A new arXiv preprint proposes a weaker assumption called 'intervention-immediacy faithfulness' that allows causal discovery to proceed correctly even when buffering or stabilizing pathways cancel out and violate the standard faithfulness assumption. The authors show that hard interventions alone can nonparametrically identify causal structures, repositioning interventions as the primary information source over conditional independence testing on observational data. The work also characterizes equivalence classes when interventions are insufficient to fully identify structure.
CausaLab is a new evaluation environment that tests LLM agents on interactive causal discovery tasks, requiring them to recover both causal graphs and structural equations from synthetic laboratory episodes governed by randomly sampled structural causal models (SCMs). The benchmark separates predictive accuracy from genuine causal understanding, revealing a persistent gap: GPT-5.2-high achieves 92% task accuracy in a 6-node observational setting but only 0.471 all-edge F1 for mechanism recovery. Mixed observation-intervention strategies improve structural fidelity, while pure intervention strategies underperform on both metrics. Premature stopping is identified as a key agent weakness, partially mitigated by prompting models to verify hypothesis-data consistency.
A new arXiv preprint proposes an extension of PCMCI+, a leading causal discovery method for multivariate time series, to handle irregularly sampled data by aggregating causal influence over temporal windows rather than fixed lags. The method is evaluated on synthetic irregular event streams with known causal structures across varying signal-to-noise ratios, consistently recovering the underlying causal graph and outperforming standard PCMCI+. This addresses a practical gap for domains like healthcare, sensor networks, and financial transactions where regular sampling is not guaranteed.
A new arXiv preprint introduces a causal auditing framework for neuron-row attribution methods in LLMs, using one-shot zeroing interventions to test whether high-scoring neurons are genuinely causally important. The authors find that attribution methods outperform activation/magnitude baselines for identifying dispensable rows, and that refusal behavior (for hate/crime prompts) can be installed or removed via attributed rows while preserving fluency. A key finding is that refusal occupies a redundant subspace — different attribution methods identify largely disjoint sufficient row sets — meaning no single method recovers a unique causal mechanism. Rank-stability, a common proxy for selector quality, is shown to be a poor indicator of causal validity.
A new arXiv paper investigates whether active exploration reduces the 'conjunctive handicap' in causal learning, using a blicket detector task with adult participants who could freely intervene to identify causal objects. Results show active exploration substantially improves human conjunctive causal reasoning, though conjunctive rules still require more tests than disjunctive ones. State-of-the-art LLMs approach human-level hypothesis inference accuracy but show less efficient exploration strategies and similar conjunctive-disjunctive performance gaps, raising questions about the nature of LLM causal reasoning.
A new arXiv paper demonstrates empirically that static retrieval relevance scores are nearly statistically independent from causal utility in multi-step agentic search (Spearman rho = -0.026 across 23,322 document observations). Using a ReAct-style agent over HotpotQA with counterfactual trajectory replays, the authors show that roughly a third of documents are causally load-bearing ('bridge documents') while appearing useless to a static reader. The mechanism is identified: bridge documents supply discriminative entities that redirect subsequent queries, meaning optimizing static RAG utility does not deliver agentic utility.
A new arXiv preprint proposes that LLMs learn causal structure through 'variational induction' — a difference-making logic — rather than through the dominant formalisms of Judea Pearl's interventionist approach or the Neyman-Rubin potential outcomes framework. The author analyzes how this logic is realized during training and maps specific architectural features (token embeddings, self-attention) to their roles in this inductive process. The argument draws a parallel between LLM causal learning and the experimental method of systematically varying circumstances. This is a theoretical contribution to understanding how LLMs represent causal and world-model structure.
Researchers introduce DKCD, a framework that augments LLM-based causal discovery with domain-specific knowledge retrieval and guided reasoning to address two failure modes: insufficient identification of latent causal factors and unreliable factor annotation. The system chains knowledge mining, knowledge-guided causal reasoning, and causal structure discovery into a pipeline that produces causal graphs from unstructured text in high-expertise domains like healthcare and finance. Experiments on two domain-specific datasets show improvements over baselines on both factor identification and graph construction quality.
A new arXiv paper studies language models trained to explain their own predictions using counterfactual supervision derived from earlier checkpoints or behaviorally similar models. Surprisingly, models frequently produce explanations more faithful to their current behavior than to the training targets — a phenomenon the authors call 'introspective coupling.' The effect persists across tasks including sycophancy and refusal, tracks concurrent behavioral shifts from other post-training objectives, and is robust to label noise, suggesting fixed counterfactual explanation datasets can serve as scalable post-training signal for introspection.