Reasoning in large language models can be disrupted by simple input perturbations (e.g. adding irrelevant distractors), suggesting reliance on spurious correlations. Identifying and suppressing the spurious features involved in long chain-of-thought reasoning remains challenging. We introduce Spurious-feature WEight EdiTing (SWEET), a training-free method that combines counterfactual interventions with mechanistic interpretability to identify and suppress spurious-features. SWEET extracts token- and sequence-level behavioral directions from activation contrasts between original problems and their counterfactual variants - capturing local and global patterns of spurious reasoning. It then localizes the responsible MLP neurons and attention heads and suppresses their contributions along these directions through weight edits. Using only 20 original-counterfactual pairs for direction extraction, SWEET improves accuracy by an average of 11.8 percentage points on perturbed inputs and 2 points on original inputs across five open-weight reasoning models and five reasoning benchmarks. It also reduces the relative gap between original and perturbed accuracies from 17.6% to 2.5%. Analyses identify two recurring forms of spurious reasoning, i.e. over-integrating irrelevant conditions and overthinking, and provide mechanistic evidence that SWEET improves reasoning robustness by suppressing both behavioral directions.
Off-policy evaluation (OPE) estimates the value a candidate decision policy would achieve, using only data logged by the policy already in production. In debt collections, unrecorded factors can influence both actions and outcomes. This paper presents JUBILEE, a benchmark environment built around a debt-collections simulator that generates million-loan panels with competing risks, rare recovery events, and hidden factors driving both outcomes and assignment. The evaluation includes an 80-run comparison of tabular and gradient-boosted estimators under two synthetic outcome mechanisms and a 190-run parameter grid. Independent policy simulations provide reference values with estimated Monte Carlo uncertainty for a synthetic reward score and for simulated cash. With hidden factors influencing action assignment, several tabular methods agree closely while underestimating the field-visit policy's score, an error that also appears with boosted models. Intervals from tabular doubly robust estimation usually contain the reference without hidden assignment dependence but mostly miss it when that dependence is present. Estimates from randomized decisions do not show the same large error, although their limited precision leaves policy choice uncertain. The decision consequences depend on how policy value is defined. The field-visit policy ranks last under the score and first under simulated cash, which excludes channel costs. Under cash, log-based estimators miss this best candidate in nearly every run with hidden assignment dependence in the paired study. JUBILEE supports reproducible study of these failure modes under specified collections mechanisms.
Safety-aligned large language models often lose their safety behaviors after fine-tuning, even when safety data are included. While this fragility is well documented, its underlying cause remains unclear. In this paper, we propose that formulaic refusal strings can function as outlier modes: narrow, weakly grounded response patterns that are easy to fit but vulnerable to subsequent updates. Controlled experiments support this account: fixed refusals are acquired much like arbitrary constant strings, in contrast to prompt-grounded natural responses. Motivated by these findings, we introduce Naturalness-Enhanced Safety Targets (NEST), a safety alignment framework that improves safety retention by constructing natural, context-aware safety targets that keep safety responses within the model's existing output distribution and grounded in semantics. We instantiate NEST in both supervised and preference-based alignment settings, using perplexity under the original model as a proxy for distributional support. Across multiple models and alignment regimes, natural safety targets improve both immediate safety and its retention under continued fine-tuning without sacrificing helpfulness. These findings suggest a practical principle for more durable safety alignment: construct safety targets that draw on the model's existing language capabilities and remain grounded in the meaning of each request.
What structure on the rule space of a cellular-automaton family exposes which rules can support which behaviours? For the 262,144 life-like rules -- the two-dimensional outer-totalistic family containing Conway's Game of Life -- the standard coordinate, Langton's activity parameter, is long known to be too coarse. I organise the family by three symmetry-derived rule-table coordinates: an unweighted activity, a birth-survival asymmetry (odd under the swap of birth and survival), and the Vichniac-Binder Boolean-derivative sensitivity in closed form, whose centre-bit component vanishes exactly on the centre-blind (B=S) submanifold. The cardinality coordinates lambda, delta are necessary but never sufficient to certify behaviour; the centre-bit sensitivity labels the B=S submanifold on which the paper's one exact exclusion holds. Combined with an exhaustive enumeration of every rule's compact gliders, still lifes and oscillators (three-by-three seed, period at most sixteen), they expose a hierarchical factorisation: the birth condition is a narrow gate deciding whether anything can propagate, the survival condition tunes storage, and every compact-glider-supporting rule is birth-light and survival-rich. One exclusion is exact: no B=S rule supports a compact glider at that footprint, proven by exhaustive enumeration and closing only for compact-footprint patterns -- 36 B=S rules ship at four-by-four seed and 40 at five-by-five. Two dynamical results frame the whole: glider-capable rules that random sampling never expresses show that capability is certifiable from the table while its expression is only estimable from dynamics; and every uncatalogued rule producing a richly Life-like ecology turns out to be a Conway or HighLife variant on the neighbour counts its trajectories actually visit.
Offline reinforcement learning can benefit from data collected in related tasks, yet indiscriminate knowledge transfer may lead to negative transfer under task mismatch and limited target-data coverage. We propose a target-aware Bayesian transfer framework for offline Q-learning based on multi-task Gaussian processes that jointly model task-specific Q-functions. Source tasks are assessed using target predictive evidence and maximum mean discrepancy (MMD), while posterior uncertainty and behavior-supported action sets guide conservative Bellman updates and policy extraction. To account for the dependence induced by Bellman bootstrapping, we distinguish Bellman innovations from bootstrap errors and establish stability and predictive source-selection results under explicit regularity conditions. We further derive selection-adaptive posterior contraction and offline policy-performance bounds, identifying the structural conditions under which source information can be pooled to achieve minimax-optimal rates in a shared-function Gaussian sequence experiment. Simulation studies with heterogeneous source tasks demonstrate improved target Q-function estimation and reduced material negative transfer compared with indiscriminate source pooling. An application to human decision-making data further shows higher average expected immediate payoff than both target-only learning and matched-size random source selection. These results highlight the benefits of combining target-aware source selection with uncertainty-aware conservative learning for reliable and sample-efficient offline reinforcement learning.
AI agents are increasingly capable of discovering new mathematical results, yet their reasoning and generated proofs can be difficult to understand and verify. We introduce an interpretable framework for AI-guided mathematical discovery, focused on improving quantitative bounds on analytic constants. Our key insight is that existing proofs often contain slack in intermediate lemmas, which can be tightened without reconstructing the entire argument. To exploit this structure, we represent proofs as partially-formalized Lean blueprints: human-interpretable dependency graphs of definitions, lemmas, and theorems. An AI agent then iteratively refines the Lean code while preserving the blueprint's dependency structure, allowing local improvements to automatically propagate to a verified final bound. This interpretable structure allows humans to inspect the agent's refinements and understand how local improvements propagate to the final theorem. For Grothendieck's constant ($\mathcal{K}_G$), we obtain $$ \mathcal{K}_G \leq \frac{\pi}{2\log(1+\sqrt{2})} - 10^{-17}, $$ an explicit numerical improvement over Krivine's 1977 bound. For Korenblum's constant ($\mathcal{K}_K$), we obtain $$ \mathcal{K}_K \geq 0.356565, $$ improving Wang's 2025 bound of $0.3554$. Our experiments demonstrate the potential of both partial formalization and iterative agentic refinement for interpretable mathematical discovery.
Medical Vision-Language Pre-training (VLP) leverages the semantic richness of radiology reports to learn 3D representations, yet current architectures remain fundamentally anatomy-agnostic. While clinical diagnostics rely on organ-specific context, standard encoders apply a uniform, dense parameterization to all volumetric patches, and existing Mixture-of-Experts (MoE) routers gate on isolated local features without explicit access to the organ composition of the scan or crop. We propose Anatomy-Activated Mixture-of-Experts (A$^{2}$MoE), a framework that incorporates anatomical inductive biases within the model's computational path. A$^{2}$MoE introduces two key components: (1) a Histogram Representation Router (HRR) that conditions expert selection on global visual context and a predicted anatomical composition histogram, encouraging anatomy-dependent expert weighting; and (2) an Organ-Query Multi-Scale Projector (OQMSP) that utilizes learnable queries to extract organ-level embeddings for fine-grained alignment without requiring manual annotations at inference. By routing at the patch level rather than the token level, A$^{2}$MoE reduces routing complexity by orders of magnitude while satisfying load balancing by construction. Pre-trained on over 60,000 multi-phase CT scans, A$^{2}$MoE achieves state-of-the-art performance across segmentation, classification, and detection benchmarks. Furthermore, we demonstrate that our anatomically grounded inductive biases support robust generalization to MRI and ultrasound, establishing a scalable foundation for medical representation learning. The code is available at https://github.com/gmum/A2MoE.
Change impact analysis (CIA) helps developers identify code entities impacted by their initial changes. While several approaches for CIA were proposed, they still struggle with either low precision or low recall, thereby increasing manual effort for developers to verify the results. Furthermore, limitations in the used datasets and experimental setups can bias the results reported by many CIA approaches. In this paper, we first address the dataset limitations by manually labeling 1,931 changed code entities stemming from 263 commits of 22 open-source Java projects. Next, we introduce two variants of katIA, our code change impact analysis approach that combines static impact analysis with the reasoning capabilities of large language models (LLMs). The first variant uses a single prompt (katIA-SP) for refining an initial static impact set while the second variant uses three specialized agents (katIA-MA). The evaluation with our curated dataset shows that katIA-SP with Gemini 2.5 Flash achieved an F1-score of 22.4% when matching impacted entities by their fully qualified names (FQN) and 27.0% when matching them by their class and entity names (CE). The latter shows an improvement of 2.0 percentage points over the state-of-the-art. This score is further improved by katIA-MA which, using Gemini 2.5 Flash, yields an FQN-matching F1-score of 29.1% and a CE-matching F1-score of 33.3%. While these improvements come at the cost of increased median generation time from 0.570 s to 11.214 s and median token usage from 1,235 to 6,462 tokens, the improved performance justifies the additional cost. The comparison of three LLMs using katIA-SP finds that Gemini 2.5 Flash is the most consistent model. These results indicate that combining static impact analysis with LLM-based refinement, particularly with the multi-agent variant, can improve code change impact analysis.
Options market makers absorb client order flow and must hedge the resulting inventory risk under trading frictions. We study whether an option hedging reinforcement learning (RL) agent trained on historical market data outperforms three alternatives in tail-risk protection: (i) traditional delta hedging, (ii) simple hedge-ratio rules, and (iii) the same agent trained in simulated markets. Most RL hedging agents are trained in model-generated markets that do not capture the crashes and volatility regimes present in real data. We introduce an empirical replay environment using S&P 500 option data from 1996 to 2017 and compare against a matched geometric Brownian motion (GBM) environment, a stochastic-volatility (SABR) simulator, and a hybrid that values historical paths with Black-Scholes, all sharing the same portfolio mechanics, stochastic arrivals, costs, and test period. We train a distributional Twin Delayed Deep Deterministic Policy Gradient (TD3) agent and evaluate it out of sample over 2018-2023. The agent trained on empirical data reduces Conditional Value-at-Risk (CVaR) tail loss by 3.1% compared to delta hedging and outperforms every hedge-ratio rule chosen during the training period and every alternative training environment. The reduction reaches 8.9% in volatile and falling markets, but reverses to a cost of 7-10% in rising markets. Empirical training induces a state-dependent short net delta: the policy under-hedges negative-delta books, most strongly after prices have fallen. This tilt disappears with GBM training and is weaker with SABR or hybrid training. These findings suggest that exposure to historical market dynamics can materially change a tail-risk hedging policy, although the evidence rests on a single test path and the edge is regime-conditional.
Modern observational biomedical datasets contain increasingly extensive baseline and outcome measurements, yet detecting a clinically meaningful treatment effect remains difficult because investigators must identify an outcome that captures treatment response while adjusting for confounding distributed across numerous noisy baseline variables. We introduce SCOUT (Score-based Causal Outcome Learning Using Treatment signals), a framework that learns a clinically interpretable, treatment-responsive outcome while adjusting for confounding throughout the learning process. SCOUT constructs low-dimensional adjustment representations of high-dimensional baseline measurements to estimate the average treatment effect (ATE) and conditional average treatment effect (CATE) for the learned outcome. Importantly, the final adjustment space is at most two-dimensional, allowing treatment-effect heterogeneity to be estimated in this space rather than in the full high-dimensional covariate space. SCOUT also introduces a fourth-order statistic for testing treatment-effect heterogeneity that emphasizes unusually large deviations from the ATE. In synthetic and semi-synthetic experiments, the algorithm achieved competitive ATE estimation accuracy and substantially lower CATE estimation error. In an observational study of patients with bipolar disorder, SCOUT identified a clinically recognizable lithium-associated adverse-effect profile, detecting both a significant ATE (Holm-adjusted $p$ = 0.0026) and significant treatment-effect heterogeneity (Holm-adjusted $p$ = 0.033).These results demonstrate how jointly learning outcomes and controlling for confounding can facilitate the discovery of treatment effects in high-dimensional observational data.
Cyber-physical smart-grid detectors are typically treated as black boxes and validated exclusively on clean, temporally shuffled datasets, which risk overly optimistic performance claims. This work presents a multi-testbed evaluation study built around an interpretable, time-windowed fusion Random Forest, adopting the pessimistic result whenever evaluation protocols disagree. (i) On a public paired cyber-physical dataset, shuffled cross-validation is optimistic: macro-F1 falls from 0.96 to 0.88 +/- 0.06 under rolling-origin temporal evaluation. A topology-free graph convolutional network (GCN) serving as a control confirms that this lightweight model is competitive on these splits. (ii) The adaptive and adaptive-joint feature-space attacks, which are infeasible as realistic grid manipulations, degrade macro-F1 to 0.67 and 0.34, respectively. Adversarial training and randomized smoothing partially restore performance, and a physics-constrained evolution-strategy evasion on the IEEE 14-Bus testbed shows the same vulnerability pattern. (iii) On real IEC-104 traffic, the cyber view reliably catches structural attacks (recall of 0.87 to 0.98) but largely fails on semantic value manipulation (recall of 0.11 to 0.35). An observability taxonomy on the IEEE 14-Bus system reveals that stealthy false data injection defeats all three observation views. (iv) On a full AC weighted least-squares state-estimation testbed, a forecasting-aided temporal test is evaded by slow ramps precisely where a rate-invariant physics-manifold test sustains detection at a common 5% false-alarm target. A full end-to-end evaluation on IEEE 118-Bus confirms the ordering but bounds its strength: the manifold test detects every trial down to 0.05 degree/step and 0.48 at 0.02 degree/step, against 0.14 for the temporal test. All code and benchmarks are released.
This paper introduces COUPLED FLOW, a new neural operator learning framework for multi-physics PDE systems that captures the co-evolution of multiple physical processes by modeling their cross-process interactions. Unlike existing neural operator methods that primarily target single-physics settings, COUPLED FLOW employs a novel structure-constrained flow matching algorithm to learn process- specific mean velocity fields conditioned on the full multi-physics state. This enables fast and accurate one-step prediction of coupled PDE systems that would otherwise require expensive numerical simulations. A key challenge in training COUPLED FLOW however lies in sampling structure-constrained conditional flow interpolants that capture cross-process correlations and serve as training examples for flow matching. To address this challenge, we derive a multi-physics mean-flow consistency identity that characterizes cross-process interactions and defines the structure-constrained mean-flow matching loss. To provide informative training samples for flow matching with this loss, we develop a principled method for sampling correlated conditional interpolants based on posterior multi-output Gaus- sian processes, which capture richer cross-process correlations than the commonly used endpoint-conditioned linear interpolants. We also establish a theoretical re- sult showing that the Wasserstein-2 distance between the predicted and oracle PDE solution distributions is upper-bounded in terms of the structure-constrained flow matching loss and the smoothness of the learned flow. Experiments across diverse coupled PDE benchmarks demonstrate that COUPLED FLOW consistently outperforms existing state-of-the-art neural operators and generative PDE solvers.
Data transformation pipelines shape the training distribution in computer vision, but designing them is a computationally expensive problem. We investigate whether large language models can enable efficient discovery by synthesizing executable transformation programs outside predefined augmentation spaces. We compare nine strategies under a unified training protocol, including standard augmentation policies, brute-force search, direct code generation with two LLMs, and performance-guided generation using LLMs fine-tuned on accuracy-ordered transformation pairs. When the best candidates are retrained for 20 epochs, LLM-generated transformation programs lead on every dataset, reaching 0.8877 accuracy on CIFAR-10, 0.5825 on CIFAR-100, and 0.9588 on SVHN, surpassing brute-force search over 6,000 candidates. The best transform on CIFAR-10 and SVHN comes from a fine-tuned generator using 270 candidates, while the best on CIFAR-100 comes from an off-the-shelf generator using only 10. This needs 22x fewer candidate evaluations in the generation step and 600x fewer for off-the-shelf generations, relative to brute-force search. Beyond transformation discovery, we analyze the reliability of the reduced-budget evaluation commonly used for augmentation search. Correlating one-epoch selection scores with 20-epoch outcomes reveals strong, weak, and absent predictive relationships depending on the dataset and architecture, with correlations ranging from 0.756 to near zero. A positive control indicates that range restriction alone is unlikely to explain the low correlations. Our results indicate that semantic synthesis with code LLMs provides an efficient alternative to search-based transformation discovery and highlight the importance of validating evaluation fidelity when optimizing data transformations.
The field of mechanistic interpretability is well-suited to collaborative projects between mathematicians and AI researchers. This work describes one such project. We trained a small transformer to learn the language of the Coxeter group affine A2, best known as the symmetries of the tiling of the Euclidean plane by equilateral triangles. We then formulated and proved a word-theoretic conjecture about affine A2 after observing the attention patterns of the trained model.
Researchers increasingly estimate political behavior based on race and gender from voter files, where race is imputed from surnames and geography. This only works where surnames suggest group membership, an assumption violated where surnames follow norms around marriage. Imputation is therefore biased against women who adopt their husband's surname, especially those who racially inter-marry, an outcome itself endogenous to partisanship. Using six US states recording self-reported race, I show that imputation is less accurate for Asian and Hispanic women, particularly those sharing the surname of a partner of another race, who differ politically from other women of their race. This violates assumptions behind downstream inference approaches, which nearly double the gender gap in partisanship among Asian voters. I derive an expression showing the bias is largest where a group is small. Using maiden rather than marriage-adopted surnames in North Carolina raises accuracy up to fourfold for women married across racial lines. Recovering maiden surnames, or omitting the married surname, removes much of the bias. In a national replication of a study of gender gaps in turnout, the correction more than triples the gap among Asian voters. In Lebanon, where wives register in their husband's locality, the error enters through geography.
Existing evaluations of AI models on high-finance work focus on retrieving answers from financial documents and producing numerically correct outputs. Real private equity (PE) and investment banking (IB) work also depends on choosing between conflicting sources, making sound assumptions, and keeping every step auditable. We introduce DeepFinance, an ultra long-horizon benchmark of 21 end-to-end PE and IB tasks that take a human professional an average of 27.1 hours. Each task pairs a request with a data room averaging 19 files and is graded against an average of 84 expert-written rubric criteria covering output, workflow, and judgment. We evaluate 22 model configurations, each in its native agent harness. The best configuration, GPT-6 Astra with max reasoning, satisfies 57.6% of rubric criteria, and the top four configurations are not statistically distinguishable. Frontier models often calculate correctly from the wrong inputs and find decision-relevant evidence without acting on it.
Binary decision-making under uncertainty arises across a wide range of applications, including energy systems scheduling, transportation, and workforce planning. While machine learning has been increasingly applied to accelerate or replace mixed-integer solvers in such settings, existing approaches rely on predictive learning, training models to reproduce optimal decisions from labeled data, which lacks knowledge of counterfactual outcomes and provides no guarantees that the resulting decisions will be feasible. This limitation is further compounded under distributional shift, where policies trained on historical data may produce systematically infeasible decisions. We propose a prescriptive learning framework for binary decision-making that addresses both limitations. Rather than predicting optimal labels, the framework jointly optimizes over linear decision policies and feasible assignments across all training scenarios, allowing feasible alternatives and their operational performance to be considered directly during training. We further extend the framework to a distributionally robust chance-constrained formulation using a Wasserstein ambiguity set, which enforces feasibility with a prescribed probability under the worst-case distribution in the ambiguity set and enables a tunable trade-off between robustness and solution quality. We demonstrate the framework on two case studies, namely a unit commitment problem and a staff allocation problem. Out-of-sample experiments under test distributions that either match or differ from the training distribution show that the robust model is more conservative when the two distributions coincide, while under distributional shift it substantially reduces unmet demand and lowers total cost by up to 76.5% relative to classical predictive learning.
Surgical video understanding is an important foundation for intelligent assistance in robot-assisted surgery. Similar surgical views can reflect different actions or procedural stages. Interpreting these views requires access to earlier observations, yet frame-centric prompting leaves this history unavailable. We introduce MemSurg, a framework that remembers surgical observations before reasoning over them. Its structured memory graph stores timestamped entity categories, locations, motion cues, and interaction summaries. Chronological links preserve recent context, while entity-associated links connect recurring observations across the procedure. Task-adaptive retrieval then supplies evidence-grounded prompts for scene and workflow reasoning, supporting both frozen inference and memory-conditioned supervised fine-tuning. Evaluation on robot-assisted prostatectomy and two complementary surgical benchmarks shows a 10.43-percentage-point gain over SurgRAW with GPT-4o. Memory conditioning adds 6.52 points under matched Qwen3-VL-8B fine-tuning, and an inference-time plug-in adds 7.04 points to EndoChat. These results support structured memory as a reusable source of procedural evidence for surgical video understanding. Our code will be publicly released.
In Bangla, a main verb is often followed by one of a small set of vector verbs. These verbs lose their own meaning and add a shade of meaning instead, such as completion. For example, likhe phello means that someone finished writing, not that they wrote and threw something. The same two verbs can also describe two separate events, so only the context shows which meaning is meant. This makes compound verbs a hard test of how language models handle multiword expressions. We present BanglaCV-Probe, a benchmark of 17 vector verbs with 1,044 minimal pairs, 204 entailment items and 197 pairs of correct and incorrect combinations, all checked by native speakers. We test 26 models (encoders, LLMs and translation systems) on four skills: telling the two meanings apart internally, understanding what a vector adds, preferring correct combinations and keeping the vector's meaning in translation. Probes on five encoders tell the two meanings apart and still reach up to 93% accuracy when trained on pairs without surface cues. Prompted models rarely use this knowledge. Entailment accuracy stays between 59% and 81%. Several LLMs answer Yes or No to almost every item, and nine of ten models miss that someone who ate up the rice left none. Most instruction-tuned LLMs translate vectors literally in 15-20% of vector sentences and in up to 55% for mora (die). On correct combinations, a small Bangla-only model (88%) matches BLOOM-7B1 (87%) and both beat Qwen2.5-7B (74%). Data and code: https://github.com/ashuvo25/Pole_OR_Vector
Forensic verification has converged on a binary "real vs. fake" label that confates fully synthetic, tampered, and AI-retouched images despite their very different consequences. We argue that GenAI manipulations decompose along two complementary channels: a camera channel of sensor fingerprints (destroyed by neural processing), and a semantic channel of scene-level coherence (disrupted by content edits). Each manipulation type leaves a distinctive two-channel signature. We instantiate this taxonomy in 2CAP (2-Channel Authenticity Protocol), pairing a contrastively-trained camera encoder with a frozen semantic encoder to serve two applications on a shared backbone: (i) four-class authenticity classification via cross-attention fusion with learnable reliability weights; and (ii) evidence generation, where an authentic image serves as a reference for a query and the two encoders produce per-channel alignment scores together with a patch-level saliency map. These signals are supplied as privileged context to a frozen Vision-Language Model (VLM) through an Observe-Generate-Refine loop, yielding explanations with localization without forensic fine-tuning. On a multi-source benchmark, 2CAP attains the best classification performance, including strong retouching detection where existing methods fail, and improves explanation quality.