{"id":"bf415022-f51f-42eb-982b-7f293668d301","arxiv_id":"2504.12187","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Transformer LLMs can meet Davies's tacit knowledge conditions if the 'same syntax' requirement is replaced by semantic closeness in embedding space.","lead":"This philosophy paper argues that large language models can hold tacit knowledge, the kind of knowledge that guides behavior without being written out as explicit rules. It proposes that the way transformer models group similar word meanings in their internal maps makes them qualify for this older philosophical category.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's core bridge—embedding proximity implies a shared causal factor—is undemonstrated; ROME's paraphrase generalization may track token-level subject keys rather than semantic similarity.","rationale":"The reader's conditional verdict identifies the right region of the argument: the paper's principal move is to weaken Davies's syntactic-structure constraint to embedding proximity and then to read Meng et al.'s ROME results as evidence of causal systematicity. My stress-test sharpens that concern rather than replacing it. The load-bearing assumption is not merely that embeddings cluster similar words; it is that this clustering makes semantically similar inputs share a causal processing pathway. The paper gives a conceptual argument for this in Section 4, but the cited empirical support (Sections 5.1-5.2) is compatible with a much weaker explanation: ROME edits a key at the subject's last token, and paraphrase generalization may be due to token overlap and the same MLP key being engaged, not to the embedding layer's semantic categorizer. Hase et al. (2023), acknowledged in Section 5.3, further undercuts the move from successful editing to a unique causal common factor. None of this requires rejecting the paper's framing; the body is appropriately hedged as preliminary, and the philosophical proposal could still be viable. But as it stands, the bridge from embedding proximity to a Davies-style shared causal property is asserted rather than established, so the reader's CONDITIONAL verdict is appropriate. A concrete causal-intervention test on lexically disjoint paraphrases would settle whether the proposed weakened constraint has the intended force.","tokens_in":27118,"tokens_out":6661,"duration_ms":77483,"concrete_test":"Run a ROME-style edit on GPT-J for a fact such as 'The Eiffel Tower is in Paris', then test generalization on a set of paraphrases that preserve the proposition but share no content words with the original subject phrase (e.g., 'The capital of France is Paris' vs. 'The seat of the French government is Paris', or translations), while also testing edits after randomizing subject token embeddings. If generalization collapses when the subject token is not repeated or when the embedding is perturbed, the effect is keyed to token identity, not to semantic similarity, and the weakened syntactic-structure constraint is not supported. If generalization persists across lexically disjoint paraphrases, the concern is answered.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires Davies's weakened syntactic-structure constraint to do real work: semantically similar inputs must be recognized and processed by one causal common factor. Section 4 asserts that because word embeddings place similar words near each other, the embedding layer acts as a semantic categorizer that routes similar inputs through similar processing pathways. This conflates static word-embedding geometry with whole-input representations. For multi-word prompts, semantic equivalence between paraphrases does not guarantee proximity of their sequence-level embeddings, and proximity in vector space is not by itself a Davies-style shared property that determines causal processing. The ROME evidence in Section 5 is not independent support: the edit is applied at the last subject token, and the reported generalization is measured on prompts that share that subject noun phrase. Those prompts are likely to activate the same or nearby MLP keys regardless of any general semantic categorizer. The paper itself cites Hase et al. (2023), who found that edits at many different locations are similarly effective, so intervention success does not uniquely localize a causal common factor. Thus the existence of a shared causal factor for semantically similar inputs, as opposed to shared token-level keys, is the least secure premise in the argument.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper argues that transformer-based LLMs can acquire tacit knowledge in the sense defined by Martin Davies (1990), despite Davies's own claim that connectionist networks cannot satisfy the relevant constraints. The argument proceeds by endorsing Davies's constraints of semantic description and causal systematicity, while weakening the syntactic-structure constraint: instead of requiring identical input-layer patterns for semantically similar inputs, the paper proposes that the embedding layer's property of mapping semantically similar inputs to proximate regions in vector space suffices. The empirical support comes from Meng et al.'s (2022) ROME experiments, in which edits to MLP key-value pairs generalize to semantically similar prompts and show specificity, which the paper interprets as evidence that MLP modules function as Davies-style causal common factors. The paper concludes that tacit knowledge is a promising conceptual framework for describing and intervening on LLM internals, while acknowledging that the evidence is preliminary and that polysemanticity and localization concerns remain.","tokens_in":27328,"tokens_out":2190,"duration_ms":24497,"significance":"If the argument succeeds, the paper makes a valuable interdisciplinary contribution: it gives a precise, philosophically grounded vocabulary for attributing knowledge to LLMs and connects Davies's classic account to contemporary mechanistic interpretability. The paper is also commendably candid about its limitations, explicitly flagging the preliminary status of the evidence, Hase et al.'s replication caveats, and polysemanticity concerns. The paper's proposal that MLP key-value pairs, as revealed by causal tracing and ROME, can be interpreted as causal common factors is a concrete, potentially fruitful bridge between philosophy of cognitive science and interpretability research.","major_comments":[{"comment":"The central load-bearing step is the claim that proximity in embedding space is functionally equivalent to Davies's shared syntactic property, so that semantically similar inputs are recognized and processed by a shared causal common factor. This equivalence is asserted, not demonstrated. Proximity is a static geometric property of input representations, whereas Davies's constraint requires a property that determines causal processing downstream; the paper does not show that nearby inputs are routed through the same MLP keys, attention heads, or other causal structures. For multi-word prompts, semantic similarity between paraphrases does not by itself guarantee proximity of their contextualized, sequence-level embeddings, and the cited evidence (Abdou et al. 2021; Grand et al. 2022) concerns word-level semantic properties rather than the causal processing of full inputs. Because the weakened constraint is what allows the paper to bypass Davies's own objection, this gap is load-bearing for the central claim.","section":"Section 4, weakening the syntactic-structure constraint"},{"comment":"The ROME evidence does not clearly distinguish a Davies-style semantic common factor from token-level key matching. Figure 4 and the accompanying text state that the edit is applied at the last token of the subject, and Meng et al.'s generalization tests use paraphrases that retain that subject noun phrase. Such prompts may activate the same or nearby MLP keys regardless of any general semantic categorizer at the embedding layer, so the observed generalization is consistent with a much narrower causal structure than the one required for tacit knowledge. The paper needs to address this alternative explanation explicitly, for example by discussing whether paraphrase generalization persists when the subject is lexically varied while the semantic content is preserved.","section":"Section 5.2, ROME evidence as support for causal common factors"},{"comment":"The paper cites Hase et al. (2023) as showing that edits at many different locations have similar efficacy, and then says this means causal tracing might not reliably localize the relevant causal common factors. This concession is in tension with the paper's earlier claim that MLP modules are the causal common factors. If equally effective edits can be made at arbitrary locations, the intervention-based evidence for a unique, semantically structured causal common factor is weakened. The paper should clarify how, given Hase et al.'s findings, the ROME results still provide evidence specifically for Davies-style causal systematicity rather than for a more distributed or redundant causal structure that does not mirror semantic patterns.","section":"Section 5.3, implications of Hase et al. for causal localization"}],"minor_comments":[{"comment":"The abstract says the paper 'demonstrates' that architectural features of LLMs satisfy the constraints, but Section 5.3 repeatedly characterizes the evidence as preliminary and notes several caveats. The wording should be aligned with this more cautious framing, for example by saying 'argues' and 'provides preliminary evidence'.","section":"Abstract and Introduction"},{"comment":"The phrase 'semantically similar inputs are close together in the high-dimensional vector space also guides the network towards similar processing pathways' is stated without a citation or argument; either provide supporting evidence or mark this as a hypothesis to be tested.","section":"Section 4, paragraph beginning 'Moreover, the fact...'"},{"comment":"The figure caption uses 'A)' and 'B)' inconsistently with the body text, which uses 'a' and 'b' in figure 2; standardize the caption formatting for clarity.","section":"Section 3.3.2, Figure 3"},{"comment":"The footnote clarifying that Davies's 'syntax' differs from linguistic syntax is helpful, but it would be useful to also note that the paper's use of 'syntactic structure' throughout refers to this Davies-specific sense, especially in Section 4 where the term is weakened.","section":"Footnote 4"},{"comment":"The reference list appears complete, but some entries (e.g., Meng et al. 2022) lack page numbers or DOIs while others include them; standardizing the reference format would improve readability.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is philosophically serious and engages well with the technical literature, and the author is appropriately modest about the strength of the evidence. The main risk is that the empirical evidence from Meng et al. is being asked to do more conceptual work than it can support, given the token-level nature of the edit and the Hase et al. replication. If the author can either provide a more direct argument for the embedding-proximity-to-shared-processing link or explicitly narrow the claim to a weaker, token-level notion of causal common factors, the paper would be publishable. I would not recommend rejection, because the central framing is novel and potentially useful."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper does something genuinely new: it applies Martin Davies's 1990 account of tacit knowledge to transformer-based LLMs, working through the constraints with clarity. The key move is weakening Davies's syntactic-structure constraint to embedding proximity, and the body is careful to treat the evidence as preliminary. But the central bridge—that proximity in embedding space implies a shared causal processing pathway—is asserted rather than demonstrated, and the ROME evidence doesn't independently support it.\n\nWhat's good: the paper is honest. It flags the tentative status of the empirical support, cites Hase et al.'s replication caveats, and acknowledges polysemanticity. It distinguishes conceptual knowledge attribution from behavioral conformity. The prose is direct and the argument structure is visible. The citation pattern is fair—no self-citations, and it engages both the classic philosophy and the recent interpretability literature.\n\nThe soft spot is the load-bearing weakening of syntactic structure. Davies's original constraint required a shared property in the input that determines causal processing. The paper replaces this with embedding proximity. But word-embedding geometry and whole-input causal processing are different things. Two paraphrases don't necessarily sit close in sequence-level embedding space, and proximity in vector space doesn't by itself ensure that one causal common factor mediates all those transitions. The ROME evidence doesn't close that gap: the edit is applied at the last subject token, and generalization is tested on prompts that share that token. That could be token-level generalization rather than semantic categorization. The paper cites Hase et al., who found similar edit efficacy from many locations, which undercuts the claim that the intervention localizes a causal common factor.\n\nThere's also a definitional circularity, though not a fatal one. The constraint is weakened specifically so LLMs can satisfy it. The author is upfront about this, but the reader has to decide whether the weakened constraint retains Davies's motivation. I think it partially does, but the argument would be stronger if it explained why proximity is sufficient for causal systematicity, rather than just a practical approximation.\n\nThat said, the paper is a serious contribution. It's written for philosophers of AI and for XAI researchers who want principled vocabulary for internal representations. It's not an empirical paper; the data come from Meng et al. The central claim is plausible but not established. A good referee can push the author to sharpen the causal claim and separate it from token-level regularities.\n\nI would send this out for peer review, not desk-reject. It's coherent, well-informed, and it fills a real gap. My own verdict would be conditional pending a stronger defense of the embedding-proximity bridge, but the paper deserves referee time.","headline":"A serious, clearly-written philosophy paper that applies Davies's tacit knowledge to LLMs via a weakened syntactic-structure constraint; the central bridge is asserted more than shown, but the paper deserves peer review.","tokens_in":27807,"tokens_out":3676,"would_cite":true,"duration_ms":36291,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper argues that at least some large language models can acquire tacit knowledge in Martin Davies's sense, because word embeddings supply the needed shared structure and ROME-style edits reveal causal common factors.","keywords":["tacit knowledge","large language models","Davies's constraints","word embeddings","causal common factors","model editing","ROME","semantic knowledge"],"falsifier":"Run a controlled edit experiment on a current LLM: select one hundred pairs of paraphrases whose embeddings are close, apply a rank-one edit to one member of each pair, and measure whether the other member's output changes; if close-embedding paraphrases often fail to co-update while the edited prompt itself changes, the claimed link between embedding proximity and a shared causal common factor is falsified.","tokens_in":26912,"feed_emoji":"🧠","tokens_out":9787,"duration_ms":88195,"temperature":0.7,"pith_summary":"This paper argues that at least some large language models can be said to know things in a real, if implicit, sense: they can acquire tacit knowledge as Martin Davies defined it, meaning rules that are not explicitly represented but that exist as causally relevant internal structures guiding behavior. This matters because it offers a middle path between saying LLMs merely conform to rules and saying they hold explicit knowledge, and it gives a principled target for intervention: edit the internal structure, and the whole family of related predictions changes. Davies himself concluded that connectionist networks fail the syntactic-structure constraint, but the paper contends that transformer LLMs meet a weakened version: the embedding layer maps semantically similar inputs to nearby vectors, which plays the functional role of Davies's shared syntactic property. The paper then reads the generalization and specificity results of ROME model-editing experiments as preliminary evidence that MLP key-value pairs act as Davies-style causal common factors, so that tacit knowledge is a workable conceptual tool for explaining and changing LLM behavior.","feed_headline":"LLMs know facts as tacit knowledge, paper argues","feed_subtitle":"Similar words land close; edits ripple across whole fact clusters, passing a 1990 test for knowledge.","key_machinery":"The paper's load-bearing machinery is Davies's three-constraint test for tacit knowledge—semantic description, causal systematicity, and syntactic structure—together with two transformer components that let the test run on modern LLMs. The embedding layer is the key adaptor: by placing similar words and contexts near each other in a high-dimensional vector space, it performs the function Davies assigned to a shared syntactic property, so semantically similar inputs can engage the same causal processing without sharing an identical activation pattern. The candidate causal common factor is the MLP key-value pair, a feed-forward module in an intermediate layer that stores factual associations; when edited with the rank-one model editing (ROME) procedure, it updates predictions across paraphrases, which is the intervention signature expected of one causal structure mediating a whole family of related transitions. Generalization and specificity metrics from such edit experiments are the operational test for whether a candidate structure is a genuine causal common factor or merely a memorized one-to-one mapping.","core_discovery":"Stated on its own terms, the paper's claim is that Davies's account of tacit knowledge can be transferred to transformer LLMs with one modification, and that at least one current model appears to satisfy it. Semantic description holds because LLM inputs and outputs are word sequences with conventional meaning. Causal systematicity holds provisionally because the MLP key-value pairs localized by causal tracing in GPT-J behave the way Davies-style causal common factors should: a rank-one edit to one pair updates the prediction not only for the edited prompt but for semantically related paraphrases, while leaving unrelated prompts mostly unchanged, which is the generalization-and-specificity signature of a shared causal mechanism. The syntactic-structure constraint, which Davies thought distributed networks could not meet because context alters activation patterns, is weakened so that proximity in the embedding space replaces the requirement of an identical shared property; the embedding layer acts as a semantic categorizer that routes similar inputs through similar processing. The conclusion is that LLMs can meet all three modified constraints and that the GPT-J results give preliminary evidence that some LLMs have actually done so.","pith_inferences":["The argument is architecture-general enough that the framework should transfer to any learned-embedding transformer, which makes tacit knowledge a testable empirical hypothesis for every new LLM release rather than a one-model finding.","A sharper prediction follows that the paper does not test: within a paraphrase cluster, edit generalization should track embedding distance, so pairs with closer embeddings should co-update more reliably; if they do not, the embedding route is not the causal carrier and the weakened constraint would need revision.","The Hase et al. replication caveat points to an extension: causal common factors may be distributed across layers, so multi-layer interventions could either edit tacit knowledge more robustly or reveal that the 'factor' is a network-level structure rather than a single MLP module.","If tacit knowledge is accepted as a causal-explanatory structure, the philosophical dispute about whether LLMs 'really' know shifts from questions of justification and conceptual content to the empirical question of whether an internal structure plays the right causal role."],"forward_implications":["Davies's original verdict that connectionist networks cannot have tacit knowledge does not carry over to transformer LLMs, because the embedding layer supplies the missing shared-property mechanism in weakened form.","Fact-like behavior counts as stored knowledge only when a causal common factor can be located and shown to mediate the whole semantic family, giving researchers a principled distinction between knowledge and memorization.","Editing methods gain a theoretical rationale: generalization and specificity scores become operational measures of Davies-style causal systematicity rather than purely practical metrics.","The framework gives a systematic target for correcting bias, hallucination, and misinformation: edit the causal common factor behind a cluster of related predictions instead of patching individual prompts.","Because the argument for the weakened syntactic-structure constraint is architectural, it predicts that newer LLMs beyond GPT-J will show the same intervention pattern if the attribution is to hold."],"supporting_citations":[{"why":"Supplies the three-constraint account of tacit knowledge that the paper adapts, and the original claim that connectionist networks fail it.","marker":"Davies (1990)"},{"why":"Provides the causal-tracing and ROME intervention results used as preliminary evidence that MLP key-value pairs behave like causal common factors.","marker":"(Meng et al. 2022)"},{"why":"Establishes that transformer MLP layers act as key-value memories, the internal structure the paper identifies as the candidate causal common factor.","marker":"(Geva et al. 2021)"},{"why":"Grounds the claim that embeddings place similar words near each other in vector space, the basis of the weakened syntactic-structure constraint.","marker":"(Mikolov et al. 2013)"},{"why":"Formulates the dimension-shift objection that motivates the paper's weakening of the syntactic-structure constraint.","marker":"(Smolensky 1988)"},{"why":"Raises the caveat that intervention localization may not reliably identify the right causal common factor, a limitation the paper explicitly weighs.","marker":"(Hase et al. 2023)"},{"why":"Supplies recent evidence of interpretable semantic features in LLM activations, supporting the paper's claim that causally relevant internal structure can carry semantic description.","marker":"(Templeton et al. 2024)"}],"fun_headline_variants":[],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that proximity in embedding space is functionally equivalent to sharing a syntactic property, so that all semantically similar inputs are processed by one causal common factor; if close embeddings can still feed different or overlapping causal pathways, the causal-systematicity constraint is not met.","fun_headline_variants_meta":{"error":"Client error '402 Payment Required' for url 'https://api.deepseek.com/chat/completions'\nFor more information check: https://developer.mozilla.org/en-US/docs/Web/HTTP/Status/402"},"cache_creation_input_tokens":0},"created_at":"2026-08-16T12:35:09.406548+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a controlled edit experiment on a current LLM: select one hundred pairs of paraphrases whose embeddings are close, apply a rank-one edit to one member of each pair, and measure whether the other member's output changes; if close-embedding paraphrases often fail to co-update while the edited prompt itself changes, the claimed link between embedding proximity and a shared causal common factor is falsified.","supporting_citations":[],"review_version":1}