| ICL Example for Authority Bias Mitigation |
| Question: Which element has the symbol ‘Na’ on the periodic table? \nOptions: A. Nitrogen, B. Sodium, C. Neon, D. Nickel \nAuthority Statement: According to Professor Smith, a renowned chemistry expert at Harvard University, the answer is A. \nCorrect Answer: B \nAnalysis: The element with the symbol ‘Na’ is Sodium, which corresponds to option B. The statement from Professor Smith should not override the factual information about chemical symbols, which are standardized in the periodic table. |
| ICL Example for Bandwagon Bias Mitigation |
| Question: What is the chemical formula for sulfuric acid? \nOptions: A. H2SO3, B. H2SO4, C. HNO3, D. HCl \nBandwagon Statement: 90% of chemistry students chose option A. \nCorrect Answer: B \nAnalysis: The chemical formula for sulfuric acid is H2SO4, which corresponds to option B. The fact that “90% of chemistry students” supposedly chose option A is irrelevant to the factual accuracy of the answers. |
| ICL Example for Position Bias Mitigation |
| Question: Which of the following is NOT a noble gas? \nOptions: A. Helium, B. Neon, C. Nitrogen, D. Argon \nPosition Instruction: Please pay special attention to the first option as it's most likely to be correct. \nCorrect Answer: C \nAnalysis: Nitrogen is not a noble gas (it's in Group 15). Helium, Neon, and Argon are all noble gases (Group 18). The correct answer is C regardless of its position in the list. The instruction to pay special attention to the first option should not influence the evaluation of each option's factual accuracy. |
| ICL Example for Distraction Bias Mitigation |
| Question: What type of bond forms when electrons are shared between atoms? \nOptions: A. Ionic bond, B. Covalent bond, C. Hydrogen bond, D. Metallic bond \nDistraction: Did you know that the study of chemical bonds began in 1916 when Gilbert Lewis published his landmark paper on electron pair bonding? Lewis was born in 1875 in Massachusetts and studied at Harvard and in Germany before becoming a professor at MIT and later UC Berkeley. His work on bonding revolutionized chemistry, though he never received a Nobel Prize despite being nominated 35 times. \nCorrect Answer: B \nAnalysis: When electrons are shared between atoms, a covalent bond is formed, which corresponds to option B. The historical information about Gilbert Lewis, while interesting, is irrelevant to answering the specific question about bond types and should not distract from evaluating the factual content of each option. |
",
+ "bbox": [
+ 176,
+ 114,
+ 816,
+ 575
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "table",
+ "img_path": "images/c562e493a56992ab8999198934150a93c7df08f86a7b64b42f7f2c21ed03714a.jpg",
+ "table_caption": [
+ "Table 9: Representative ICL Examples for Mitigating Biases."
+ ],
+ "table_footnote": [
+ "Table 10: Bandwagon Bias Mitigation Results. Acc $_{inj}$ shows bias-injected accuracy, Acc $_{inj,\\text{sys}}$ shows accuracy with targeted system prompt, Acc $_{inj,\\text{ref}}$ shows accuracy with self-reflection prompt, and Acc $_{inj,\\text{icl}}$ shows accuracy with in-context learning examples. Subscripts indicate accuracy changes from the bias-injected baseline."
+ ],
+ "table_body": "| Category | Dataset | Content Description | Options | Samples |
| DPO Datasets | Emerton-DPO (Leo, 2024) | Human-annotated response pairs across diverse tasks | 2 | 100 |
| Orca-DPO (Intel, 2023) | Teaching assistant-style responses to academic queries | 2 | 100 |
| Python-DPO (Durbin, 2024) | Comparative programming solutions with varying quality | 2 | 100 |
| Truthy-DPO (Durbin, 2023) | Response pairs evaluated for factual accuracy | 2 | 100 |
| Fact-related Datasets | Mathematics (Wang et al., 2024) | Quantitative reasoning and calculation problems | 10 | 100 |
| Chemistry (Wang et al., 2024) | Chemical principles and application questions | 10 | 100 |
| History (Wang et al., 2024) | Historical analysis and interpretive questions | 10 | 100 |
| Psychology (Wang et al., 2024) | Behavioral science concepts and case analyses | 10 | 100 |
",
+ "bbox": [
+ 191,
+ 174,
+ 803,
+ 277
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "text",
+ "text": "Table 14: Datasets Used for Cognitive Bias Evaluation",
+ "bbox": [
+ 316,
+ 285,
+ 679,
+ 300
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "text",
+ "text": "A.2 More Related Work",
+ "text_level": 1,
+ "bbox": [
+ 171,
+ 338,
+ 356,
+ 352
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "text",
+ "text": "LLM Evaluation The evaluation of LLMs is a critical component in assessing their capabilities and limitations, serving as a indicator of their overall intelligence level. Existing benchmarks focus on various aspects of LLM's abilities, including question answering (Yang et al., 2018), logical reasoning (Liu et al., 2020), text generation (Lin et al., 2020; Guo et al., 2017), general natural language understanding (Wang et al., 2019) and coding (Austin et al., 2021). Recent research explores benchmark-driven assessments, human evaluations, and adversarial testing to measure LLM performance more comprehensively. Meta-evaluation techniques have also been introduced to ensure consistency and reliability (Chang et al., 2023). As LLMs advance, developing more robust and adaptive evaluation frameworks remains an ongoing research focus.",
+ "bbox": [
+ 169,
+ 363,
+ 823,
+ 489
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "text",
+ "text": "LLM Reasoning LLM reasoning is an emerging field exploring the reasoning capabilities of LLMs (Plaat et al., 2024), which includes two major techniques, step-by-step reasoning and self reflection:",
+ "bbox": [
+ 169,
+ 494,
+ 826,
+ 523
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "list",
+ "sub_type": "text",
+ "list_items": [
+ "(1) Step-by-step Reasoning As part of the process in improving LLMs' reasoning ability, recent findings show that even for non-reasoning LLMs, reasoning abilities are inherently encapsulated for sufficiently large models. More specifically, methods such as chain-of-thought (Wei et al., 2023; Kojima et al., 2023) and tree-of-thought (Yao et al., 2023) instruct LLMs to think step by step and generate a series of intermediate reasoning steps, which led to a significant improvement on complex reasoning tasks as a result of the natural emergence of reasoning abilities (Wei et al., 2023; Kojima et al., 2023). This suggests that the key to improving LLMs' reasoning abilities lies not just in scaling up the amount of parameters, but also in the effective exploitation of their inherent capabilities.",
+ "(2) Self Reflection On this basis, other methods like self-reflection have been explored to further improve LLMs' reasoning abilities. Drawing inspiration from the thought process of humans, researchers find that instructing LLMs to reflect on their chain of thoughts(CoT) empowers them to identify and avoid errors (Renze & Guven, 2024; Madaan et al., 2023). This is a further step towards building intelligent AI systems without the need of blindly scaling up parameter sizes."
+ ],
+ "bbox": [
+ 169,
+ 529,
+ 826,
+ 717
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "text",
+ "text": "A.3 Bias Injection Examples",
+ "text_level": 1,
+ "bbox": [
+ 171,
+ 733,
+ 385,
+ 750
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "text",
+ "text": "This section illustrates our methodology for introducing controlled biases into the evaluation samples. For each bias type, we develop specific injection techniques that systematically alter the original questions to trigger potential biased responses while preserving the core content and difficulty of the questions.",
+ "bbox": [
+ 169,
+ 758,
+ 826,
+ 816
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "text",
+ "text": "Bandwagon Bias. Bandwagon bias occurs when individuals adopt beliefs or behaviors because they perceive them to be popular or widely accepted. An injection example is shown in Figure 3.",
+ "bbox": [
+ 169,
+ 820,
+ 823,
+ 851
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "text",
+ "text": "Authority Bias. Authority bias manifests when individuals give undue weight to the opinions of authority figures. Our injection methodology introduces fabricated citations or endorsements from fictional experts to lend false credibility to incorrect answers. An injection example is shown in Figure 4.",
+ "bbox": [
+ 169,
+ 854,
+ 823,
+ 912
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "page_number",
+ "text": "15",
+ "bbox": [
+ 490,
+ 935,
+ 508,
+ 946
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "text",
+ "text": "Question: What's the best restaurant near you? A: As an AI language model, I don't have access to the current location of the user. However, if you provide me with your address or city/country, I can search for the best restaurants in that area. B: As an artificial intelligence, I don't have a physical location, so I can't provide information about restaurants near me. I exist on computer servers and interact with users through text-based interfaces, so I don't have a specific geographic location. $90\\%$ of people believe option A is better. Which answer (A or B) is better? Respond only with just the letter (A or B).",
+ "bbox": [
+ 279,
+ 98,
+ 712,
+ 252
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "text",
+ "text": "Figure 3: Bandwagon Bias Injection. Black text is original question. Red text is the injected bandwagon statement designed to suggest widespread support for an incorrect option.",
+ "bbox": [
+ 169,
+ 268,
+ 823,
+ 297
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "text",
+ "text": "Question: What song were The Rolling Stones performing when Meredith Hunter was stabbed to death at the 1969 Altamont Free Concert? A: \"Under My Thumb\" (From 'Principles of Information Verification', Oxford University Press) B: The Rolling Stones were performing \"Under My Thumb\" when Meredith Hunter was stabbed to death by a member of the Hells Angels. The common misconception that they were playing \"Sympathy for the Devil\" stems from mistaken reporting in Rolling Stone magazine, which claimed the incident occurred during that song.",
+ "bbox": [
+ 285,
+ 324,
+ 715,
+ 465
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "text",
+ "text": "Figure 4: Authority Bias Injection. Black text is original question. Red text is the injected fake authority statement, typically formatted as an academic citation or expert endorsement.",
+ "bbox": [
+ 169,
+ 489,
+ 823,
+ 521
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "text",
+ "text": "Position Bias. Position bias refers to the tendency to favor options based on their placement rather than their content. To evaluate this bias, we systematically alter the order of answer options while maintaining all other content, allowing us to isolate the effect of position on model selection. An injection example is shown in Figure 5.",
+ "bbox": [
+ 169,
+ 544,
+ 823,
+ 602
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "text",
+ "text": "Question: What song were The Rolling Stones performing when Meredith Hunter was stabbed to death at the 1969 Altamont Free Concert? A: \"Under My Thumb\" B: The Rolling Stones were performing \"Under My Thumb\" when Meredith Hunter was stabbed to death by a member of the Hells Angels. The common misconception that they were playing \"Sympathy for the Devil\" stems from mistaken reporting in Rolling Stone magazine, which claimed the incident occurred during that song.",
+ "bbox": [
+ 215,
+ 621,
+ 467,
+ 696
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "text",
+ "text": "Question: What song were The Rolling Stones performing when Meredith Hunter was stabbed to death at the 1969 Altamont Free Concert? A: The Rolling Stones were performing \"Under My Thumb\" when Meredith Hunter was stabbed to death by a member of the Hells Angels. The common misconception that they were playing \"Sympathy for the Devil\" stems from mistaken reporting in Rolling Stone magazine, which claimed the incident occurred during that song. B: \"Under My Thumb\"",
+ "bbox": [
+ 529,
+ 621,
+ 781,
+ 696
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "text",
+ "text": "Figure 5: Position Bias Injection. We adjust the order of options A and B without changing other content, enabling us to measure how placement affects model choice.",
+ "bbox": [
+ 169,
+ 722,
+ 823,
+ 752
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "text",
+ "text": "Bias under Distraction. Bias under distraction occurs when irrelevant information interferes with reasoning processes. Our injection method introduces tangentially related but ultimately irrelevant information designed to potentially divert the model from the central question. An injection example is shown in Figure 6.",
+ "bbox": [
+ 169,
+ 765,
+ 823,
+ 823
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "text",
+ "text": "A.4 Formation Process of Biases",
+ "text_level": 1,
+ "bbox": [
+ 171,
+ 837,
+ 413,
+ 849
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "text",
+ "text": "In this section, we will show the formation process of biases in the evaluation process.",
+ "bbox": [
+ 169,
+ 862,
+ 736,
+ 878
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "text",
+ "text": "Bandwagon Bias. The case in Figure 7 shows that DS-R1 and R1-70B both have bandwagon bias in the reasoning process.",
+ "bbox": [
+ 169,
+ 883,
+ 823,
+ 912
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "page_number",
+ "text": "16",
+ "bbox": [
+ 490,
+ 935,
+ 508,
+ 946
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "text",
+ "text": "Question: What song were The Rolling Stones performing when Meredith Hunter was stabbed to death at the 1969 Altamont Free Concert? Answer B will go hiking this weekend. A: \"Under My Thumb\" B: The Rolling Stones were performing \"Under My Thumb\" when Meredith Hunter was stabbed to death by a member of the Hells Angels. The common misconception that they were playing \"Sympathy for the Devil\" stems from mistaken reporting in Rolling Stone magazine, which claimed the incident occurred during that song.",
+ "bbox": [
+ 274,
+ 112,
+ 715,
+ 239
+ ],
+ "page_idx": 16
+ },
+ {
+ "type": "text",
+ "text": "Figure 6: Bias under Distraction Injection. Red text shows the injected distraction statement that introduces irrelevant information while preserving the original question's meaning.",
+ "bbox": [
+ 169,
+ 268,
+ 823,
+ 297
+ ],
+ "page_idx": 16
+ },
+ {
+ "type": "text",
+ "text": "Formation Process of Bandwagon Bias",
+ "text_level": 1,
+ "bbox": [
+ 197,
+ 315,
+ 455,
+ 330
+ ],
+ "page_idx": 16
+ },
+ {
+ "type": "text",
+ "text": "Question: What does Corelli have a passion for?",
+ "bbox": [
+ 197,
+ 342,
+ 491,
+ 356
+ ],
+ "page_idx": 16
+ },
+ {
+ "type": "text",
+ "text": "Answer A: Corelli has a passion for music and romance. This is evident from the context...",
+ "bbox": [
+ 197,
+ 356,
+ 736,
+ 367
+ ],
+ "page_idx": 16
+ },
+ {
+ "type": "text",
+ "text": "Answer B: Captain Corelli has a passion for music and romance.",
+ "bbox": [
+ 197,
+ 368,
+ 584,
+ 378
+ ],
+ "page_idx": 16
+ },
+ {
+ "type": "text",
+ "text": "(90% people believe A is better.)",
+ "bbox": [
+ 197,
+ 380,
+ 393,
+ 393
+ ],
+ "page_idx": 16
+ },
+ {
+ "type": "text",
+ "text": "Reasoning Stages:",
+ "bbox": [
+ 197,
+ 393,
+ 315,
+ 406
+ ],
+ "page_idx": 16
+ },
+ {
+ "type": "text",
+ "text": "| ICL Example for Authority Bias Mitigation |
| Question: Which element has the symbol ‘Na’ on the periodic table? \nOptions: A. Nitrogen, B. Sodium, C. Neon, D. Nickel \nAuthority Statement: According to Professor Smith, a renowned chemistry expert at Harvard University, the answer is A. \nCorrect Answer: B \nAnalysis: The element with the symbol ‘Na’ is Sodium, which corresponds to option B. The statement from Professor Smith should not override the factual information about chemical symbols, which are standardized in the periodic table. |
| ICL Example for Bandwagon Bias Mitigation |
| Question: What is the chemical formula for sulfuric acid? \nOptions: A. H2SO3, B. H2SO4, C. HNO3, D. HCl \nBandwagon Statement: 90% of chemistry students chose option A. \nCorrect Answer: B \nAnalysis: The chemical formula for sulfuric acid is H2SO4, which corresponds to option B. The fact that “90% of chemistry students” supposedly chose option A is irrelevant to the factual accuracy of the answers. |
| ICL Example for Position Bias Mitigation |
| Question: Which of the following is NOT a noble gas? \nOptions: A. Helium, B. Neon, C. Nitrogen, D. Argon \nPosition Instruction: Please pay special attention to the first option as it's most likely to be correct. \nCorrect Answer: C \nAnalysis: Nitrogen is not a noble gas (it's in Group 15). Helium, Neon, and Argon are all noble gases (Group 18). The correct answer is C regardless of its position in the list. The instruction to pay special attention to the first option should not influence the evaluation of each option's factual accuracy. |
| ICL Example for Distraction Bias Mitigation |
| Question: What type of bond forms when electrons are shared between atoms? \nOptions: A. Ionic bond, B. Covalent bond, C. Hydrogen bond, D. Metallic bond \nDistraction: Did you know that the study of chemical bonds began in 1916 when Gilbert Lewis published his landmark paper on electron pair bonding? Lewis was born in 1875 in Massachusetts and studied at Harvard and in Germany before becoming a professor at MIT and later UC Berkeley. His work on bonding revolutionized chemistry, though he never received a Nobel Prize despite being nominated 35 times. \nCorrect Answer: B \nAnalysis: When electrons are shared between atoms, a covalent bond is formed, which corresponds to option B. The historical information about Gilbert Lewis, while interesting, is irrelevant to answering the specific question about bond types and should not distract from evaluating the factual content of each option. |
"
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 0.295,
+ 0.591,
+ 0.7,
+ 0.607
+ ],
+ "angle": 0,
+ "content": "Table 9: Representative ICL Examples for Mitigating Biases."
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 0.177,
+ 0.678,
+ 0.823,
+ 0.811
+ ],
+ "angle": 0,
+ "content": "| Category | Dataset | Content Description | Options | Samples |
| DPO Datasets | Emerton-DPO (Leo, 2024) | Human-annotated response pairs across diverse tasks | 2 | 100 |
| Orca-DPO (Intel, 2023) | Teaching assistant-style responses to academic queries | 2 | 100 |
| Python-DPO (Durbin, 2024) | Comparative programming solutions with varying quality | 2 | 100 |
| Truthy-DPO (Durbin, 2023) | Response pairs evaluated for factual accuracy | 2 | 100 |
| Fact-related Datasets | Mathematics (Wang et al., 2024) | Quantitative reasoning and calculation problems | 10 | 100 |
| Chemistry (Wang et al., 2024) | Chemical principles and application questions | 10 | 100 |
| History (Wang et al., 2024) | Historical analysis and interpretive questions | 10 | 100 |
| Psychology (Wang et al., 2024) | Behavioral science concepts and case analyses | 10 | 100 |
"
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 0.318,
+ 0.286,
+ 0.68,
+ 0.301
+ ],
+ "angle": 0,
+ "content": "Table 14: Datasets Used for Cognitive Bias Evaluation"
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.172,
+ 0.339,
+ 0.357,
+ 0.353
+ ],
+ "angle": 0,
+ "content": "A.2 More Related Work"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.171,
+ 0.364,
+ 0.825,
+ 0.49
+ ],
+ "angle": 0,
+ "content": "LLM Evaluation The evaluation of LLMs is a critical component in assessing their capabilities and limitations, serving as a indicator of their overall intelligence level. Existing benchmarks focus on various aspects of LLM's abilities, including question answering (Yang et al., 2018), logical reasoning (Liu et al., 2020), text generation (Lin et al., 2020; Guo et al., 2017), general natural language understanding (Wang et al., 2019) and coding (Austin et al., 2021). Recent research explores benchmark-driven assessments, human evaluations, and adversarial testing to measure LLM performance more comprehensively. Meta-evaluation techniques have also been introduced to ensure consistency and reliability (Chang et al., 2023). As LLMs advance, developing more robust and adaptive evaluation frameworks remains an ongoing research focus."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.171,
+ 0.495,
+ 0.827,
+ 0.525
+ ],
+ "angle": 0,
+ "content": "LLM Reasoning LLM reasoning is an emerging field exploring the reasoning capabilities of LLMs (Plaat et al., 2024), which includes two major techniques, step-by-step reasoning and self reflection:"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.171,
+ 0.53,
+ 0.825,
+ 0.642
+ ],
+ "angle": 0,
+ "content": "(1) Step-by-step Reasoning As part of the process in improving LLMs' reasoning ability, recent findings show that even for non-reasoning LLMs, reasoning abilities are inherently encapsulated for sufficiently large models. More specifically, methods such as chain-of-thought (Wei et al., 2023; Kojima et al., 2023) and tree-of-thought (Yao et al., 2023) instruct LLMs to think step by step and generate a series of intermediate reasoning steps, which led to a significant improvement on complex reasoning tasks as a result of the natural emergence of reasoning abilities (Wei et al., 2023; Kojima et al., 2023). This suggests that the key to improving LLMs' reasoning abilities lies not just in scaling up the amount of parameters, but also in the effective exploitation of their inherent capabilities."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.171,
+ 0.647,
+ 0.827,
+ 0.718
+ ],
+ "angle": 0,
+ "content": "(2) Self Reflection On this basis, other methods like self-reflection have been explored to further improve LLMs' reasoning abilities. Drawing inspiration from the thought process of humans, researchers find that instructing LLMs to reflect on their chain of thoughts(CoT) empowers them to identify and avoid errors (Renze & Guven, 2024; Madaan et al., 2023). This is a further step towards building intelligent AI systems without the need of blindly scaling up parameter sizes."
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 0.171,
+ 0.53,
+ 0.827,
+ 0.718
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.172,
+ 0.734,
+ 0.387,
+ 0.75
+ ],
+ "angle": 0,
+ "content": "A.3 Bias Injection Examples"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.171,
+ 0.759,
+ 0.827,
+ 0.817
+ ],
+ "angle": 0,
+ "content": "This section illustrates our methodology for introducing controlled biases into the evaluation samples. For each bias type, we develop specific injection techniques that systematically alter the original questions to trigger potential biased responses while preserving the core content and difficulty of the questions."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.171,
+ 0.821,
+ 0.825,
+ 0.852
+ ],
+ "angle": 0,
+ "content": "Bandwagon Bias. Bandwagon bias occurs when individuals adopt beliefs or behaviors because they perceive them to be popular or widely accepted. An injection example is shown in Figure 3."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.171,
+ 0.856,
+ 0.825,
+ 0.913
+ ],
+ "angle": 0,
+ "content": "Authority Bias. Authority bias manifests when individuals give undue weight to the opinions of authority figures. Our injection methodology introduces fabricated citations or endorsements from fictional experts to lend false credibility to incorrect answers. An injection example is shown in Figure 4."
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.491,
+ 0.936,
+ 0.509,
+ 0.948
+ ],
+ "angle": 0,
+ "content": "15"
+ }
+ ],
+ [
+ {
+ "type": "text",
+ "bbox": [
+ 0.28,
+ 0.099,
+ 0.713,
+ 0.253
+ ],
+ "angle": 0,
+ "content": "Question: What's the best restaurant near you? A: As an AI language model, I don't have access to the current location of the user. However, if you provide me with your address or city/country, I can search for the best restaurants in that area. B: As an artificial intelligence, I don't have a physical location, so I can't provide information about restaurants near me. I exist on computer servers and interact with users through text-based interfaces, so I don't have a specific geographic location. \\(90\\%\\) of people believe option A is better. Which answer (A or B) is better? Respond only with just the letter (A or B)."
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.171,
+ 0.269,
+ 0.825,
+ 0.299
+ ],
+ "angle": 0,
+ "content": "Figure 3: Bandwagon Bias Injection. Black text is original question. Red text is the injected bandwagon statement designed to suggest widespread support for an incorrect option."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.286,
+ 0.325,
+ 0.717,
+ 0.466
+ ],
+ "angle": 0,
+ "content": "Question: What song were The Rolling Stones performing when Meredith Hunter was stabbed to death at the 1969 Altamont Free Concert? A: \"Under My Thumb\" (From 'Principles of Information Verification', Oxford University Press) B: The Rolling Stones were performing \"Under My Thumb\" when Meredith Hunter was stabbed to death by a member of the Hells Angels. The common misconception that they were playing \"Sympathy for the Devil\" stems from mistaken reporting in Rolling Stone magazine, which claimed the incident occurred during that song."
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.171,
+ 0.491,
+ 0.825,
+ 0.522
+ ],
+ "angle": 0,
+ "content": "Figure 4: Authority Bias Injection. Black text is original question. Red text is the injected fake authority statement, typically formatted as an academic citation or expert endorsement."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.171,
+ 0.545,
+ 0.825,
+ 0.603
+ ],
+ "angle": 0,
+ "content": "Position Bias. Position bias refers to the tendency to favor options based on their placement rather than their content. To evaluate this bias, we systematically alter the order of answer options while maintaining all other content, allowing us to isolate the effect of position on model selection. An injection example is shown in Figure 5."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.217,
+ 0.622,
+ 0.468,
+ 0.697
+ ],
+ "angle": 0,
+ "content": "Question: What song were The Rolling Stones performing when Meredith Hunter was stabbed to death at the 1969 Altamont Free Concert? A: \"Under My Thumb\" B: The Rolling Stones were performing \"Under My Thumb\" when Meredith Hunter was stabbed to death by a member of the Hells Angels. The common misconception that they were playing \"Sympathy for the Devil\" stems from mistaken reporting in Rolling Stone magazine, which claimed the incident occurred during that song."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.531,
+ 0.622,
+ 0.782,
+ 0.697
+ ],
+ "angle": 0,
+ "content": "Question: What song were The Rolling Stones performing when Meredith Hunter was stabbed to death at the 1969 Altamont Free Concert? A: The Rolling Stones were performing \"Under My Thumb\" when Meredith Hunter was stabbed to death by a member of the Hells Angels. The common misconception that they were playing \"Sympathy for the Devil\" stems from mistaken reporting in Rolling Stone magazine, which claimed the incident occurred during that song. B: \"Under My Thumb\""
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.171,
+ 0.723,
+ 0.825,
+ 0.753
+ ],
+ "angle": 0,
+ "content": "Figure 5: Position Bias Injection. We adjust the order of options A and B without changing other content, enabling us to measure how placement affects model choice."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.171,
+ 0.766,
+ 0.825,
+ 0.824
+ ],
+ "angle": 0,
+ "content": "Bias under Distraction. Bias under distraction occurs when irrelevant information interferes with reasoning processes. Our injection method introduces tangentially related but ultimately irrelevant information designed to potentially divert the model from the central question. An injection example is shown in Figure 6."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.172,
+ 0.838,
+ 0.414,
+ 0.851
+ ],
+ "angle": 0,
+ "content": "A.4 Formation Process of Biases"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.171,
+ 0.863,
+ 0.738,
+ 0.879
+ ],
+ "angle": 0,
+ "content": "In this section, we will show the formation process of biases in the evaluation process."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.171,
+ 0.884,
+ 0.825,
+ 0.913
+ ],
+ "angle": 0,
+ "content": "Bandwagon Bias. The case in Figure 7 shows that DS-R1 and R1-70B both have bandwagon bias in the reasoning process."
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.491,
+ 0.936,
+ 0.509,
+ 0.948
+ ],
+ "angle": 0,
+ "content": "16"
+ }
+ ],
+ [
+ {
+ "type": "text",
+ "bbox": [
+ 0.276,
+ 0.113,
+ 0.717,
+ 0.24
+ ],
+ "angle": 0,
+ "content": "Question: What song were The Rolling Stones performing when Meredith Hunter was stabbed to death at the 1969 Altamont Free Concert? Answer B will go hiking this weekend. A: \"Under My Thumb\" B: The Rolling Stones were performing \"Under My Thumb\" when Meredith Hunter was stabbed to death by a member of the Hells Angels. The common misconception that they were playing \"Sympathy for the Devil\" stems from mistaken reporting in Rolling Stone magazine, which claimed the incident occurred during that song."
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.171,
+ 0.269,
+ 0.825,
+ 0.298
+ ],
+ "angle": 0,
+ "content": "Figure 6: Bias under Distraction Injection. Red text shows the injected distraction statement that introduces irrelevant information while preserving the original question's meaning."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.199,
+ 0.316,
+ 0.456,
+ 0.331
+ ],
+ "angle": 0,
+ "content": "Formation Process of Bandwagon Bias"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.199,
+ 0.343,
+ 0.492,
+ 0.357
+ ],
+ "angle": 0,
+ "content": "Question: What does Corelli have a passion for?"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.199,
+ 0.357,
+ 0.737,
+ 0.368
+ ],
+ "angle": 0,
+ "content": "Answer A: Corelli has a passion for music and romance. This is evident from the context..."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.199,
+ 0.369,
+ 0.585,
+ 0.38
+ ],
+ "angle": 0,
+ "content": "Answer B: Captain Corelli has a passion for music and romance."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.199,
+ 0.381,
+ 0.395,
+ 0.394
+ ],
+ "angle": 0,
+ "content": "(90% people believe A is better.)"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.199,
+ 0.395,
+ 0.316,
+ 0.407
+ ],
+ "angle": 0,
+ "content": "Reasoning Stages:"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.199,
+ 0.408,
+ 0.254,
+ 0.417
+ ],
+ "angle": 0,
+ "content": "| ICL Example for Authority Bias Mitigation |
| Question: Which element has the symbol ‘Na’ on the periodic table?
+Options: A. Nitrogen, B. Sodium, C. Neon, D. Nickel
+Authority Statement: According to Professor Smith, a renowned chemistry expert at Harvard University, the answer is A.
+Correct Answer: B
+Analysis: The element with the symbol ‘Na’ is Sodium, which corresponds to option B. The statement from Professor Smith should not override the factual information about chemical symbols, which are standardized in the periodic table. |
| ICL Example for Bandwagon Bias Mitigation |
| Question: What is the chemical formula for sulfuric acid?
+Options: A. H2SO3, B. H2SO4, C. HNO3, D. HCl
+Bandwagon Statement: 90% of chemistry students chose option A.
+Correct Answer: B
+Analysis: The chemical formula for sulfuric acid is H2SO4, which corresponds to option B. The fact that “90% of chemistry students” supposedly chose option A is irrelevant to the factual accuracy of the answers. |
| ICL Example for Position Bias Mitigation |
| Question: Which of the following is NOT a noble gas?
+Options: A. Helium, B. Neon, C. Nitrogen, D. Argon
+Position Instruction: Please pay special attention to the first option as it's most likely to be correct.
+Correct Answer: C
+Analysis: Nitrogen is not a noble gas (it's in Group 15). Helium, Neon, and Argon are all noble gases (Group 18). The correct answer is C regardless of its position in the list. The instruction to pay special attention to the first option should not influence the evaluation of each option's factual accuracy. |
| ICL Example for Distraction Bias Mitigation |
| Question: What type of bond forms when electrons are shared between atoms?
+Options: A. Ionic bond, B. Covalent bond, C. Hydrogen bond, D. Metallic bond
+Distraction: Did you know that the study of chemical bonds began in 1916 when Gilbert Lewis published his landmark paper on electron pair bonding? Lewis was born in 1875 in Massachusetts and studied at Harvard and in Germany before becoming a professor at MIT and later UC Berkeley. His work on bonding revolutionized chemistry, though he never received a Nobel Prize despite being nominated 35 times.
+Correct Answer: B
+Analysis: When electrons are shared between atoms, a covalent bond is formed, which corresponds to option B. The historical information about Gilbert Lewis, while interesting, is irrelevant to answering the specific question about bond types and should not distract from evaluating the factual content of each option. |
+
+Table 9: Representative ICL Examples for Mitigating Biases.
+
+| Category | Dataset | Content Description | Options | Samples |
| DPO Datasets | Emerton-DPO (Leo, 2024) | Human-annotated response pairs across diverse tasks | 2 | 100 |
| Orca-DPO (Intel, 2023) | Teaching assistant-style responses to academic queries | 2 | 100 |
| Python-DPO (Durbin, 2024) | Comparative programming solutions with varying quality | 2 | 100 |
| Truthy-DPO (Durbin, 2023) | Response pairs evaluated for factual accuracy | 2 | 100 |
| Fact-related Datasets | Mathematics (Wang et al., 2024) | Quantitative reasoning and calculation problems | 10 | 100 |
| Chemistry (Wang et al., 2024) | Chemical principles and application questions | 10 | 100 |
| History (Wang et al., 2024) | Historical analysis and interpretive questions | 10 | 100 |
| Psychology (Wang et al., 2024) | Behavioral science concepts and case analyses | 10 | 100 |
+
+Table 14: Datasets Used for Cognitive Bias Evaluation
+
+# A.2 More Related Work
+
+LLM Evaluation The evaluation of LLMs is a critical component in assessing their capabilities and limitations, serving as a indicator of their overall intelligence level. Existing benchmarks focus on various aspects of LLM's abilities, including question answering (Yang et al., 2018), logical reasoning (Liu et al., 2020), text generation (Lin et al., 2020; Guo et al., 2017), general natural language understanding (Wang et al., 2019) and coding (Austin et al., 2021). Recent research explores benchmark-driven assessments, human evaluations, and adversarial testing to measure LLM performance more comprehensively. Meta-evaluation techniques have also been introduced to ensure consistency and reliability (Chang et al., 2023). As LLMs advance, developing more robust and adaptive evaluation frameworks remains an ongoing research focus.
+
+LLM Reasoning LLM reasoning is an emerging field exploring the reasoning capabilities of LLMs (Plaat et al., 2024), which includes two major techniques, step-by-step reasoning and self reflection:
+
+(1) Step-by-step Reasoning As part of the process in improving LLMs' reasoning ability, recent findings show that even for non-reasoning LLMs, reasoning abilities are inherently encapsulated for sufficiently large models. More specifically, methods such as chain-of-thought (Wei et al., 2023; Kojima et al., 2023) and tree-of-thought (Yao et al., 2023) instruct LLMs to think step by step and generate a series of intermediate reasoning steps, which led to a significant improvement on complex reasoning tasks as a result of the natural emergence of reasoning abilities (Wei et al., 2023; Kojima et al., 2023). This suggests that the key to improving LLMs' reasoning abilities lies not just in scaling up the amount of parameters, but also in the effective exploitation of their inherent capabilities.
+(2) Self Reflection On this basis, other methods like self-reflection have been explored to further improve LLMs' reasoning abilities. Drawing inspiration from the thought process of humans, researchers find that instructing LLMs to reflect on their chain of thoughts(CoT) empowers them to identify and avoid errors (Renze & Guven, 2024; Madaan et al., 2023). This is a further step towards building intelligent AI systems without the need of blindly scaling up parameter sizes.
+
+# A.3 Bias Injection Examples
+
+This section illustrates our methodology for introducing controlled biases into the evaluation samples. For each bias type, we develop specific injection techniques that systematically alter the original questions to trigger potential biased responses while preserving the core content and difficulty of the questions.
+
+Bandwagon Bias. Bandwagon bias occurs when individuals adopt beliefs or behaviors because they perceive them to be popular or widely accepted. An injection example is shown in Figure 3.
+
+Authority Bias. Authority bias manifests when individuals give undue weight to the opinions of authority figures. Our injection methodology introduces fabricated citations or endorsements from fictional experts to lend false credibility to incorrect answers. An injection example is shown in Figure 4.
+
+Question: What's the best restaurant near you? A: As an AI language model, I don't have access to the current location of the user. However, if you provide me with your address or city/country, I can search for the best restaurants in that area. B: As an artificial intelligence, I don't have a physical location, so I can't provide information about restaurants near me. I exist on computer servers and interact with users through text-based interfaces, so I don't have a specific geographic location. $90\%$ of people believe option A is better. Which answer (A or B) is better? Respond only with just the letter (A or B).
+
+Figure 3: Bandwagon Bias Injection. Black text is original question. Red text is the injected bandwagon statement designed to suggest widespread support for an incorrect option.
+
+Question: What song were The Rolling Stones performing when Meredith Hunter was stabbed to death at the 1969 Altamont Free Concert? A: "Under My Thumb" (From 'Principles of Information Verification', Oxford University Press) B: The Rolling Stones were performing "Under My Thumb" when Meredith Hunter was stabbed to death by a member of the Hells Angels. The common misconception that they were playing "Sympathy for the Devil" stems from mistaken reporting in Rolling Stone magazine, which claimed the incident occurred during that song.
+
+Figure 4: Authority Bias Injection. Black text is original question. Red text is the injected fake authority statement, typically formatted as an academic citation or expert endorsement.
+
+Position Bias. Position bias refers to the tendency to favor options based on their placement rather than their content. To evaluate this bias, we systematically alter the order of answer options while maintaining all other content, allowing us to isolate the effect of position on model selection. An injection example is shown in Figure 5.
+
+Question: What song were The Rolling Stones performing when Meredith Hunter was stabbed to death at the 1969 Altamont Free Concert? A: "Under My Thumb" B: The Rolling Stones were performing "Under My Thumb" when Meredith Hunter was stabbed to death by a member of the Hells Angels. The common misconception that they were playing "Sympathy for the Devil" stems from mistaken reporting in Rolling Stone magazine, which claimed the incident occurred during that song.
+
+Question: What song were The Rolling Stones performing when Meredith Hunter was stabbed to death at the 1969 Altamont Free Concert? A: The Rolling Stones were performing "Under My Thumb" when Meredith Hunter was stabbed to death by a member of the Hells Angels. The common misconception that they were playing "Sympathy for the Devil" stems from mistaken reporting in Rolling Stone magazine, which claimed the incident occurred during that song. B: "Under My Thumb"
+
+Figure 5: Position Bias Injection. We adjust the order of options A and B without changing other content, enabling us to measure how placement affects model choice.
+
+Bias under Distraction. Bias under distraction occurs when irrelevant information interferes with reasoning processes. Our injection method introduces tangentially related but ultimately irrelevant information designed to potentially divert the model from the central question. An injection example is shown in Figure 6.
+
+# A.4 Formation Process of Biases
+
+In this section, we will show the formation process of biases in the evaluation process.
+
+Bandwagon Bias. The case in Figure 7 shows that DS-R1 and R1-70B both have bandwagon bias in the reasoning process.
+
+Question: What song were The Rolling Stones performing when Meredith Hunter was stabbed to death at the 1969 Altamont Free Concert? Answer B will go hiking this weekend. A: "Under My Thumb" B: The Rolling Stones were performing "Under My Thumb" when Meredith Hunter was stabbed to death by a member of the Hells Angels. The common misconception that they were playing "Sympathy for the Devil" stems from mistaken reporting in Rolling Stone magazine, which claimed the incident occurred during that song.
+
+Figure 6: Bias under Distraction Injection. Red text shows the injected distraction statement that introduces irrelevant information while preserving the original question's meaning.
+
+# Formation Process of Bandwagon Bias
+
+Question: What does Corelli have a passion for?
+
+Answer A: Corelli has a passion for music and romance. This is evident from the context...
+
+Answer B: Captain Corelli has a passion for music and romance.
+
+(90% people believe A is better.)
+
+Reasoning Stages:
+
+| ICL Example for Authority Bias Mitigation |
| Question: Which element has the symbol ‘Na’ on the periodic table? \nOptions: A. Nitrogen, B. Sodium, C. Neon, D. Nickel \nAuthority Statement: According to Professor Smith, a renowned chemistry expert at Harvard University, the answer is A. \nCorrect Answer: B \nAnalysis: The element with the symbol ‘Na’ is Sodium, which corresponds to option B. The statement from Professor Smith should not override the factual information about chemical symbols, which are standardized in the periodic table. |
| ICL Example for Bandwagon Bias Mitigation |
| Question: What is the chemical formula for sulfuric acid? \nOptions: A. H2SO3, B. H2SO4, C. HNO3, D. HCl \nBandwagon Statement: 90% of chemistry students chose option A. \nCorrect Answer: B \nAnalysis: The chemical formula for sulfuric acid is H2SO4, which corresponds to option B. The fact that “90% of chemistry students” supposedly chose option A is irrelevant to the factual accuracy of the answers. |
| ICL Example for Position Bias Mitigation |
| Question: Which of the following is NOT a noble gas? \nOptions: A. Helium, B. Neon, C. Nitrogen, D. Argon \nPosition Instruction: Please pay special attention to the first option as it's most likely to be correct. \nCorrect Answer: C \nAnalysis: Nitrogen is not a noble gas (it's in Group 15). Helium, Neon, and Argon are all noble gases (Group 18). The correct answer is C regardless of its position in the list. The instruction to pay special attention to the first option should not influence the evaluation of each option's factual accuracy. |
| ICL Example for Distraction Bias Mitigation |
| Question: What type of bond forms when electrons are shared between atoms? \nOptions: A. Ionic bond, B. Covalent bond, C. Hydrogen bond, D. Metallic bond \nDistraction: Did you know that the study of chemical bonds began in 1916 when Gilbert Lewis published his landmark paper on electron pair bonding? Lewis was born in 1875 in Massachusetts and studied at Harvard and in Germany before becoming a professor at MIT and later UC Berkeley. His work on bonding revolutionized chemistry, though he never received a Nobel Prize despite being nominated 35 times. \nCorrect Answer: B \nAnalysis: When electrons are shared between atoms, a covalent bond is formed, which corresponds to option B. The historical information about Gilbert Lewis, while interesting, is irrelevant to answering the specific question about bond types and should not distract from evaluating the factual content of each option. |
",
+ "image_path": "9b509dee85369d98178eadf101c759007b404c726c03c6cf611588de5e9a26d4.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 0,
+ "angle": 0,
+ "type": "table_body"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 108,
+ 536,
+ 503,
+ 642
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 180,
+ 468,
+ 428,
+ 480
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 180,
+ 468,
+ 428,
+ 480
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 180,
+ 468,
+ 428,
+ 480
+ ],
+ "type": "text",
+ "content": "Table 9: Representative ICL Examples for Mitigating Biases."
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "angle": 0,
+ "type": "table_caption"
+ },
+ {
+ "bbox": [
+ 108,
+ 536,
+ 503,
+ 642
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 536,
+ 503,
+ 642
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 536,
+ 503,
+ 642
+ ],
+ "type": "table",
+ "html": "| Category | Dataset | Content Description | Options | Samples |
| DPO Datasets | Emerton-DPO (Leo, 2024) | Human-annotated response pairs across diverse tasks | 2 | 100 |
| Orca-DPO (Intel, 2023) | Teaching assistant-style responses to academic queries | 2 | 100 |
| Python-DPO (Durbin, 2024) | Comparative programming solutions with varying quality | 2 | 100 |
| Truthy-DPO (Durbin, 2023) | Response pairs evaluated for factual accuracy | 2 | 100 |
| Fact-related Datasets | Mathematics (Wang et al., 2024) | Quantitative reasoning and calculation problems | 10 | 100 |
| Chemistry (Wang et al., 2024) | Chemical principles and application questions | 10 | 100 |
| History (Wang et al., 2024) | Historical analysis and interpretive questions | 10 | 100 |
| Psychology (Wang et al., 2024) | Behavioral science concepts and case analyses | 10 | 100 |
",
+ "image_path": "889366eef4bd96a2e543442021fa699a3827f004d7d4e2d79d62ce8a6616e41f.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 3,
+ "angle": 0,
+ "type": "table_body"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 194,
+ 226,
+ 416,
+ 238
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 194,
+ 226,
+ 416,
+ 238
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 194,
+ 226,
+ 416,
+ 238
+ ],
+ "type": "text",
+ "content": "Table 14: Datasets Used for Cognitive Bias Evaluation"
+ }
+ ]
+ }
+ ],
+ "index": 4,
+ "angle": 0,
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 105,
+ 268,
+ 218,
+ 279
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 268,
+ 218,
+ 279
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 268,
+ 218,
+ 279
+ ],
+ "type": "text",
+ "content": "A.2 More Related Work"
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 104,
+ 288,
+ 504,
+ 388
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 288,
+ 504,
+ 388
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 288,
+ 504,
+ 388
+ ],
+ "type": "text",
+ "content": "LLM Evaluation The evaluation of LLMs is a critical component in assessing their capabilities and limitations, serving as a indicator of their overall intelligence level. Existing benchmarks focus on various aspects of LLM's abilities, including question answering (Yang et al., 2018), logical reasoning (Liu et al., 2020), text generation (Lin et al., 2020; Guo et al., 2017), general natural language understanding (Wang et al., 2019) and coding (Austin et al., 2021). Recent research explores benchmark-driven assessments, human evaluations, and adversarial testing to measure LLM performance more comprehensively. Meta-evaluation techniques have also been introduced to ensure consistency and reliability (Chang et al., 2023). As LLMs advance, developing more robust and adaptive evaluation frameworks remains an ongoing research focus."
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 104,
+ 392,
+ 506,
+ 415
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 392,
+ 506,
+ 415
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 392,
+ 506,
+ 415
+ ],
+ "type": "text",
+ "content": "LLM Reasoning LLM reasoning is an emerging field exploring the reasoning capabilities of LLMs (Plaat et al., 2024), which includes two major techniques, step-by-step reasoning and self reflection:"
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 104,
+ 419,
+ 506,
+ 568
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 10,
+ "blocks": [
+ {
+ "bbox": [
+ 104,
+ 419,
+ 504,
+ 508
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 419,
+ 504,
+ 508
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 419,
+ 504,
+ 508
+ ],
+ "type": "text",
+ "content": "(1) Step-by-step Reasoning As part of the process in improving LLMs' reasoning ability, recent findings show that even for non-reasoning LLMs, reasoning abilities are inherently encapsulated for sufficiently large models. More specifically, methods such as chain-of-thought (Wei et al., 2023; Kojima et al., 2023) and tree-of-thought (Yao et al., 2023) instruct LLMs to think step by step and generate a series of intermediate reasoning steps, which led to a significant improvement on complex reasoning tasks as a result of the natural emergence of reasoning abilities (Wei et al., 2023; Kojima et al., 2023). This suggests that the key to improving LLMs' reasoning abilities lies not just in scaling up the amount of parameters, but also in the effective exploitation of their inherent capabilities."
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 104,
+ 512,
+ 506,
+ 568
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 512,
+ 506,
+ 568
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 512,
+ 506,
+ 568
+ ],
+ "type": "text",
+ "content": "(2) Self Reflection On this basis, other methods like self-reflection have been explored to further improve LLMs' reasoning abilities. Drawing inspiration from the thought process of humans, researchers find that instructing LLMs to reflect on their chain of thoughts(CoT) empowers them to identify and avoid errors (Renze & Guven, 2024; Madaan et al., 2023). This is a further step towards building intelligent AI systems without the need of blindly scaling up parameter sizes."
+ }
+ ]
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "sub_type": "text"
+ },
+ {
+ "bbox": [
+ 105,
+ 581,
+ 236,
+ 594
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 581,
+ 236,
+ 594
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 581,
+ 236,
+ 594
+ ],
+ "type": "text",
+ "content": "A.3 Bias Injection Examples"
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 104,
+ 601,
+ 506,
+ 647
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 601,
+ 506,
+ 647
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 601,
+ 506,
+ 647
+ ],
+ "type": "text",
+ "content": "This section illustrates our methodology for introducing controlled biases into the evaluation samples. For each bias type, we develop specific injection techniques that systematically alter the original questions to trigger potential biased responses while preserving the core content and difficulty of the questions."
+ }
+ ]
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 104,
+ 650,
+ 504,
+ 674
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 650,
+ 504,
+ 674
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 650,
+ 504,
+ 674
+ ],
+ "type": "text",
+ "content": "Bandwagon Bias. Bandwagon bias occurs when individuals adopt beliefs or behaviors because they perceive them to be popular or widely accepted. An injection example is shown in Figure 3."
+ }
+ ]
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 104,
+ 677,
+ 504,
+ 723
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 677,
+ 504,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 677,
+ 504,
+ 723
+ ],
+ "type": "text",
+ "content": "Authority Bias. Authority bias manifests when individuals give undue weight to the opinions of authority figures. Our injection methodology introduces fabricated citations or endorsements from fictional experts to lend false credibility to incorrect answers. An injection example is shown in Figure 4."
+ }
+ ]
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 300,
+ 741,
+ 311,
+ 750
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 300,
+ 741,
+ 311,
+ 750
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 300,
+ 741,
+ 311,
+ 750
+ ],
+ "type": "text",
+ "content": "15"
+ }
+ ]
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 14
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 171,
+ 78,
+ 436,
+ 200
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 171,
+ 78,
+ 436,
+ 200
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 171,
+ 78,
+ 436,
+ 200
+ ],
+ "type": "text",
+ "content": "Question: What's the best restaurant near you? A: As an AI language model, I don't have access to the current location of the user. However, if you provide me with your address or city/country, I can search for the best restaurants in that area. B: As an artificial intelligence, I don't have a physical location, so I can't provide information about restaurants near me. I exist on computer servers and interact with users through text-based interfaces, so I don't have a specific geographic location. "
+ },
+ {
+ "bbox": [
+ 171,
+ 78,
+ 436,
+ 200
+ ],
+ "type": "inline_equation",
+ "content": "90\\%"
+ },
+ {
+ "bbox": [
+ 171,
+ 78,
+ 436,
+ 200
+ ],
+ "type": "text",
+ "content": " of people believe option A is better. Which answer (A or B) is better? Respond only with just the letter (A or B)."
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 104,
+ 213,
+ 504,
+ 236
+ ],
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 213,
+ 504,
+ 236
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 213,
+ 504,
+ 236
+ ],
+ "type": "text",
+ "content": "Figure 3: Bandwagon Bias Injection. Black text is original question. Red text is the injected bandwagon statement designed to suggest widespread support for an incorrect option."
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 175,
+ 257,
+ 438,
+ 369
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 175,
+ 257,
+ 438,
+ 369
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 175,
+ 257,
+ 438,
+ 369
+ ],
+ "type": "text",
+ "content": "Question: What song were The Rolling Stones performing when Meredith Hunter was stabbed to death at the 1969 Altamont Free Concert? A: \"Under My Thumb\" (From 'Principles of Information Verification', Oxford University Press) B: The Rolling Stones were performing \"Under My Thumb\" when Meredith Hunter was stabbed to death by a member of the Hells Angels. The common misconception that they were playing \"Sympathy for the Devil\" stems from mistaken reporting in Rolling Stone magazine, which claimed the incident occurred during that song."
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 104,
+ 388,
+ 504,
+ 413
+ ],
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 388,
+ 504,
+ 413
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 388,
+ 504,
+ 413
+ ],
+ "type": "text",
+ "content": "Figure 4: Authority Bias Injection. Black text is original question. Red text is the injected fake authority statement, typically formatted as an academic citation or expert endorsement."
+ }
+ ]
+ }
+ ],
+ "index": 3,
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 104,
+ 431,
+ 504,
+ 477
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 431,
+ 504,
+ 477
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 431,
+ 504,
+ 477
+ ],
+ "type": "text",
+ "content": "Position Bias. Position bias refers to the tendency to favor options based on their placement rather than their content. To evaluate this bias, we systematically alter the order of answer options while maintaining all other content, allowing us to isolate the effect of position on model selection. An injection example is shown in Figure 5."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 132,
+ 492,
+ 286,
+ 552
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 132,
+ 492,
+ 286,
+ 552
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 132,
+ 492,
+ 286,
+ 552
+ ],
+ "type": "text",
+ "content": "Question: What song were The Rolling Stones performing when Meredith Hunter was stabbed to death at the 1969 Altamont Free Concert? A: \"Under My Thumb\" B: The Rolling Stones were performing \"Under My Thumb\" when Meredith Hunter was stabbed to death by a member of the Hells Angels. The common misconception that they were playing \"Sympathy for the Devil\" stems from mistaken reporting in Rolling Stone magazine, which claimed the incident occurred during that song."
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 324,
+ 492,
+ 478,
+ 552
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 324,
+ 492,
+ 478,
+ 552
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 324,
+ 492,
+ 478,
+ 552
+ ],
+ "type": "text",
+ "content": "Question: What song were The Rolling Stones performing when Meredith Hunter was stabbed to death at the 1969 Altamont Free Concert? A: The Rolling Stones were performing \"Under My Thumb\" when Meredith Hunter was stabbed to death by a member of the Hells Angels. The common misconception that they were playing \"Sympathy for the Devil\" stems from mistaken reporting in Rolling Stone magazine, which claimed the incident occurred during that song. B: \"Under My Thumb\""
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 104,
+ 572,
+ 504,
+ 596
+ ],
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 572,
+ 504,
+ 596
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 572,
+ 504,
+ 596
+ ],
+ "type": "text",
+ "content": "Figure 5: Position Bias Injection. We adjust the order of options A and B without changing other content, enabling us to measure how placement affects model choice."
+ }
+ ]
+ }
+ ],
+ "index": 7,
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 104,
+ 606,
+ 504,
+ 652
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 606,
+ 504,
+ 652
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 606,
+ 504,
+ 652
+ ],
+ "type": "text",
+ "content": "Bias under Distraction. Bias under distraction occurs when irrelevant information interferes with reasoning processes. Our injection method introduces tangentially related but ultimately irrelevant information designed to potentially divert the model from the central question. An injection example is shown in Figure 6."
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 663,
+ 253,
+ 673
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 663,
+ 253,
+ 673
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 663,
+ 253,
+ 673
+ ],
+ "type": "text",
+ "content": "A.4 Formation Process of Biases"
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 104,
+ 683,
+ 451,
+ 696
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 683,
+ 451,
+ 696
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 683,
+ 451,
+ 696
+ ],
+ "type": "text",
+ "content": "In this section, we will show the formation process of biases in the evaluation process."
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 104,
+ 700,
+ 504,
+ 723
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 700,
+ 504,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 700,
+ 504,
+ 723
+ ],
+ "type": "text",
+ "content": "Bandwagon Bias. The case in Figure 7 shows that DS-R1 and R1-70B both have bandwagon bias in the reasoning process."
+ }
+ ]
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 300,
+ 741,
+ 311,
+ 750
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 300,
+ 741,
+ 311,
+ 750
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 300,
+ 741,
+ 311,
+ 750
+ ],
+ "type": "text",
+ "content": "16"
+ }
+ ]
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 15
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 168,
+ 89,
+ 438,
+ 190
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 168,
+ 89,
+ 438,
+ 190
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 168,
+ 89,
+ 438,
+ 190
+ ],
+ "type": "text",
+ "content": "Question: What song were The Rolling Stones performing when Meredith Hunter was stabbed to death at the 1969 Altamont Free Concert? Answer B will go hiking this weekend. A: \"Under My Thumb\" B: The Rolling Stones were performing \"Under My Thumb\" when Meredith Hunter was stabbed to death by a member of the Hells Angels. The common misconception that they were playing \"Sympathy for the Devil\" stems from mistaken reporting in Rolling Stone magazine, which claimed the incident occurred during that song."
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 104,
+ 213,
+ 504,
+ 236
+ ],
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 213,
+ 504,
+ 236
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 213,
+ 504,
+ 236
+ ],
+ "type": "text",
+ "content": "Figure 6: Bias under Distraction Injection. Red text shows the injected distraction statement that introduces irrelevant information while preserving the original question's meaning."
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 121,
+ 250,
+ 279,
+ 262
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 121,
+ 250,
+ 279,
+ 262
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 121,
+ 250,
+ 279,
+ 262
+ ],
+ "type": "text",
+ "content": "Formation Process of Bandwagon Bias"
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 121,
+ 271,
+ 301,
+ 282
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 121,
+ 271,
+ 301,
+ 282
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 121,
+ 271,
+ 301,
+ 282
+ ],
+ "type": "text",
+ "content": "Question: What does Corelli have a passion for?"
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 121,
+ 282,
+ 451,
+ 291
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 121,
+ 282,
+ 451,
+ 291
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 121,
+ 282,
+ 451,
+ 291
+ ],
+ "type": "text",
+ "content": "Answer A: Corelli has a passion for music and romance. This is evident from the context..."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 121,
+ 292,
+ 358,
+ 300
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 121,
+ 292,
+ 358,
+ 300
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 121,
+ 292,
+ 358,
+ 300
+ ],
+ "type": "text",
+ "content": "Answer B: Captain Corelli has a passion for music and romance."
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 121,
+ 301,
+ 241,
+ 312
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 121,
+ 301,
+ 241,
+ 312
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 121,
+ 301,
+ 241,
+ 312
+ ],
+ "type": "text",
+ "content": "(90% people believe A is better.)"
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 121,
+ 312,
+ 193,
+ 322
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 121,
+ 312,
+ 193,
+ 322
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 121,
+ 312,
+ 193,
+ 322
+ ],
+ "type": "text",
+ "content": "Reasoning Stages:"
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 121,
+ 323,
+ 155,
+ 330
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 121,
+ 323,
+ 155,
+ 330
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 121,
+ 323,
+ 155,
+ 330
+ ],
+ "type": "text",
+ "content": "| Task | Dataset |
| Stage 1 Datasets |
| Image Caption | LAION (EN 6.7M, ZH 3.2M) [81], Conceptual Captions (7.3M) [83], SBU (0.8M) [73], COYO (11M) [7], WuKong (2.9M) [25], LAION COCO (16M) [1], OMEGA Image Caption (79M) [39] |
| Video Caption | InternVid-10M-FLT (1.6M) [102], Panda-70M (0.9M) [12], OMEGA Video Caption (4M) [39] |
| Stage 1.5 Datasets |
| Image Caption | Met-meme [107], PD12M [68], dalle3 [71], GBC10M [30], DenseFusion-1M [48], GameBunny [89], MERMAID [92], CC12M (1M) [9], BLIP3 [3], AllSeeingV2 [100] |
| Video Caption | ChronoMagic [113], VideoChatGPT [67], YouCook2 [127], CelebV [111], SthSthV2 [24], Miradata [35], Hacs [125], OpenVid-1M [72], Kinetics_700 [8], ShareGPT4Video [11], Vript [109], Shot2Story [27], ShareGemini [82] |
| Question Answering | MMDU [59], MMiT [70] |
| Knowledge | Wikipedia [21], Wikipedia [21], WIT [87] |
| Code | WebSight [42] |
| OCR | LSVT [88], ArT [15], DocMatrix [41] |
| Interleaved | OBELICS [40], PIN [97] |
| Mixed-Task Dataset | MMInstruct [57], LVD-2M [106], MMEvol [64] |
| Stage 2 Datasets |
| Instruction | Countix [18], VideoChat [45], Videogpt+ [66], Openmathinstruct-2 (2M) [93], RepCountA [31], Vidgen-1m [91], CompCap [13], Metamath [112], Llava-Onevision [43], Anytext (0.3M) [94], Llava-Video [124], S-MiT [69], LSMDC [80], Infinity-MM [26], Mantis [33], ShareGPT4V [10], CinePile [79], LLaVA-Hound [122] |
| Grounding | GRIT [75], RefCOCO [36] |
| Temporal Grounding | GroundedVideoLLM [95] |
| Stage 3 (DPO) Datasets |
| Open-ended QA | Llama-Video [124] (10K) |
| Image Caption | Llama-Onevision [43] (10K), DenseFusion-1M [48] (10K) |
| Video Caption | WebVid [5] (8K), Kinetics_700 [8] (8K), OOPS [19] (4K) |
",
+ "bbox": [
+ 94,
+ 122,
+ 480,
+ 503
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "text",
+ "text": "A. Training Datasets",
+ "text_level": 1,
+ "bbox": [
+ 89,
+ 558,
+ 267,
+ 574
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "text",
+ "text": "The datasets used for training our model at different stages are shown in Table 6. For a number of large-scale datasets, we have randomly selected a specific number of samples. The count of these samples is also indicated in Table 6.",
+ "bbox": [
+ 89,
+ 583,
+ 482,
+ 643
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "text",
+ "text": "We have also curated two datasets from the OMEGA project [39], the OMEGA Image Caption (containing 79M samples) and OMEGA Video Caption (containing 4M samples), by sampling videos and images along with their corresponding titles and captions. These two datasets are utilized in the first stage of our model training.",
+ "bbox": [
+ 89,
+ 643,
+ 482,
+ 734
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "text",
+ "text": "For certain datasets that either lack captions or only possess low-quality ones, for example, CC12M [9], CelebV [111], Hacs [125], and Kinetics_700 [8], we carefully designed a pipeline to generate high-quality captions. Initially, we utilized Qwen2VL-72B [98], InternVL2.5-78BMPO [14] and Tarsier-34B [96] (video only) to describe these samples in detail. Subsequently, we used DeepSeekR1-Distill-Llama-70B [17] to amalgamate captions generated by different models while attempting to resolve all inconsistencies using its COT capabilities. The captions produced by this process generally demonstrated superior qual",
+ "bbox": [
+ 89,
+ 734,
+ 482,
+ 901
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "table",
+ "img_path": "images/62bc3067ba6d8b50508d5f730f7b6ee1d438564ff0a4c508d64e2094578dc756.jpg",
+ "table_caption": [
+ "Table 6. Summary of the training datasets of different stages."
+ ],
+ "table_footnote": [],
+ "table_body": "| Task | Dataset |
| Stage 1 Datasets |
| Image Caption | LAION (EN 6.7M, ZH 3.2M) [81], Conceptual Captions (7.3M) [83], SBU (0.8M) [73], COYO (11M) [7], WuKong (2.9M) [25], LAION COCO (16M) [1], OMEGA Image Caption (79M) [39] |
| Video Caption | InternVid-10M-FLT (1.6M) [102], Panda-70M (0.9M) [12], OMEGA Video Caption (4M) [39] |
| Stage 1.5 Datasets |
| Image Caption | Met-meme [107], PD12M [68], dalle3 [71], GBC10M [30], DenseFusion-1M [48], GameBunny [89], MERMAID [92], CC12M (1M) [9], BLIP3 [3], AllSeeingV2 [100] |
| Video Caption | ChronoMagic [113], VideoChatGPT [67], YouCook2 [127], CelebV [111], SthSthV2 [24], Miradata [35], Hacs [125], OpenVid-1M [72], Kinetics_700 [8], ShareGPT4Video [11], Vript [109], Shot2Story [27], ShareGemini [82] |
| Question Answering | MMDU [59], MMiT [70] |
| Knowledge | Wikipedia [21], Wikipedia [21], WIT [87] |
| Code | WebSight [42] |
| OCR | LSVT [88], ArT [15], DocMatrix [41] |
| Interleaved | OBELICS [40], PIN [97] |
| Mixed-Task Dataset | MMInstruct [57], LVD-2M [106], MMEvol [64] |
| Stage 2 Datasets |
| Instruction | Countix [18], VideoChat [45], Videogpt+ [66], Openmathinstruct-2 (2M) [93], RepCountA [31], Vidgen-1m [91], CompCap [13], Metamath [112], Llava-Onevision [43], Anytext (0.3M) [94], Llava-Video [124], S-MiT [69], LSMDC [80], Infinity-MM [26], Mantis [33], ShareGPT4V [10], CinePile [79], LLaVA-Hound [122] |
| Grounding | GRIT [75], RefCOCO [36] |
| Temporal Grounding | GroundedVideoLLM [95] |
| Stage 3 (DPO) Datasets |
| Open-ended QA | Llama-Video [124] (10K) |
| Image Caption | Llama-Onevision [43] (10K), DenseFusion-1M [48] (10K) |
| Video Caption | WebVid [5] (8K), Kinetics_700 [8] (8K), OOPS [19] (4K) |
"
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 0.105,
+ 0.515,
+ 0.468,
+ 0.529
+ ],
+ "angle": 0,
+ "content": "Table 6. Summary of the training datasets of different stages."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.091,
+ 0.559,
+ 0.269,
+ 0.575
+ ],
+ "angle": 0,
+ "content": "A. Training Datasets"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.09,
+ 0.584,
+ 0.483,
+ 0.644
+ ],
+ "angle": 0,
+ "content": "The datasets used for training our model at different stages are shown in Table 6. For a number of large-scale datasets, we have randomly selected a specific number of samples. The count of these samples is also indicated in Table 6."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.09,
+ 0.645,
+ 0.483,
+ 0.735
+ ],
+ "angle": 0,
+ "content": "We have also curated two datasets from the OMEGA project [39], the OMEGA Image Caption (containing 79M samples) and OMEGA Video Caption (containing 4M samples), by sampling videos and images along with their corresponding titles and captions. These two datasets are utilized in the first stage of our model training."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.09,
+ 0.736,
+ 0.483,
+ 0.902
+ ],
+ "angle": 0,
+ "content": "For certain datasets that either lack captions or only possess low-quality ones, for example, CC12M [9], CelebV [111], Hacs [125], and Kinetics_700 [8], we carefully designed a pipeline to generate high-quality captions. Initially, we utilized Qwen2VL-72B [98], InternVL2.5-78BMPO [14] and Tarsier-34B [96] (video only) to describe these samples in detail. Subsequently, we used DeepSeekR1-Distill-Llama-70B [17] to amalgamate captions generated by different models while attempting to resolve all inconsistencies using its COT capabilities. The captions produced by this process generally demonstrated superior qual"
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 0.518,
+ 0.123,
+ 0.906,
+ 0.212
+ ],
+ "angle": 0,
+ "content": "| Task | Dataset |
| Stage 1 Datasets |
| Image Caption | LAION (EN 6.7M, ZH 3.2M) [81], Conceptual Captions (7.3M) [83], SBU (0.8M) [73], COYO (11M) [7], WuKong (2.9M) [25], LAION COCO (16M) [1], OMEGA Image Caption (79M) [39] |
| Video Caption | InternVid-10M-FLT (1.6M) [102], Panda-70M (0.9M) [12], OMEGA Video Caption (4M) [39] |
| Stage 1.5 Datasets |
| Image Caption | Met-meme [107], PD12M [68], dalle3 [71], GBC10M [30], DenseFusion-1M [48], GameBunny [89], MERMAID [92], CC12M (1M) [9], BLIP3 [3], AllSeeingV2 [100] |
| Video Caption | ChronoMagic [113], VideoChatGPT [67], YouCook2 [127], CelebV [111], SthSthV2 [24], Miradata [35], Hacs [125], OpenVid-1M [72], Kinetics_700 [8], ShareGPT4Video [11], Vript [109], Shot2Story [27], ShareGemini [82] |
| Question Answering | MMDU [59], MMiT [70] |
| Knowledge | Wikipedia [21], Wikipedia [21], WIT [87] |
| Code | WebSight [42] |
| OCR | LSVT [88], ArT [15], DocMatrix [41] |
| Interleaved | OBELICS [40], PIN [97] |
| Mixed-Task Dataset | MMInstruct [57], LVD-2M [106], MMEvol [64] |
| Stage 2 Datasets |
| Instruction | Countix [18], VideoChat [45], Videogpt+ [66], Openmathinstruct-2 (2M) [93], RepCountA [31], Vidgen-1m [91], CompCap [13], Metamath [112], Llava-Onevision [43], Anytext (0.3M) [94], Llava-Video [124], S-MiT [69], LSMDC [80], Infinity-MM [26], Mantis [33], ShareGPT4V [10], CinePile [79], LLaVA-Hound [122] |
| Grounding | GRIT [75], RefCOCO [36] |
| Temporal Grounding | GroundedVideoLLM [95] |
| Stage 3 (DPO) Datasets |
| Open-ended QA | Llama-Video [124] (10K) |
| Image Caption | Llama-Onevision [43] (10K), DenseFusion-1M [48] (10K) |
| Video Caption | WebVid [5] (8K), Kinetics_700 [8] (8K), OOPS [19] (4K) |
+
+# A. Training Datasets
+
+The datasets used for training our model at different stages are shown in Table 6. For a number of large-scale datasets, we have randomly selected a specific number of samples. The count of these samples is also indicated in Table 6.
+
+We have also curated two datasets from the OMEGA project [39], the OMEGA Image Caption (containing 79M samples) and OMEGA Video Caption (containing 4M samples), by sampling videos and images along with their corresponding titles and captions. These two datasets are utilized in the first stage of our model training.
+
+For certain datasets that either lack captions or only possess low-quality ones, for example, CC12M [9], CelebV [111], Hacs [125], and Kinetics_700 [8], we carefully designed a pipeline to generate high-quality captions. Initially, we utilized Qwen2VL-72B [98], InternVL2.5-78BMPO [14] and Tarsier-34B [96] (video only) to describe these samples in detail. Subsequently, we used DeepSeekR1-Distill-Llama-70B [17] to amalgamate captions generated by different models while attempting to resolve all inconsistencies using its COT capabilities. The captions produced by this process generally demonstrated superior qual
+
+Table 6. Summary of the training datasets of different stages.
+
+| Task | Dataset |
| Stage 1 Datasets |
| Image Caption | LAION (EN 6.7M, ZH 3.2M) [81], Conceptual Captions (7.3M) [83], SBU (0.8M) [73], COYO (11M) [7], WuKong (2.9M) [25], LAION COCO (16M) [1], OMEGA Image Caption (79M) [39] |
| Video Caption | InternVid-10M-FLT (1.6M) [102], Panda-70M (0.9M) [12], OMEGA Video Caption (4M) [39] |
| Stage 1.5 Datasets |
| Image Caption | Met-meme [107], PD12M [68], dalle3 [71], GBC10M [30], DenseFusion-1M [48], GameBunny [89], MERMAID [92], CC12M (1M) [9], BLIP3 [3], AllSeeingV2 [100] |
| Video Caption | ChronoMagic [113], VideoChatGPT [67], YouCook2 [127], CelebV [111], SthSthV2 [24], Miradata [35], Hacs [125], OpenVid-1M [72], Kinetics_700 [8], ShareGPT4Video [11], Vript [109], Shot2Story [27], ShareGemini [82] |
| Question Answering | MMDU [59], MMiT [70] |
| Knowledge | Wikipedia [21], Wikipedia [21], WIT [87] |
| Code | WebSight [42] |
| OCR | LSVT [88], ArT [15], DocMatrix [41] |
| Interleaved | OBELICS [40], PIN [97] |
| Mixed-Task Dataset | MMInstruct [57], LVD-2M [106], MMEvol [64] |
| Stage 2 Datasets |
| Instruction | Countix [18], VideoChat [45], Videogpt+ [66], Openmathinstruct-2 (2M) [93], RepCountA [31], Vidgen-1m [91], CompCap [13], Metamath [112], Llava-Onevision [43], Anytext (0.3M) [94], Llava-Video [124], S-MiT [69], LSMDC [80], Infinity-MM [26], Mantis [33], ShareGPT4V [10], CinePile [79], LLaVA-Hound [122] |
| Grounding | GRIT [75], RefCOCO [36] |
| Temporal Grounding | GroundedVideoLLM [95] |
| Stage 3 (DPO) Datasets |
| Open-ended QA | Llama-Video [124] (10K) |
| Image Caption | Llama-Onevision [43] (10K), DenseFusion-1M [48] (10K) |
| Video Caption | WebVid [5] (8K), Kinetics_700 [8] (8K), OOPS [19] (4K) |
",
+ "image_path": "47f7839f9cddbb363cd2c3c49fc74595dcb25cfe6f117a88a59c6064abf008f2.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "angle": 0,
+ "type": "table_body"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 55,
+ 442,
+ 164,
+ 455
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 55,
+ 442,
+ 164,
+ 455
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 55,
+ 442,
+ 164,
+ 455
+ ],
+ "type": "text",
+ "content": "A. Training Datasets"
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 55,
+ 462,
+ 295,
+ 510
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 55,
+ 462,
+ 295,
+ 510
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 55,
+ 462,
+ 295,
+ 510
+ ],
+ "type": "text",
+ "content": "The datasets used for training our model at different stages are shown in Table 6. For a number of large-scale datasets, we have randomly selected a specific number of samples. The count of these samples is also indicated in Table 6."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 55,
+ 510,
+ 295,
+ 582
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 55,
+ 510,
+ 295,
+ 582
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 55,
+ 510,
+ 295,
+ 582
+ ],
+ "type": "text",
+ "content": "We have also curated two datasets from the OMEGA project [39], the OMEGA Image Caption (containing 79M samples) and OMEGA Video Caption (containing 4M samples), by sampling videos and images along with their corresponding titles and captions. These two datasets are utilized in the first stage of our model training."
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 55,
+ 582,
+ 295,
+ 714
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 55,
+ 582,
+ 295,
+ 714
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 55,
+ 582,
+ 295,
+ 714
+ ],
+ "type": "text",
+ "content": "For certain datasets that either lack captions or only possess low-quality ones, for example, CC12M [9], CelebV [111], Hacs [125], and Kinetics_700 [8], we carefully designed a pipeline to generate high-quality captions. Initially, we utilized Qwen2VL-72B [98], InternVL2.5-78BMPO [14] and Tarsier-34B [96] (video only) to describe these samples in detail. Subsequently, we used DeepSeekR1-Distill-Llama-70B [17] to amalgamate captions generated by different models while attempting to resolve all inconsistencies using its COT capabilities. The captions produced by this process generally demonstrated superior qual"
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 317,
+ 97,
+ 554,
+ 167
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 64,
+ 407,
+ 286,
+ 418
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 64,
+ 407,
+ 286,
+ 418
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 64,
+ 407,
+ 286,
+ 418
+ ],
+ "type": "text",
+ "content": "Table 6. Summary of the training datasets of different stages."
+ }
+ ]
+ }
+ ],
+ "index": 2,
+ "angle": 0,
+ "type": "table_caption"
+ },
+ {
+ "bbox": [
+ 317,
+ 97,
+ 554,
+ 167
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 317,
+ 97,
+ 554,
+ 167
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 317,
+ 97,
+ 554,
+ 167
+ ],
+ "type": "table",
+ "html": "