{"id":"a6a5b2ca-3c58-4d84-ae0f-4170b8cdd47a","arxiv_id":"2411.17992","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"The thesis shows that randomly masking input tokens during fine-tuning makes post-hoc explanations of NLP models consistently faithful under an erasure-based faithfulness metric.","lead":"This PhD thesis proposes two new ways to make AI explanations trustworthy: faithfulness measurable models, trained so their explanations can be cheaply and reliably checked, and self-explanations by large language models. It finds that faithfulness is usually model- and task-dependent, but simple training modifications can make explanations consistently faithful.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Near-optimal faithfulness claim is only as strong as the RACU erasure proxy; FMM experiments lack an independent ground-truth check that high-RACU explanations track real reasoning.","rationale":"The reader's weakest_assumption correctly identifies the erasure-metric proxy as the foundation of the faithfulness guarantee, and I agree with the conditional verdict. My concern sharpens this: the key missing evidence is not just a philosophical point about proxies, but an empirical check that RACU-optimal explanations recover known causal structure in the FMM setting. The Chapter 3 synthetic validation is reassuring for the metric on a linear problem, but it does not transfer automatically to masked-fine-tuned non-linear models, where training changes could make erasure performance high for reasons unrelated to the model's internal reasoning. The MaSF validation addresses out-of-distribution masking but not the gap between RACU and ground truth. I am not claiming the FMM result is wrong; the peer-reviewed empirical work is substantial and the mechanism is plausible. The concern is scoping: the headline claim should be framed as near-optimal relative to the RACU definition unless a ground-truth synthetic check is added. Because the thesis already includes limitations acknowledging the proxy and because the requested condition is a strengthening of evidence rather than a demonstration of internal inconsistency, I do not move the verdict; I recommend retaining the conditional acceptance with that condition made explicit.","tokens_in":51730,"tokens_out":7490,"duration_ms":78882,"concrete_test":"Construct a synthetic nonlinear classification task with known ground-truth token importance: generate inputs from a known rule (e.g., sentiment depends on a conjunction or disjunction of two token sets, with distractor and redundant tokens), and train a RoBERTa-base FMM with masked fine-tuning as in §4.2.1 plus a plain fine-tuned baseline on the same data. Compute RACU for each explanation method (Gradient, IG, LOO, Beam) as in Table 4.3, and separately compute the Spearman rank correlation and top-k overlap between each explanation's token ranking and the known ground-truth causal token weights. If FMM explanations show high RACU but no better ground-truth recovery than baseline explanations, the near-optimal-faithfulness claim is an artifact of the erasure metric; if RACU and ground-truth recovery move together across methods, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that FMMs yield near-theoretically-optimal faithful explanations rests entirely on the erasure metric operationalized as RACU (Ch. 3, eq. 3.4; Ch. 4, §4.1.4). The thesis explicitly acknowledges that the model's true reasoning is unknown and that erasure-based faithfulness is a proxy, which is honest but also identifies the load-bearing point: nothing in the FMM experiments independently checks whether explanations scoring high on RACU correspond to the model's actual cause structure. The synthetic validation in §3.2.3 shows Recursive ROAR recovers known features for a linear problem, but the FMM experiments use non-linear RoBERTa models and report RACU against random-mask baselines; no ground-truth token importance is available there. Masked fine-tuning is exactly the kind of training change that could make erasure-based faithfulness easy to achieve, for example by making predictions depend on individually removable surface cues, without making explanations reflect higher-order or interactive reasoning. RACU certifies only that removing the top-ranked tokens hurts more than removing random tokens. The MaSF in-distribution check (§4.1.3) reduces the confound of OOD degradation masquerading as importance, but it does not validate that RACU-optimal explanations track the model's actual reasoning, nor does it certify all masked inputs along the full beam-search trajectory. The abstract's wording 'near theoretical optimal' is therefore stronger than what is demonstrated: the evidence supports near-optimal performance relative to the RACU proxy, not unqualified faithfulness.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This PhD thesis proposes two new interpretability paradigms: faithfulness measurable models (FMMs) and self-explanations. Chapter 3 develops Recursive ROAR and the RACU metric for erasure-based faithfulness of importance measures. Chapter 4 introduces masked fine-tuning to make erasure evaluation in-distribution and a Beam search that optimizes RACU, reporting that FMMs achieve near-optimal RACU and consistent faithfulness across tasks and models. Chapter 5 evaluates self-explanations from instruction-tuned LLMs via self-consistency checks and finds faithfulness to be model-, task-, and explanation-dependent. The thesis concludes that post-hoc and intrinsic explanations are model- and task-dependent by default, while FMMs yield consistently faithful explanations.","tokens_in":52040,"tokens_out":5760,"duration_ms":47964,"significance":"The FMM idea—reformulating interpretability as designing models for cheap and reliable faithfulness measurement rather than architectural explainability—is a valuable contribution. The thesis benefits from extensive experiments (eight datasets, RoBERTa-base/large and BiLSTM-Attention, confidence intervals, MaSF out-of-distribution checks, and synthetic ground-truth validation for Recursive ROAR) and is honest about limitations. If the central claim is accepted, it would give practitioners a recipe for obtaining post-hoc token-importance explanations that are consistently faithful under the erasure metric without sacrificing predictive performance. However, the headline claim goes beyond what is demonstrated, because faithfulness is operationalized solely through RACU and the Beam optimizer targets that same metric.","major_comments":[{"comment":"The abstract's claim of 'near theoretical optimal in terms of faithfulness' is stronger than the evidence. The Beam method in §4.1.5 explicitly optimizes an explanation to maximize the RACU score defined in Eq. (3.4), and Table 4.3 reports that same RACU score; therefore the near-optimal result is partly a consequence of optimizing and evaluating on the same objective. Please rephrase the claim as 'near-optimal under the RACU erasure metric,' or provide evidence that high-RACU explanations correspond to the model's actual reasoning independently of this metric.","section":"Abstract; §4.1.5; Table 4.3"},{"comment":"The synthetic ground-truth validation in §3.2.3 uses a linear problem, while the FMM experiments use non-linear RoBERTa models; the thesis acknowledges in §3.6 and §4.4 that the model's true reasoning is unknown and that erasure is a proxy. Without an independent check that high-RACU explanations track true token importance for these non-linear models, the statement that FMMs yield 'consistently faithful explanations' is established only with respect to the erasure proxy, not with respect to faithfulness as 'reflecting the model's reasoning' as defined in Chapter 2.","section":"§3.2.3; §4.2.3; §3.6; §4.4"},{"comment":"The MaSF in-distribution validation is performed for masked inputs at evaluation time, but the Beam search in §4.1.5 generates many partially masked inputs along its trajectory that are not all certified by the MaSF check. Because RACU's validity rests on masked inputs being in-distribution, an out-of-distribution response during the search could inflate RACU even for tokens that are not causally important. The authors should either run the MaSF check on the full set of inputs explored by Beam, or restrict the claim to the subset of masked inputs verified to be in-distribution.","section":"§4.1.3; §4.1.5"}],"minor_comments":[{"comment":"The word 'intrisic' is a typo and should read 'intrinsic'.","section":"Abstract"},{"comment":"The column header 'defintion' should be 'definition'.","section":"Table 1.1"},{"comment":"The sentence 'this paradigm archives the goal of taking the best part from both paradigms' should use 'achieves' instead of 'archives'.","section":"§1.4.1"},{"comment":"The phrase 'randomly masking the training dataset' is imprecise: the method is a two-stage masked fine-tuning procedure with specific masking ratios and a mixed-strategy schedule, which matters for reproducibility.","section":"Abstract; §4.1.2"},{"comment":"The masked fine-tuning hyperparameters (masking ratio and schedule) are not analyzed for sensitivity; documenting this would strengthen the practical claim that 'simple modifications' to the model suffice.","section":"§4.2.1"}],"recommendation":"major_revision","confidential_remarks":"This is a PhD thesis compiled from several published papers with already peer-reviewed components. The editor should consider whether the arXiv version's abstract and headline claims match the more carefully scoped peer-reviewed results, and whether the journal's audience expects a thesis-style multi-chapter manuscript or a more focused article."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere's the short version: this is a PhD thesis that stitches together four already-published papers (EMNLP Findings 2022, ICML 2024, ACL Findings 2024, ACM Computing Surveys) around the idea of 'faithfulness measurable models' (FMMs). The technical core is not new, but the synthesis and the paradigm framing are. The empirical work is genuinely extensive: eight datasets, two architectures, confidence intervals, MaSF out-of-distribution checks, and a synthetic ground-truth validation for the Recursive ROAR metric. That part deserves credit; it is careful, reproducible work that has already survived peer review.\n\nThe main soft spot is exactly where the reader put the finger: the headline claim that FMMs produce explanations that are 'near theoretical optimal in terms of faithfulness' is measured with RACU, the same metric that the Beam search optimizer explicitly maximizes. So the near-optimality is partly by construction. The thesis is honest that erasure-based faithfulness is a proxy—it says the model's true reasoning is unknown—but the abstract's wording goes beyond that honesty. The stress-test note is right that nothing in the FMM experiments independently checks whether high-RACU explanations correspond to the model's actual cause structure; the synthetic validation is on a linear problem, and the RoBERTa experiments only compare against random-mask baselines. So a reader should read 'near theoretical optimal' as 'near optimal with respect to this particular erasure proxy.'\n\nAnother minor point: the self-consistency check in Chapter 5 is a consistency test, not a causal faithfulness proof. The chapter acknowledges this in its limitations, but it's worth flagging that 'faithfulness' there is even more indirect.\n\nNet: if this is submitted as a journal paper or thesis defense document, it deserves a serious referee. The underlying experiments are solid, the FMM idea is plausible, and the limitations are mostly acknowledged. But the abstract should be reworded to say 'near-optimal under the RACU metric,' and the authors should add a paragraph explaining what would be needed to ground-truth the FMM explanations (e.g., mechanistic probes or controlled task structures). That revision is doable, and the paper would be stronger for it.\n\nFor a reading group: maybe. If you want the paradigm discussion and a compact review of faithfulness metrics, it's a useful read. If you want the actual methods, read the ICML paper and the EMNLP Findings paper directly. I'd cite those, not the thesis, in my own work.\n\nRecommendation: engage with it—send it to referees, ask for the claim-softening and a ground-truth discussion. But don't let the 'near theoretical optimal' phrasing survive as is.","headline":"Solid thesis packaging solid prior work into a provocative but imperfect paradigm; the 'near-optimal faithfulness' claim needs a metric-relative disclaimer.","tokens_in":52567,"tokens_out":3129,"would_cite":false,"duration_ms":28701,"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":"This Ph.D. thesis claims that faithfulness-measurable models, built by randomly masking tokens during fine-tuning, yield token-importance explanations that are near-theoretically-optimal faithful under erasure, without architectural…","keywords":["faithfulness","interpretability","importance measures","masked fine-tuning","erasure metric","self-explanations","natural language processing","post-hoc explanations"],"falsifier":"Take a trained FMM on a task with known ground-truth important tokens, for instance a synthetic dataset where only certain tokens determine the label. If the faithfulness-optimized explanation fails to recover those tokens, or if the model's performance after masking its top tokens is not below the random-masking baseline, the central claim of near-optimal faithfulness would be refuted. A simpler check: find any dataset where masked fine-tuning leaves out-of-distribution p-values below the 5% threshold, showing masked inputs are still out-of-distribution.","tokens_in":51534,"feed_emoji":"🔍","tokens_out":5607,"duration_ms":48897,"temperature":0.7,"pith_summary":"This thesis tries to establish that faithfulness—whether an explanation reflects the model's actual reasoning—can be engineered rather than hoped for. It argues that the two standard paradigms, intrinsic and post-hoc, are both unproductive: post-hoc explanations are often no better than random, and intrinsic models sacrifice generality or overstate their interpretability. The proposed alternative, faithfulness measurable models (FMMs), reframes the goal from \"design a model that can be explained\" to \"design a model for which faithfulness can be measured cheaply and reliably.\" With masked fine-tuning, removing tokens becomes an in-distribution operation, so the erasure metric can be applied directly; explanations can then be optimized toward maximum faithfulness. The central empirical claim is that FMMs give near-theoretical-optimal faithfulness on synthetic tasks and consistently faithful explanations across real tasks, whereas the same post-hoc methods on plain models are model- and task-dependent.","feed_headline":"Masked training makes NLP explanations faithful near limit","feed_subtitle":"A simple fine-tuning tweak makes token-importance explanations consistently faithful under erasure checks.","key_machinery":"The central mechanism is masked fine-tuning: during training, random input tokens are replaced with a mask token, so the model learns to treat masked inputs as ordinary in-distribution inputs. This is what unlocks the erasure metric, because removing allegedly important tokens no longer sends the input out of distribution. On top of this, beam search over token subsets optimizes an importance measure toward maximal faithfulness, turning the FMM into an indirectly self-explaining model without any architectural constraint. The faithfulness score is the relative area between curves, which compares the performance loss from removing the explanation's top tokens against removing tokens at random.","core_discovery":"The discovery is that a small, architecture-free change to training—randomly masking input tokens during fine-tuning—makes the faithfulness of token-importance explanations measurable and optimizable. The thesis defines an FMM as a model built so that the erasure faithfulness metric (\"if a token is truly important, removing it should hurt the prediction\") is cheap and reliable, by making masked inputs in-distribution. It validates this with an out-of-distribution test and then optimizes explanations with beam search. Measured by the relative area between curves, FMMs reach near theoretical optimal faithfulness on synthetic problems and, unlike plain fine-tuning, produce consistently faithful explanations across models, tasks, and explanation methods. The thesis also examines self-explanations from large language models, proposing self-consistency checks, and finds those explanations remain model- and task-dependent.","pith_inferences":["The FMM principle is not tied to token masking in principle; if faithfulness can be made cheap to measure for other explanation types, the same optimize-toward-faithfulness recipe could apply, and the thesis's future-work sketch points at causal language models.","If the erasure metric is accepted as the definition of faithfulness, then FMMs provide a practical audit protocol: before relying on an explanation in a high-stakes application, one can compute its faithfulness score on the deployed model and reject explanations that fall below the random baseline.","The result suggests that interpretability research may be better served by designing for measurability rather than for explainability, since measurability is what makes optimization and guarantees possible without constraining the architecture."],"forward_implications":["Practitioners can get consistently faithful token-importance explanations from a standard transformer by adding random masking to fine-tuning, with no loss in predictive performance.","Faithfulness of post-hoc and intrinsic explanations should no longer be assumed transferable: it is model- and task-dependent unless the model is designed to be faithfulness-measurable.","Because faithfulness is cheap to measure on an FMM, explanations can be audited per model instance before deployment, and optimized toward the theoretical optimum.","Self-explanations from large language models are not generally trustworthy; their faithfulness depends on model, task, and explanation type, so they need per-use validation.","The same base model, datasets, and post-hoc methods become consistently faithful when switched from plain fine-tuning to masked fine-tuning, isolating the training modification as the cause."],"supporting_citations":[{"why":"Supplies the ROAR remove-and-retrain baseline and the random-token comparison that Recursive ROAR and the FMM faithfulness metric build on.","marker":"[14]"},{"why":"Defines the erasure metric that an explanation is faithful when removing important tokens degrades performance relative to removing random tokens.","marker":"[60]"},{"why":"Provides the beam-search method used to optimize an importance measure toward maximum faithfulness on an FMM.","marker":"[35]"},{"why":"MaSF is the out-of-distribution detector used to validate that masked fine-tuning keeps masked inputs in-distribution.","marker":"[46]"},{"why":"RoBERTa is the base model that is fine-tuned with plain and masked fine-tuning in the FMM experiments.","marker":"[4]"}],"fun_headline_variants":["Fine-tune with random masks for near-optimal NLP explanations","Masked fine-tuning yields consistently faithful token explanations","Simple training tweak makes NLP explanations provably faithful","Random masking during fine-tuning ensures faithful explainability","Masked training makes NLP explanations near-optimal in faithfulness"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire faithfulness guarantee rests on the assumption that \"faithful\" means the erasure test: an explanation is faithful if removing its top tokens hurts the model more than removing random tokens, and that masked fine-tuning indeed makes all masked inputs in-distribution.","fun_headline_variants_meta":{"raw":{"variants":["Fine-tune with random masks for near-optimal NLP explanations","Masked fine-tuning yields consistently faithful token explanations","Simple training tweak makes NLP explanations provably faithful","Random masking during fine-tuning ensures faithful explainability","Masked training makes NLP explanations near-optimal in faithfulness"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000817,"raw_usage":{"total_tokens":3604,"prompt_tokens":996,"completion_tokens":2608,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":612,"completion_tokens_details":{"reasoning_tokens":2531}},"tokens_in":612,"tokens_out":2608,"duration_ms":16167,"temperature":1.0,"reasoning_tokens":2531,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:37:18.672914+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a trained FMM on a task with known ground-truth important tokens, for instance a synthetic dataset where only certain tokens determine the label. If the faithfulness-optimized explanation fails to recover those tokens, or if the model's performance after masking its top tokens is not below the random-masking baseline, the central claim of near-optimal faithfulness would be refuted. A simpler check: find any dataset where masked fine-tuning leaves out-of-distribution p-values below the 5% threshold, showing masked inputs are still out-of-distribution.","supporting_citations":[],"review_version":1}