{"id":"d45ce77e-f29d-4ea1-a9ab-9f775d5eff66","arxiv_id":"2412.11196","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"InBoL improves MLLM trustworthiness by training models to refuse questions that lie beyond their extrinsic (visual) or intrinsic (knowledge) information boundaries, using confidence-aware DPO.","lead":"This paper trains multimodal AI models to say 'I don't know' when an image or the model's own knowledge is insufficient, and introduces a new way to score how trustworthy that behavior is. The approach, called InBoL, uses confidence estimates and preference learning to raise refusal accuracy while trying to preserve helpful answering.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 2's in-domain trustworthiness scores are inconsistent with Eq. (5): for LLaVA1.5-7B, Acc=12.00 and RefR=46.10 give strust=-29.9, not -6.50; the headline example is unsupported until corrected.","rationale":"The reader's confidence-proxy concern is real but not the most load-bearing, because the paper's own numbers can be checked independently of any proxy. Eq. (5) defines the metric; Table 1 reproduces Eq. (5), but Table 2 does not. This is an internal contradiction in the primary result, not a matter of external consensus or missing ablations. A corrected computation may still favor CA-DPO, so I would not reject outright; but the paper must release corrected Table 2 values or state the exact formula used. The comparison in Table 4 makes the problem sharper: the CA-DPO row there reports Acc=49.10, RefR=30.30, Strust=28.50 (consistent with Eq. 5), while Table 2 reports Acc=87.00, RefR=49.10, Strust=28.50 for the same model; both cannot be correct. This also undermines the specific example in the strongest_claim. The confidence-proxy issue in Section 3.2 is secondary: if the metric is wrong, the method cannot be evaluated; if the metric is fixed, the proxy issue still needs reporting of the number of samples, temperature, and variance, but it does not by itself overturn the comparison.","tokens_in":18141,"tokens_out":10822,"duration_ms":97820,"concrete_test":"Recompute every row of Table 2 from the reported Acc and RefR using Eq. (5) on the percentage scale: Strust = 2*Acc + RefR - 100. (i) Check the 7B CA-DPO inconsistency directly: Table 2 gives Acc=87.00, RefR=49.10, Strust=28.50, while Table 4 gives Acc=49.10, RefR=30.30, Strust=28.50; both cannot be correct. (ii) If a different normalization was used for the in-domain column, state it explicitly. (iii) Recompute the IDK-IT vs CA-DPO comparison and the reported '-6.50 to 28.50' improvement; if the corrected ranking or effect size changes materially, revise the central claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central evidence for InBoL is the claimed score improvement in Table 2, but that table does not follow the paper's own metric. Eq. (5) defines strust = 2*Acc + RefR - 1, and every OOD row in Table 1 is consistent with this definition (e.g., AOKVQA: 2*0.7856 + 0 - 1 = 0.5713). The in-domain rows in Table 2 cannot be reproduced: for LLaVA1.5-7B, 2*0.12 + 0.461 - 1 = -0.299 (-29.9 on the percentage scale), not -6.50; for IDK-IT, 2*0.92 + 0.386 - 1 = 1.226 (+122.6), not +16.00. The problem is sharpened by comparing the 7B CA-DPO row across tables: Table 2 reports Acc=87.00, RefR=49.10, Strust=28.50, while Table 4 reports Acc=49.10, RefR=30.30, Strust=28.50 for the same condition; both cannot follow from Eq. (5). If the in-domain column uses a different normalization, the paper must state it; without that, the headline claim that CA-DPO raises the score from -6.50 to 28.50 cannot be verified. The confidence-proxy assumption in Section 3.2 is also fragile, but it is secondary: even a perfect confidence estimator would not resolve this internal contradiction.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes InBoL, an information-boundary-aware learning framework for multimodal large language models (MLLMs) that trains models to refuse answering when information is insufficient. The framework introduces a data construction pipeline that labels training questions as known, mixed, or unknown based on a confidence estimate derived from the base model's sampled-response accuracy, and then applies IDK instruction tuning (IDK-IT) and confidence-aware direct preference optimization (CA-DPO). The paper also proposes a user-centric trustworthiness score, strust = 2*Acc + RefR - 1, and evaluates it on in-domain and out-of-domain datasets with LLaVA1.5-7B/13B, claiming significant trustworthiness improvements while preserving helpfulness.","tokens_in":18462,"tokens_out":8459,"duration_ms":65980,"significance":"If the reported gains are reliable, the paper addresses an important problem in MLLM trustworthiness by systematically combining refusal behavior with confidence-aware training. The proposed hybrid evaluator (string matching plus LLM judgment) is validated against human annotations with Cohen's Kappa 0.885, which is a useful practical contribution. The data-generation pipeline is shown to work with an open-source model (Qwen2-VL) as well as GPT-4o, enhancing reproducibility. However, the central experimental claim is undermined by an internal inconsistency in the main in-domain results table, which prevents the reader from verifying the reported trustworthiness scores.","major_comments":[{"comment":"","section":"Table 2 and Eq. (5)"},{"comment":"","section":"Table 4 vs Table 2"},{"comment":"","section":"Section 3.2 and Figure 4"},{"comment":"","section":"Section 5.5"},{"comment":"","section":"Appendix E and Section 4.2"}],"minor_comments":[{"comment":"","section":"Section 3.1"},{"comment":"","section":"Throughout"},{"comment":"","section":"Section 3.2"},{"comment":"","section":"Tables 5 and 6"},{"comment":"","section":"Section 2.2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript has a serious internal inconsistency in its main results table (Table 2), which is not a mere typo: the reported Strust values cannot be derived from the reported Acc and RefR using the paper's defining equation, and Acc+RefR exceeds 100 in several rows, suggesting that the definitions of these metrics are unclear or the numbers are erroneous. The paper also contains a missing results table (Section 5.5) and placeholder references, indicating the manuscript is not fully finalized. These issues are load-bearing for the central claim, but they appear correctable with a careful revision and re-analysis; I therefore recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: Table 2, the in-domain results that drive the paper's headline claim, does not match the paper's own scoring equation. Eq. (5) gives strust = 2*Acc + RefR - 1, and this works for Table 1 (e.g., AOKVQA: 2*0.7856 + 0 - 1 = 0.5712). But Table 2's in-domain numbers cannot be reproduced: for LLaVA1.5-7B, Acc=12.00 and RefR=46.10 give -29.9, not -6.50; for IDK-IT, 2*0.92 + 0.386 - 1 = 1.226, not 16.00. Table 4 even lists the same Strust=28.50 for CA-DPO 7B with different Acc and RefR. So the claim that CA-DPO raises the score from -6.50 to 28.50 is currently unsupported.\n\nWhat is genuinely new here: the intrinsic/extrinsic boundary split is a sensible way to organize refusal training for MLLMs, and the data generation pipeline is a real contribution—it produces unanswerable questions from any VQA set, and the authors show it works with Qwen2-VL-72B as well as GPT-4o. The CA-DPO loss, weighing preference pairs by the model's own confidence, is a reasonable idea. The hybrid evaluator (string match plus Llama2-13B) is a practical detail, and the reported Kappa of 0.885 against humans is good evidence it works.\n\nThe soft spots beyond Table 2: the confidence thresholds are inconsistent (main text says 0.8/0.2; Appendix F.1 varies them over 2–8). Section 5.5 has a missing Table?? placeholder for the linear-probing accuracy. OOD evaluation uses string matching only, while in-domain uses the hybrid evaluator—so OOD accuracy may be undercounted. The confidence estimator (sample counts, temperature, variance) is not documented. The 'maintaining helpfulness' claim has no defined helpfulness metric. And the trustworthiness score is used both to select hyperparameters (Appendices F.1/F.2) and to report final results, so the evaluation is partly circular.\n\nNone of these individually kill the framework, but Table 2 is load-bearing: the abstract's central improvement is unverifiable until the authors explain the calculation. If they fix that, release the numbers, and add code/data, this becomes a solid paper. As it stands, I would send it to review with a request for major revision—not desk-reject it, because the OOD results are consistent and the framework is useful. I would not cite it in its current form.\n\nWho it is for: people working on hallucination mitigation and refusal behaviors in MLLMs. A serious referee should spend time on it.\n\nRecommendation: conditional—engage, but require the Table 2 fix and a documented confidence estimator.","headline":"Table 2 contradicts Eq. (5), so the headline trustworthiness gain is unverified; the framework and OOD results are promising enough to send to review with major-revision demands.","tokens_in":19048,"tokens_out":3989,"would_cite":false,"duration_ms":30841,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that teaching MLLMs to refuse questions outside their information boundaries sharply improves trustworthiness without sacrificing helpfulness.","keywords":["multimodal large language models","trustworthiness","refusal","information boundaries","hallucination mitigation","direct preference optimization","confidence estimation","visual question answering"],"falsifier":"Give the pipeline a model whose confidence is demonstrably miscalibrated on a controlled topic (for example, images modified so that answerability changes while the question text stays identical), and check where IDK-IT and CA-DPO draw the refusal boundary; if refusals track the biased confidence rather than true answerability, the learned boundary is an artifact of the estimate. A cheaper check on the quantitative claim: recompute the in-domain trustworthiness scores from the reported Acc and RefR values with the paper's own formula $s_{trust} = 2\\cdot Acc + RefR - 1$, since the published in-domain numbers do not reproduce and the magnitude of the gain should be re-verified with the released evaluation code or per-sample scores.","tokens_in":2076,"feed_emoji":"🤷","tokens_out":8367,"duration_ms":180862,"temperature":0.7,"pith_summary":"The paper sets out to show that a multimodal large language model (MLLM) becomes more trustworthy when it is trained to refuse answering questions that lie outside its information boundaries. It introduces the Information Boundary-aware Learning Framework (InBoL), which defines extrinsic boundaries (what the image actually contains) and intrinsic boundaries (what the model itself can know and perceive), then trains the model with instruction tuning and confidence-aware direct preference optimization. The paper's key reported result is a large jump in a user-centric trustworthiness score: LLaVA1.5-7B's in-domain score rises from -6.50 to 28.50 with the CA-DPO variant, while accuracy is preserved. This matters because it offers a concrete path to reducing hallucinations in vision-language assistants without forcing them to be silent or unhelpful.","feed_headline":"Refusal training lifts MLLM trust score by 35 points","feed_subtitle":"Refusal-aware training lifts LLaVA1.5-7B's in-domain trust score from -6.50 to 28.50.","key_machinery":"The central object is a pair of information boundaries: the extrinsic boundary, which asks whether the needed information is present in the visual input, and the intrinsic boundary, which asks whether this particular model's knowledge and perceptual capacity can supply the answer. Confidence estimation locates the intrinsic boundary by sampling multiple responses from the base model and measuring the accuracy rate with a hybrid evaluator (string matching plus an LLM semantic check), then splitting questions into Known, Mixed, and Unknown using thresholds $\\delta_k$ and $\\delta_{uk}$. The training machinery is Confidence-aware DPO, a loss that weights two preference pairs - 'correct over incorrect' and 'refusal over incorrect' - by the estimated confidence $conf_x$, so high-confidence questions push toward answering and low-confidence ones push toward refusal. The evaluation machinery is a user-centric trustworthiness score $s_{trust} = 2\\cdot Acc + RefR - 1$, which rewards correct answers, rewards refusals over wrong ones, and lets different models be compared without knowing each model's boundary.","core_discovery":"The paper's central claim is that an MLLM should refuse exactly when a question crosses either of two information boundaries, and that this refusal behavior is learnable from the model's own signals. The extrinsic boundary separates questions answerable from the image from those that are not, such as questions about absent objects or false premises; the intrinsic boundary separates what the model's own knowledge and perception can supply from what they cannot. The paper claims that both boundaries can be taught with data generated from any VQA set: confidence estimation for the intrinsic side and GPT-4o-generated unanswerable questions for the extrinsic side. The reported evidence is that IDK-IT and especially the confidence-aware DPO variant (CA-DPO) raise the trustworthiness score above baselines on every benchmark tested, with CA-DPO preserving accuracy while boosting refusal, for example lifting LLaVA1.5-7B's in-domain trustworthiness from -6.50 to 28.50 at a 49.1% refusal rate.","pith_inferences":["The paper does not spell out that the refusal boundary is per-model: applying InBoL to a different base model requires regenerating the IDK data with that model's own confidence estimates, so the framework transfers as a procedure, not as a fixed refusal line.","A natural extension of the linear-probing result is to condition generated refusals on whether the missing information is visual or knowledge-based, and to score the resulting explanations with the same user-centric metric; the paper lists such explanatory refusals as future work.","Because the CA-DPO loss weights the answer-versus-refusal trade-off through a single confidence factor, varying that factor should produce a predictable monotone sweep of helpfulness against caution, which would let deployers set the refusal threshold to the cost of error in their application.","The framework's ceiling is set by the calibration of the sampling-based confidence estimate; replacing it with a calibrated uncertainty measure (ensembles, logit-based tests) would be a direct upgrade path that the paper does not explore."],"forward_implications":["Training with IDK-IT and CA-DPO raises the trustworthiness score on all out-of-domain benchmarks tested (general VQA, knowledge-intensive VQA, and unanswerable VQA) for both LLaVA1.5-7B and 13B, so the effect is not confined to the training distribution.","Trained models refuse more on low-confidence questions and answer more on high-confidence ones, and their accuracy on the questions they do answer exceeds the original model's, indicating that boundary awareness improves both caution and usefulness.","On unanswerable out-of-domain sets (VizWiz-unanswerable, filtered VQAv2-IDK, BeyondVisQA), refusal rates jump from single digits to roughly 70-81%, showing the model can recognize the extrinsic boundary.","CA-DPO outperforms plain DPO trained on the same preference data, implying that the confidence-weighted balancing is what drives the improvement rather than the preference data alone.","The data-generation pipeline works with an open-source generator (Qwen2-VL-72B) as well as GPT-4o, so the framework does not depend on a closed API."],"supporting_citations":[{"why":"Supplies the confidence-estimation-by-sampling method and the IDK dataset construction paradigm that the pipeline follows.","marker":"Cheng et al., 2024"},{"why":"Provides the user-preference value function that the trustworthiness objective and score are built on.","marker":"Xu et al., 2024"},{"why":"Prior honesty-alignment work whose confidence-estimation approach is used to label Known, Mixed, and Unknown questions.","marker":"Yang et al., 2023"},{"why":"The DPO loss that CA-DPO modifies by weighting preference pairs with the estimated confidence.","marker":"Rafailov et al., 2024"},{"why":"Provides the VQAv2-IDK dataset used to evaluate refusal on ambiguous or unanswerable questions, and a prior refusal-training baseline.","marker":"Cha et al., 2024"},{"why":"The VizWiz dataset's unanswerable subset is used to test the model's awareness of the extrinsic information boundary.","marker":"Gurari et al., 2018"},{"why":"Provides the BeyondVisQA subset and the self-awareness framing used for out-of-domain unanswerable evaluation.","marker":"Wang et al., 2024b"},{"why":"Source of VQA v2, one of the three datasets the data-construction pipeline runs on.","marker":"Goyal et al., 2017"}],"fun_headline_variants":["MLLMs learn to refuse, trust score jumps 35 points","Refusal-aware training boosts MLLM trust by 35 points","Teaching AI to say 'I don't know' lifts trust score 35 pts","InBoL: MLLMs refuse when info is lacking, trust up 35","Refusal training improves MLLM trustworthiness by 35 points"],"cache_read_input_tokens":20992,"weakest_assumption_plain":"The framework rests on the assumption that the accuracy of the base model's sampled outputs, judged by a hybrid string-matching-plus-LLM checker, truthfully measures whether the model can answer the question; if that estimate is biased, the learned refusal boundary is trained at the wrong place.","fun_headline_variants_meta":{"raw":{"variants":["MLLMs learn to refuse, trust score jumps 35 points","Refusal-aware training boosts MLLM trust by 35 points","Teaching AI to say 'I don't know' lifts trust score 35 pts","InBoL: MLLMs refuse when info is lacking, trust up 35","Refusal training improves MLLM trustworthiness by 35 points"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000317,"raw_usage":{"total_tokens":1789,"prompt_tokens":940,"completion_tokens":849,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":556,"completion_tokens_details":{"reasoning_tokens":751}},"tokens_in":556,"tokens_out":849,"duration_ms":6225,"temperature":1.0,"reasoning_tokens":751,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T15:11:24.489561+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Give the pipeline a model whose confidence is demonstrably miscalibrated on a controlled topic (for example, images modified so that answerability changes while the question text stays identical), and check where IDK-IT and CA-DPO draw the refusal boundary; if refusals track the biased confidence rather than true answerability, the learned boundary is an artifact of the estimate. A cheaper check on the quantitative claim: recompute the in-domain trustworthiness scores from the reported Acc and RefR values with the paper's own formula $s_{trust} = 2\\cdot Acc + RefR - 1$, since the published in-domain numbers do not reproduce and the magnitude of the gain should be re-verified with the released evaluation code or per-sample scores.","supporting_citations":[],"review_version":1}