{"id":"b66c276b-6ddd-4290-9e1c-93c100c71154","arxiv_id":"2506.02946","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Counterfactuals for LM agents computed over a high-level abstraction of the action, instead of its tokens, preserve the observed action's meaning across counterfactual contexts far more often than token-level counterfactuals.","lead":"Language-model agents choose actions as strings of text, so counterfactual 'what if' analysis is usually done on individual tokens. This paper instead computes counterfactuals over the meaning of the action, and reports that meaning stays stable across contexts far more often than with token-level methods.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proposition 1's consistency proof averages over the prior P(UY), but the ACF algorithm uses the posterior P(UY|s,y); for deterministic abstractions (Machiavelli) the method therefore cannot transmit the observed y, making Figure 3 inconsistent with Eqs. (3)-(4).","rationale":"The reader's weakest assumption identified the reliance on conditioning on Y alone and the degenerate Machiavelli case, but the reader accepted Proposition 1 as correctly proved. My concern pinpoints a more specific and more damaging formal issue: the proof of Proposition 1 is correct only for a quantity that averages over the prior on UY, not for the posterior-conditioned distribution that the algorithm actually uses. If this concern lands, the paper's central theoretical justification for ACF is vacuous, and the Machiavelli experiment is unexplained by the stated method. I nevertheless keep the reader's CONDITIONAL verdict rather than moving to REJECT, because the underlying abstraction-mediated idea is promising and the issues are addressable in revision: restate or replace Proposition 1 with a property of the posterior-conditioned distribution, specify how deterministic abstractions are handled (e.g., by explicitly targeting Y'=y), release code, and add metrics that do not restate the design objective. The empirical comparisons, especially ST and the qualitative gender-steering example, provide some independent support, but the formal gap and the inconsistent degenerate-case demonstration mean the current submission is not yet acceptable as is. Hence no verdict change from the reader's conditional acceptance posture.","tokens_in":15287,"tokens_out":13573,"duration_ms":159072,"concrete_test":"Re-implement Eqs. (3)-(4) for the Machiavelli setup with degenerate f_Y (deterministic annotation) and U'_Y drawn from P(UY|s,y). Compute P_{A'}(a'|s') for two different observed values y_a and y_b in the same factual state; if the two outputs are identical (as they must be when UY is deterministic), then Figure 3 cannot be reproduced by the stated algorithm. Independently, re-derive Prop 1 replacing the prior P(UY) in the appendix with the posterior P(UY|s,y); exhibit a simple two-action, two-abstraction SCM where the equality fails, confirming the proposition as written is not the relevant statement.","verdict_should_be":"UNCHANGED","load_bearing_attack":"ACF's only theoretical guarantee is Proposition 1, but the proposition is not about the distribution the algorithm computes. In Section 3.1, step 1 sets U'_Y to the posterior P(UY | s,y), and Eq. (4) uses this posterior when forming P_{Y'}(y'|s'). Proposition 1, however, states and proves the identity for U'_Y drawn from the prior P(UY): the appendix's first line is Σ_{u'} P(u')... with no conditioning on y. Averaging the ACF distribution over the prior is interventional by construction, so the result reduces to a tautology and says nothing about whether conditioning on y produces consistent counterfactuals. The defect is not just notational: when f_Y is deterministic (Sec 4.1 says PY is degenerate), UY has no entropy, the posterior equals the prior, and the ACF output in Eqs. (3)-(4) is independent of the observed y. Yet Figure 3 reports that ACF increases probability of actions with the observed annotation 'physical harm:1'. No extra step that would produce such dependence (e.g., setting Y'=y before the mapping-back step) is described or derived. Thus the main formal guarantee is unsupported as stated, and the Machiavelli demonstration is not an instance of the stated method.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Abstract Counterfactuals (ACF), a framework for counterfactual inference on language-model agents. Instead of performing abduction on the token-level Gumbel mechanism, ACF introduces an abstraction variable Y that represents the high-level meaning of an action in a given state, performs abduction on Y's exogenous noise, predicts a counterfactual abstraction Y' in the counterfactual state, and then maps Y' back to an action distribution via Bayes' rule (Eqs. 3-4). The authors claim Proposition 1 as a consistency property: averaging the ACF distribution over the prior of the exogenous noise recovers the interventional distribution. The method is evaluated on three settings: MACHIAVELLI choice-based games with deterministic annotation-based abstractions, Bios biography generation with a gender-steering latent intervention, and GoEmotions with token replacement interventions. The paper reports that ACF maintains much higher semantic consistency than token-level Gumbel counterfactuals (TLCF), with lower Abstraction Change Rate and higher Counterfactual Probability Increase Rate.","tokens_in":15482,"tokens_out":11064,"duration_ms":101743,"significance":"If the central claim were sound, ACF would be a valuable contribution to causal analysis of LM agents: it is black-box, avoids token-level semantic drift, and provides an interpretable abstraction layer. The empirical results are impressive in magnitude (ACR 0.02-0.12 vs 0.32-0.67), and the unsupervised abstraction pipeline using an auxiliary LLM is a useful methodological addition. The paper also correctly identifies real failure modes of token-level counterfactuals. However, the main theoretical guarantee, Proposition 1, is not about the distribution that the algorithm actually computes, and the deterministic-abstraction experiment in MACHIAVELLI appears to degenerate to the interventional policy. These issues directly affect the claim that the approach 'produces consistent and meaningful counterfactuals'. As a result, the paper as it stands does not yet substantiate its central contribution, though the underlying idea remains promising and potentially fixable.","major_comments":[{"comment":"The statement of Proposition 1 concerns the expectation of the ACF distribution over U'_Y drawn from the prior P(UY), but the algorithm's Step 1 derives the posterior P(UY | s, y). The proof's first line, Σ_{u'} P(u') ..., confirms that the proposition is about prior-averaged behavior, not about the posterior-conditioned distribution that the method actually computes. Consequently, the proposition does not establish any consistency property for the output of Eqs. (3)-(4). The defect is load-bearing: in the MACHIAVELLI experiment (Sec. 4.1) P_Y is degenerate, so UY has a point-mass posterior equal to the prior; then Step 2 yields Y' distributed as the interventional P(y'|s'), and Eq. (4) collapses to PA'(a'|s') = PA(a'|s'). The reported behavior in Figure 3 (increased probability of actions with the observed annotation 'physical harm:1') therefore cannot be produced by the method as described; the figure is inconsistent with Eqs. (3)-(4) unless an additional step (e.g., setting Y'=y) is introduced. The main formal guarantee is thus unsupported, and the Machiavelli demonstration is not an instance of the stated method.","section":"Section 3.1 and Appendix A"},{"comment":"The SCM in Sec. 3 posits Yt = fY(At, St, UYt) as a structural node of the agent's data-generating process. In the experiments, however, Y is realized as an external artifact: benchmark annotations in Sec. 4.1, a fine-tuned DistilBERT classifier in Appendix E, and an LLM topic assignment in Appendix F. The paper never provides evidence that these external distributions are faithful stand-ins for the agent's own abstraction mechanism. In particular, the premise that Y is a context-stable semantic proxy (Sec. 3) is asserted but not tested, and the validity of conditioning on Y alone (dropping the observed action a) is not verified. A concrete test would be to compare counterfactual predictions made with the external Y against observed behavior in a held-out counterfactual state, or to measure whether the learned PY is invariant across contexts. Without such validation, the claim that the resulting counterfactuals are 'meaningful' for the agent is not established.","section":"Section 3 and Sections 4.1-4.3"},{"comment":"The implementation of the abduction step is not specified. The algorithm requires computing or estimating the posterior P(UY | s, y) and the quantities P(y | s, u'_Y) in Step 1, but the paper's descriptions of the abstraction models (a fine-tuned DistilBERT outputting P(y | s, a), an LLM providing logits over topic indices) do not explain how an explicit exogenous noise UY is defined or how the posterior is obtained. Without these details the method is not reproducible, and it is unclear whether the reported experimental results actually instantiate the three-step procedure of Section 3.1.","section":"Section 3.1 and Appendices E-F"}],"minor_comments":[{"comment":"The proof's notation conflates the exogenous noise of A and Y and drops the conditioning on s throughout; while the algebraic identity is correct, the notation should be cleaned up to make clear that P(y'|u') means P(y' | s', u').","section":"Appendix A"},{"comment":"The statement that 'we pad the Gumbel noise vector with 0 values when the counterfactual action space is larger than the factual one' is an ad hoc choice for the TLCF baseline; the paper should justify it or at least acknowledge that it is an arbitrary convention affecting the comparison.","section":"Section 4.1"},{"comment":"The ST rows in Tables 1 and 2 report win rates, but the captions do not state this; please clarify how the ST comparison is computed and what the t-statistics correspond to.","section":"Section 4.2 and 4.3"},{"comment":"The notation Pcf(y' | a, s') in the ACR definition is not defined; it should be stated that Pcf is the counterfactual distribution over the abstraction given a counterfactual action and state.","section":"Appendix B"},{"comment":"Since the Machiavelli results are explicitly illustrative, the figure should include a note that the probabilities are point estimates from a single scene and lack error bars or variance information.","section":"Figure 3"}],"recommendation":"major_revision","confidential_remarks":"The Proposition 1 mismatch is a serious issue that likely requires reworking either the algorithm or the theorem. If the authors can show that the posterior-averaged distribution also satisfies an appropriate consistency property (e.g., averaging over the factual y distribution recovers the interventional distribution), the paper could become acceptable. However, the deterministic-abstraction case (Machiavelli) needs to be addressed explicitly, since the current method degenerates to the interventional policy there. I would also encourage the authors to provide a concrete validation of the external abstraction Y as a faithful model of the agent's semantics, as the current leap from SCM abstraction to external classifier is a gap that undermines the claim of meaningful counterfactuals."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea is genuinely new: instead of doing counterfactuals on the token mechanism, condition on a semantic abstraction Y of the action, predict Y' in the counterfactual state, and map back. The three-step procedure is clean and black-box. The paper is also honest about needing to define an abstraction and about compute cost. The comparisons against TLCF on gender steering and token replacement show large differences, and the semantic tightness metric plus the worked gender example are independent evidence that something real is happening.\n\nBut the formal guarantee as stated does not cover the algorithm. Proposition 1 averages over the prior of UY, while the abduction step in Eqs. (3)-(4) uses the posterior. The stress-test note is right: that makes the proposition a prior-averaged identity, not a statement about the posterior-conditioned counterfactual. For a deterministic abstraction (MACHIAVELLI), the posterior equals the prior, so the method cannot transmit the observed y; the ACF distribution would be interventional. Yet Figure 3 shows precisely a dependence on the observed abstraction. The implementation must be doing something else (maybe setting Y'=y before mapping back), but it isn't described. That is the paper's most serious soft spot.\n\nTwo more concerns, in proportion. The headline metrics ACR and CPIR measure whether the observed abstraction is preserved—which is exactly what the conditioning step forces. Part of the advantage over TLCF is by construction; the ST metric and the qualitative examples carry more weight than the tables. And the abstraction classifiers (DistilBERT, LLM) are treated as if they were the agent's own f_Y; the paper never checks this fidelity assumption. Minor: no code, no sampling details, no confidence intervals on the main numbers.\n\nWho benefits: anyone working on causal explanations or interpretability of LM agents. The idea is worth airing, and the flaws are addressable. A serious referee should see it, but with the expectation of heavy revision: fix or rescope the proposition, explain or correct the Machiavelli experiment, and add code and error bars.","headline":"New abstraction-level counterfactual idea for LM agents; the consistency proof is misstated and the Machiavelli experiment needs explanation, but the core deserves a serious referee.","tokens_in":16181,"tokens_out":4126,"would_cite":true,"duration_ms":39194,"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":"The paper argues that counterfactual inference for language-model agents should be performed over high-level semantic abstractions of actions, not over the tokens the model outputs, and presents experiments showing this yields far more…","keywords":["abstract counterfactuals","language model agents","structural causal models","counterfactual inference","semantic abstraction","token-level counterfactuals","text generation","choice-based environments"],"falsifier":"Train the abstraction on labels that are known not to reflect the agent's action semantics, or on random labels, and rerun the same comparisons: if the consistency advantage over token-level methods survives, the claimed mechanism—semantic fidelity—is not what drives it.","tokens_in":14909,"feed_emoji":"🤖","tokens_out":7590,"duration_ms":67390,"temperature":0.7,"pith_summary":"Counterfactual questions about language-model agents—\"what would this agent have done in a different situation?\"—are hard because agents act by producing arbitrary strings, and the same token can mean different things in different contexts. This paper proposes asking the question at the level of an abstraction $Y$ that captures the high-level meaning of the action (cautious or reckless, profession, emotion), and then mapping the counterfactual abstraction back to concrete actions. The experiments claim that this abstract procedure keeps the semantic content of the counterfactual aligned with the factual action far more often than token-level methods do, while still delivering the intervention's intended effect, for example changing gender while preserving profession. If the claim holds, ACF gives a black-box method for meaningful \"what if\" analysis of LM agents without inverting their token-generation internals.","feed_headline":"For language agents, what-ifs work better over meaning than tokens","feed_subtitle":"New method anchors what-if questions in high-level action meaning, keeping counterfactual content consistent where token-level methods…","key_machinery":"The load-bearing object is the abstraction variable $Y_t = f_Y(A_t, S_t, U^Y_t)$, a context-dependent semantic summary of the agent's action, together with the mapping-back formula (3) that reweights candidate actions by $P_Y(y' \\mid s', a')$, i.e. by how likely each action is to produce the counterfactual abstraction. The argument is carried by two structural choices: abduction is done on the combined mechanism $f_Y \\circ f_A$ while conditioning only on the observed abstraction value, and the final action is sampled from the posterior $P_{A'}$ that marginalizes over the counterfactual abstraction. Proposition 1 guarantees that averaging these individual counterfactual distributions over the prior noise of $U_Y$ recovers the interventional distribution, which anchors the method to standard causal semantics.","core_discovery":"The paper's central claim is that a counterfactual action for an LM agent can and should be derived through the action's semantic abstraction, not through the token-sampling mechanism. After observing action $a$ in state $s$ and its abstraction $y = f_Y(a,s)$, the method abducts the noise $U'_Y = U_Y \\mid s, y$, predicts a counterfactual abstraction distribution in the new state $s'$, and maps back to actions with the Bayes reweighting $P_{A'}(a' \\mid y', s') = P_Y(y' \\mid s', a')\\, P_A(a' \\mid s') / P_Y(y' \\mid s')$. The observed token $a$ is deliberately omitted from the conditioning, so the procedure cannot merely inflate the probability of repeating the same token; it must instead favour actions that express the same meaning in the new context. This is what the paper means by preserving high-level action meaning across counterfactual contexts.","pith_inferences":["An implication the paper leaves implicit: anchoring counterfactuals on $Y$ changes the unit of explanation from words to properties, so questions about blame or responsibility become about intents and features rather than about particular strings; downstream causal-explanations discussions would inherit that shift.","The paper never verifies that the external classifier's $P_Y$ matches the agent's own abstraction mechanism $f_Y$; a natural next experiment is to vary the abstraction scheme while fixing everything else and check whether ACF's advantage tracks the abstraction's fidelity to the agent's semantics.","Because abduction conditions on the abstraction value only, ACF predicts the same counterfactual distribution for any two observed actions that map to the same $y$; this is a testable signature—two actions with identical labels but different token-level identities should nonetheless yield identical ACF counterfactuals.","If $Y$ is degenerate (one class), ACF collapses to the interventional distribution, and if $Y$ becomes as detailed as the action itself it approaches token-level behaviour; charting this spectrum would map precisely where token-level methods start to fail."],"forward_implications":["In choice-based games, ACF remains well defined when the factual and counterfactual action menus differ in size or share no token labels, a case where token-level Gumbel counterfactuals require padding or truncating noise.","In open-text generation, ACF can change the surface text while holding the high-level content fixed; the reported Abstraction Change Rate drops from 0.38–0.67 with token-level methods to 0.02–0.41 with ACF across the two text tasks.","The method needs only black-box samples from the language model; the token-level mechanism $f_A$ is never inverted, so no internal logits or weights are required.","Counterfactual outputs stay coherent with each other: ACF wins the semantic-tightness comparison against token-level counterfactuals in all six reported configurations.","Abstractions can be supplied in three ways—expert annotations, supervised classifiers, or unsupervised LLM-discovered topic clusters—so the framework is usable in settings without a predefined action space."],"supporting_citations":[{"why":"Defines the token-level Gumbel counterfactual approach that ACF is compared against and whose failure modes ACF is designed to overcome.","marker":"[26]"},{"why":"Introduces the SCM-based token-level counterfactual method for LLMs that motivates the need for abstraction-level inference.","marker":"[7]"},{"why":"Provides the Gumbel-Max SCM used to formalize the token sampling mechanism the paper deliberately bypasses.","marker":"[22]"},{"why":"Supplies the text-based game benchmark with morality annotations used as degenerate abstractions in the first evaluation.","marker":"[23]"},{"why":"Supplies the biography dataset and profession labels used for the supervised abstraction and gender-steering intervention.","marker":"[8]"},{"why":"Supplies the Reddit emotion dataset and labels used for the emotion abstraction in token-replacement counterfactual generation.","marker":"[9]"},{"why":"Supplies the MiMiC affine steering transformation used as the latent-space gender intervention.","marker":"[30]"},{"why":"Supplies the DistilBERT architecture fine-tuned to implement the supervised abstraction classifiers.","marker":"[29]"},{"why":"Motivates the unsupervised abstraction through semantic equivalence classes of generated text.","marker":"[10]"}],"fun_headline_variants":["What-if reasoning for AI agents, anchored in meaning","Semantic counterfactuals beat token-level for language agents","New counterfactual method focuses on action meaning, not tokens","Language agent what-ifs: abstract actions, better results","Meaning-based counterfactuals for language agents"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"ACF assumes that the abstraction variable $Y$ faithfully and stably captures the high-level meaning of the agent's action across contexts, and that the distribution $P_Y$ used in the procedure (certified by annotations, a classifier, or an LLM) is a valid stand-in for the agent's own abstraction mechanism $f_Y$.","fun_headline_variants_meta":{"raw":{"variants":["What-if reasoning for AI agents, anchored in meaning","Semantic counterfactuals beat token-level for language agents","New counterfactual method focuses on action meaning, not tokens","Language agent what-ifs: abstract actions, better results","Meaning-based counterfactuals for language agents"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000161,"raw_usage":{"total_tokens":1231,"prompt_tokens":934,"completion_tokens":297,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":550,"completion_tokens_details":{"reasoning_tokens":219}},"tokens_in":550,"tokens_out":297,"duration_ms":3326,"temperature":1.0,"reasoning_tokens":219,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:16:50.559956+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the abstraction on labels that are known not to reflect the agent's action semantics, or on random labels, and rerun the same comparisons: if the consistency advantage over token-level methods survives, the claimed mechanism—semantic fidelity—is not what drives it.","supporting_citations":[{"cited_title":"Gumbel Counter- factual Generation From Language Models","cited_arxiv_id":null,"evidence_quote":"Defines the token-level Gumbel counterfactual approach that ACF is compared against and whose failure modes ACF is designed to overcome."},{"cited_title":"Counterfactual Off-Policy Evaluation with Gumbel-Max Structural Causal Models","cited_arxiv_id":null,"evidence_quote":"Provides the Gumbel-Max SCM used to formalize the token sampling mechanism the paper deliberately bypasses."},{"cited_title":"Do the Rewards Justify the Means? Measuring Trade-Offs Between Rewards and Ethical Behavior in the Machiavelli Benchmark","cited_arxiv_id":null,"evidence_quote":"Supplies the text-based game benchmark with morality annotations used as degenerate abstractions in the first evaluation."},{"cited_title":"Representation Surgery: Theory and Practice of Affine Steering, July","cited_arxiv_id":null,"evidence_quote":"Supplies the MiMiC affine steering transformation used as the latent-space gender intervention."}],"review_version":1}