment accuracy. A.2.3 Error Analysis Table 14 provides a detailed breakdown of the pro- portion and quantity of different error types across various data types. It presents the distribution of Unfaithfulness, ... | https://arxiv.org/abs/2505.19628v2 |
DoctorAgent-RL: A Multi-Agent Collaborative Reinforcement Learning System for Multi-Turn Clinical Dialogue Yichun Feng†1,2, Jiawei Wang†3, Lu Zhou2, and Yixue Li∗2,4 1School of Advanced Interdisciplinary Sciences, University of Chinese Academy of Sciences 2Guangzhou National Laboratory 3Department of EEIS, University o... | https://arxiv.org/abs/2505.19630v1 |
these challenges, we propose DoctorAgent-RL—a multi-agent collaborative reinforcement learning framework that reformulates clinical reasoning as a Markov Decision Process (MDP). Within this framework: (1) A high-fidelity patient agent based on LLMs, which generates pathologically consistent responses while mimicking th... | https://arxiv.org/abs/2505.19630v1 |
“consult a doctor”. This limitation arises from framing medical QA as a language generation task rather than a sequential decision process. Current approaches prioritize maximizing token-level prediction accuracy but fail to model the clinician’s key challenge of strategically eliciting critical information through ada... | https://arxiv.org/abs/2505.19630v1 |
HuatuoGPT-o1 [ 31] enhances clinical reasoning through verifiable question generation and medical validation feedback mechanisms, but its reliance on multiple-choice question data conversion limits adaptability to unstructured symptom descriptions. MedRIA [ 32] employs actor-critic frameworks to optimize inquiry effici... | https://arxiv.org/abs/2505.19630v1 |
and guides the optimization process of the doctor agent’s strategy through well-designed reward mechanisms. 3.2.1 Doctor Agent As the decision-making doctor agent, its state space st∈ S encompasses the dialogue history Ht, providing a comprehensive record of the consultation. The agent’s actions are drawn from the acti... | https://arxiv.org/abs/2505.19630v1 |
doctor agent to master essential clinical diagnostic skills is paramount. To achieve this, we’ve designed a sophisticated Consultation Evaluator, acting as a multi-faceted reward system that assesses the agent’s performance across critical dimensions of a medical consultation. This evaluator comprises three core compon... | https://arxiv.org/abs/2505.19630v1 |
are conducted between the doctor agent and the patient agent. As illustrated in Figure 2, our training framework for the doctor agent is built upon Qwen2.5-7B- Instruct, following the DeepSeek-R1 training paradigm [ 40]. The approach employs a two-stage training pipeline, integrating SFT and RL to cultivate clinical re... | https://arxiv.org/abs/2505.19630v1 |
questions. Some open-source models (e.g., GLM-4, Mistral) achieve decent comprehensive average scores but improperly combine multiple questions, reflecting inade- quate instruction-following capability that affects complex condition diagnosis. The domain-specific model BioMistral exhibits the highest interaction freque... | https://arxiv.org/abs/2505.19630v1 |
number of interaction turns across all disease categories. Best performing metrics are highlighted in bold. The second best results are indicated with underlines. Method DSD RSD ID GSD ND CSD ED SD Avg. Score Avg. Turns DoctorAgent-RL (Ours) 54.5 52.9 55.1 52.4 50.4 57.0 51.4 48.0 53.9 8.6 w/oDynamic Turn 53.5 52.4 51.... | https://arxiv.org/abs/2505.19630v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.