Papers

Many applications of physical AI run within finite or closed physical worlds with a limited set of physical laws governing object behavior. Examples include robots working in a warehouse and agents moving around in a video game. In order to better organize, filter, and curate data for physical AI applications, we propose a new approach centered on the unique settings and physical laws of individual datasets. We train LeCuration, a small world model intended to serve as a data curation tool for a separate, larger downstream model. To build this model, we choose LeWorldModel (LeWM)as our latent encoder and predictor, adding a diffusion transformer (DiT) decoder to add visuals to autoregressive gameplay rollout. We find that the embeddings of this model can be used as an anomaly detection signal and as a content-based clustering heuristic, and that auto-regressively predicting the game state with this model allows us to qualitatively check for action-state consistency. This paper presents a qualitative, proof-of-concept case study on CS:GO gameplay data; we do not yet report quantitative curation metrics or downstream training results, which we identify as the key next step.

Attention routing requires a signal that predicts the value of attention on the current prefix. We evaluate whether hidden-state extrapolation error, curvature and error change improve this prediction beyond uncertainty, one-step displacement, position and state projections. Paired executions of the final attention layer supply signed next-token loss differences in frozen SmolLM3-3B-Base and Qwen3.5-4B-Base checkpoints. Utility-supervised routers are tested on 100 held-out PG-19 books at an identical causal 20 percent invocation quota. None of six prespecified comparisons shows a positive gain after familywise correction. In Qwen3.5, a parameter-matched fixed-projection control lowers NLL by 0.00356 nats/token relative to the trajectory router (95 percent interval 0.00218 to 0.00487). Secondary results depend on the operation removed, feature location and scoring horizon; frozen thresholds also drift substantially at longer horizons. Actual selected-query execution yields small long-sequence latency reductions with increased NLL, while learned routers remain slower during cached continuation. The study identifies limits on the incremental value of these trajectory summaries and separates allocation quality from measured inference benefit.

Modern agentic coding frameworks increasingly rely on community-shared rule files (e.g., AGENTS.md or .cursorrules) to guide autonomous code generation, yet the security risks of this pipeline remain underexplored. To bridge this gap, we introduce the package hallucination attack, where an attacker injects malicious prompts into benign rule files to induce coding agents to replace legitimate dependencies with attacker-controlled packages. To obtain effective malicious prompts injected into rule files, we propose PackHallu, an evolutionary optimization framework that iteratively rewrites these injected prompts using trajectory-level feedback and LLM-guided mutations. Evaluations across multiple benchmarks, LLMs, and agent frameworks show that PackHallu achieves high attack success rates and strong transferability across diverse models and agent combinations. Our findings demonstrate that coding agents are vulnerable to package hallucination attacks, highlighting the urgent need for stronger security safeguards in autonomous coding systems.

Best-of-N (BoN) is a common inference-time alignment method that selects the highest-scoring response among N samples from a reference model. Evaluating BoN policies from logged data is challenging under sample-only access because standard off-policy estimators require density ratios that depend on unavailable response likelihoods. In this paper, we propose a sample-only framework for evaluating and selecting BoN policies without access to these likelihoods. We show that the order-statistic structure of BoN allows the required density ratios to be expressed through score-rank probabilities that are estimable from samples alone. We then develop a doubly robust estimator of the BoN policy value (BoN-DR) that efficiently reuses a shared auxiliary sample pool across candidate budgets. We establish valid asymptotic inference even under reward estimator misspecification and prove the efficiency of our BoN-DR estimator. Since larger budgets can amplify errors in the score function and lead to reward overoptimization, we derive two selection rules: (i) maximizing the estimated policy value and (ii) maximizing a lower confidence bound on the improvement over the reference policy, which accounts for estimation uncertainty and provides a no-harm guarantee. Across synthetic experiments and GSM8K with multiple reference and reward models, our framework accurately estimates BoN policy values and selects effective sampling budgets.

The difficulty of learning goal-reaching policies is often attributed to a "curse of horizon" that manifests as bias accumulation in temporal-difference backups and noisy advantage estimates. In this work, we identify an additional informational curse of horizon in goal-conditioned policy learning, where increasing the goal relabeling horizon can significantly reduce policy generalization and performance. Through a series of controlled experiments with oracle planners, we decouple the goal horizons sampled during training from those that the policy is asked to reach at test time. Even when evaluated only on a sequence of nearby subgoals, goal-conditioned behavioral cloning (BC) policies suffer from severe, training horizon-dependent performance degradation that is mitigated by reinforcement learning (RL) objectives. We explain this phenomenon as a horizon-dependent decrease in the conditional mutual information between actions and hindsight-relabeled goals, and find empirically that both BC and RL policies trained on longer-horizon goals exhibit a shift in sensitivity from goal to state information, as measured by the policy's input Jacobians. Motivated by this observation, we find that distilling the input Jacobians of short-horizon policies into long-horizon policies yields significant performance gains, especially in combinatorial manipulation tasks. Taken together, our results highlight goal relabeling horizon as an important consideration when learning generalist policies from offline data.

Agentic systems use large language models (LLMs) to carry out concrete tasks. Prior work often borrows abstractions such as scheduling, caching or isolation piecemeal from operating systems, so the mechanisms it builds share little common ground, and the shared view of the two forms of agentic system, Workflows and Agents, is limited. We construct an abstract machine that provides both. We treat the LLM as an Oracle and extend a two-stack pushdown automaton with one instruction, which hands the Oracle a whole stack as its query and appends the answer to that same stack. The machine thus performs two computations, the Oracle's and a Turing-complete one that we call the Priestess. A stack that the program only appends to grows autoregressively, as an agent's context does. Two symmetry breakings, S in storage and T in transitions, make a Priestess program the operating system of the programs the Oracle runs, and produce the Agent and the Workflow as the two placements of a task's program. For internally autoregressive Oracles, the two computations synchronize at the end of every answer under certain conditions, and through that synchronization we model caching and analyse scheduling. No guarantee that holds for every Oracle can fix which content crosses between the two computations, but such a guarantee does fix the boundary itself. The construction V fits the machine to a von Neumann computer. To show that it is realizable, we propose ArchNights, an extended RISC-V ISA and a Linux-style operating system implementing the machine by design. ArchNights-SE runs on gem5 as a computer system, becomes an agentic system when it runs an LLM as the Oracle, and will be open source. Agentic systems can then be designed as computer systems are. With a foundation built and a unified view, future work can share invariants and bounds, each with its conditions.

Modern web agents built on large vision-language models process webpages, select relevant UI elements, and translate model outputs into browser actions. Existing visual red-teaming approaches use adversarial visual content to manipulate this process. However, they primarily target model inference and do not explicitly account for structured input processing or action post-processing. Consequently, model-level success does not establish control over browser execution and cannot reliably characterize end-to-end agent robustness. To address this gap, we formulate red teaming for vision-grounded web agents as an end-to-end grounding-to-execution problem, and introduce WebMirage, a framework that crafts localized visual perturbations that cause agents to select attacker-controlled content and execute the corresponding browser action across varying webpage renderings. It uses a role-slot abstraction and webpage recomposition to capture competition among webpage elements, and dataflow analysis to align optimization with action post-processing. We evaluate WebMirage across four agent configurations and six VLM backbones on 2,250 tasks covering 13 public websites and a sandbox benchmark. WebMirage achieves an average attack success rate of 91.9%, compared with 17.4% for the strongest baseline, and remains effective against three agent-level defenses.

Self-improving LLM systems propose changes to themselves and keep those that score better on a small evaluation set. We treat this keep-if-better step as selection under measurement noise, model the correlated errors of the candidates in a single decision, and study empirically what happens when the evaluation set is reused. In runs where Qwen models rewrite their own instructions and every candidate is also scored on 600 held-out items, most proposals after the first are harmful, and the model gives the size of the winner's curse of a generation's best candidate. With a prior from a separate pilot, it matches the average overstatement of first-generation commits in native loops, though not setting by setting. In a pre-registered study, the final selection-set score of greedy loops exceeded held-out accuracy by 13 to 20 points with 16 selection items and by 1 to 5 points with 256. Held-out gains grew with the selection set on TREC but not on GSM8K, and the tested acceptance rules did not beat greedy acceptance over whole runs. Gains measured on the selection set also exceeded held-out gains when a current model refined a competent instruction, and in the validation scores of GEPA and MIPROv2. Scoring the starting and the current instruction on 64 items never used for selection removes the average bias of a loop's reported gain, but single estimates remain off by about 6 points. Self-improvement studies should report held-out gains with their uncertainty.

Rapid advances in AI have opened a new era for physics, enabling unprecedented capabilities in research and discovery. However, these advances also bring challenges in verifying, trusting, and reusing the results they so rapidly produce. Formalization is a necessary part of the solution to these challenges. In this paper, as well as giving an overview of the role of AI and formalization in physics, we make the case that: (1) formalization is an essential layer for AI-driven discoveries in both computational and theoretical physics; (2) AI is lowering the barrier to, and increasing the efficiency of, formalization in physics, though it remains weaker at choosing definitions and designing libraries than at writing proofs; (3) the speed AI brings to formalization in physics currently comes at the cost of trust, credit, and community; (4) shared, community-reviewed formal libraries such as Physlib should be central in making progress in this field, both as a trusted, reusable foundation for physics results and as the environment in which AI systems can reason about physics; (5) formalizing physics raises challenges beyond those in mathematics, most importantly that a checked proof does not establish that its definitions faithfully capture the intended physics. We also give our perspectives on near-term and long-term directions in these areas.

Sensing-aided beam prediction can reduce the beam-training overhead of millimeter-wave (mmWave) systems by leveraging environmental information from complementary modalities such as vision and position. However, the resulting sensing and inference pipeline can be energy-intensive, motivating the use of energy-efficient spiking neural networks (SNNs) as an alternative to conventional artificial neural networks (ANNs). In this work, we develop SNN architectures for beam prediction and systematically compare them with ANNs using position-only, vision-only, and multimodal inputs. Our results show that compact ANNs consistently outperform SNNs for position-only sensing. In contrast, for vision and multimodal sensing, SNNs achieve higher accuracy than ANNs at similar energy costs and require up to an order of magnitude less energy at a comparable accuracy. For example, the multimodal SNN achieves 58.08% top-1 accuracy, outperforming the closest-energy ANN by 18.11 percentage points. We further show that SNN accuracy is limited by the binary communication between layers rather than by model size. These results demonstrate the potential of SNNs for energy-efficient sensing-aided beam prediction while highlighting their accuracy limitation.

Public leaderboards for AI models are read continuously, and attackers can see every published standing. Vote rigging, selective disclosure of private variants, and benchmark contamination can each move a ranking. Existing guarantees assume genuine records or bound the corruption per step, which an attacker who corrupts in bursts evades. We introduce the certified corruption budget, a tolerance $\widehat{B}_t$ computed after $t$ records and published with each pairwise claim. With probability at least $1-\alpha$, simultaneously at all times, the claim is correct or more than $\widehat{B}_t$ records were corrupted. It holds against attackers who watch every certificate, with no bound on their budget. Forged records and records altered once seen require different certificates: the certificate for forgeries fails, with probability approaching one, against an attacker who flips votes it has seen, while one that charges roughly twice as much per record remains valid, with constant bets even against attackers who see the future, and no smaller charge is valid at every level. The certified budget grows nearly as fast as any valid method allows: with a win fraction $\frac{1}{2}+\delta$, each new record adds close to $2\delta$ to the number of forged records the claim can withstand ($\delta$ flipped). Publishing the best of $V$ private variants costs only an amount growing like $\log V$. In replays on 1.8 million Chatbot Arena votes, a few hundred rigged votes make standard confidence intervals certify false orderings, while ours stays valid. On real votes, our certificate shows that clearly separated models withstand about 2,000 forged votes.

Tool-using agents repeatedly carry observations whose useful content can be much smaller than their original payload. We study agent-controlled forgetting: the acting model selects previously observed tool results, replaces each with a short note at its original position, and retains the exact original in a recoverable archive. A Python harness exposes batch archival and explicit recovery without task-specific model training, while protecting user instructions and assistant messages from these operations. In an exploratory OpenTelemetry debugging case followed by an unrelated implementation task, the method ended with 231,951 provider-reported prompt tokens versus 912,492 under retained history, used 50% fewer cumulative input tokens, and had an estimated API cost of USD 1.28-1.44 versus approximately USD 4.38. Both arms passed the two-case primary behavioral oracle; neither fully satisfied the follow-up evaluation. The method made more requests and took 17% longer. A contrasting application-development pair produced no context or cost saving, and an earlier continuation exhibited lower manually assessed quality despite reduced context. These observations demonstrate substantial resource savings in noisy tool-use trajectories and identify workload dependence as a central consideration for reversible context management.

Large language models (LLMs) are increasingly being integrated into hardware verification to automate specification interpretation, verification-artifact generation, debugging, formal reasoning, and tool orchestration. This survey provides a systematic review of LLM-assisted hardware functional verification across SystemVerilog assertion generation, stimulus and testbench generation, bug localization and design repair, model checking and equivalence checking, SAT/SMT optimization, and emerging agentic verification workflows. We organize the literature by methodology, verification objective, tool interaction, benchmark, and evaluation criterion, and examine both inference-time techniques–including prompting, retrieval, structured reasoning, and agentic workflows–and training-time adaptation. Across these areas, a common pattern emerges: LLMs are most effective as semantic reasoning, search, and orchestration components embedded within verification-aware workflows, while simulators, formal engines, coverage tools, and solvers provide executable feedback and correctness evidence. However, tool acceptance alone does not establish verification correctness, since assertions, tests, repairs, or proofs may satisfy available checks without faithfully capturing the complete design intent. We therefore identify semantic alignment between specifications and verification evidence, scalable integration with deterministic tools, generalization to unseen designs, and rigorous evaluation of correctness, cost, robustness, and human effort as key challenges. Finally, we discuss emerging directions toward specification-centered, neuro-symbolic, and persistent agentic verification systems that combine LLM flexibility with independently checkable verification evidence.

Scientific studies of language agents need behavioral variables that support hypotheses across tasks and models. We formulate this research problem as learning and testing a hierarchy of trajectory abstractions. A concrete recursive procedure first measures role- and phase-indexed events, proposes temporally constrained relations, and tests their stability across conditions. It then constructs episode-level motif variables from selected relations and repeats the analysis on those variables. Explicit measurement functions connect every abstraction level to the original trajectories. Observations and randomized protocol experiments assess the resulting hypotheses, while comparisons between intervention realizations determine whether an abstraction should be retained, refined, or restricted. We derive a finite-depth bound for accepted reductions, identify protocol effects on fixed abstractions, and characterize realization disagreement and composition of abstraction error. A finite-sample test makes projected intervention consistency operational, and constructed examples illustrate motif construction and abstraction refinement. The formulation distinguishes this experimental approach from semantic taxonomies, qualitative theory induction, and behavior-model recovery. It specifies a proposed research procedure for discovering generalizable behavioral hypotheses, with literature-relative novelty assessed separately from model-relative surprise.

LLM-as-judge is now a standard tool for scalable evaluation, but judge performance is still often summarized by a single accuracy number. This aggregate view hides the deployment conditions under which a judge succeeds or fails. We introduce Conditional Accuracy Profiling (CAP), a post-hoc diagnostic framework that decomposes pairwise LLM-judge accuracy into eight conditions organized into content sensitivity, robustness, and rationale quality. CAP is benchmark-agnostic: it can be applied directly when a benchmark provides the required annotations, approximately through task-subset proxies, or through controlled augmentation when perturbation pairs can be generated. We instantiate CAP on seven LLM judges across six pairwise judging benchmarks, including judgerEva-Standard, a controlled testbed we created to support all eight conditions. CAP exposes profile differences hidden by aggregate accuracy: on judgerEva's judge-independent Hard-Constructed subset, the two judges most sensitive to omitted qualifications rank in the bottom three of seven by overall accuracy, so omission sensitivity is not predicted by aggregate accuracy. Across benchmarks, Position Robustness shows the strongest rank stability (mean Spearman $\bar{\rho}{=}0.87$) but is itself fragile under JudgeBench-Pro adversarial stress, showing the largest mean accuracy drop among the shared conditions, though the dominant degradation channel varies by judge. Condition-level profiles provide a more actionable basis than aggregate accuracy for selecting LLM judges.

Flow and diffusion models suffer from slow inference due to computationally expensive numerical integration. Distillation provides a promising way for a student model to learn from a teacher's dynamics, enabling one-step or few-step generation. However, existing methods often depend on curated distillation datasets, costly teacher rollouts, or auxiliary proxy networks, which complicate model training and scaling. In this work, we propose Consistent Distribution Matching, a simulation-free and data-free distillation method for accelerating diffusion and flow models while preserving strong generative capacity. Our key insight is to unify sample generation and score estimation with one student network. Thus, our framework uses only two models, a frozen teacher and a trainable student, and optimizes one objective. We prove that minimizing our objective indicates Wasserstein convergence of the student flow-map pushforwards to the teacher marginals. On ImageNet 256$\times$256, our method attains an FID of 2.04 with a single function evaluation (1-NFE) and a 4-NFE FID of 1.37 within 40 epochs of training, surpassing the state-of-the-art distillation baselines without data. Our code code and model are available at https://consistentdmd.github.io/.

Direct Preference Optimization (DPO) treats all constraint violations equally: a $1 budget overshoot and a $1,000 overshoot induce the same training signal. It is also susceptible to length and style bias when preference pairs come from different model families. We introduce Constraint-Margin DPO (CM-DPO), which replaces DPO's binary preference signal with a continuous margin derived from a deterministic symbolic verifier and scaled by violation severity. Hard and soft constraints are separated through a lexicographic objective, ensuring hard constraints are never traded off against preferences. To supply CM-DPO with bias-reduced training pairs, we generate preference data through procedurally generated constraint profiles (DCCG) and minimal-edit distillation from a reasoning teacher (RT-MED), within a framework we call SynPlan-R. On TravelPlanner, NaturalPlan, and out-of-distribution PlanBench, an 8B model fine-tuned with CM-DPO achieves 89.2% pass rate and 93.4% solve rate, matching multi-agent systems at 13x lower latency while outperforming GPT-4o on unseen Blocksworld by 9.2 points.

LLM-guided program evolution has enabled discoveries in mathematics and computational optimization, raising the prospect of reinforcement learning (RL) algorithms that self-evolve to improve how agents learn. However, realizing this prospect faces two obstacles. Joint search over coupled algorithmic components is difficult to scale: simultaneous changes can disrupt learning, while isolated changes overlook their dependencies. Evaluating candidate algorithms also requires costly training, with fitness remaining uncertain across random seeds. We introduce RLDiscover, a framework for the self-evolution of model-free deep RL algorithms. Progressive Co-Evolution advances from targeted component edits to joint evolution, while Progressive Probabilistic Evaluation balances search breadth and evaluation fidelity through staged training and repeated evaluation. Experiments across SAC, PPO, and DQN on four benchmark suites show substantial improvements in mean return, with per-family median gains of 32%-84% and a peak return ratio of approximately 363x over a near-zero baseline. These gains include transitions from failed learning to successful task completion, and improvements persist when evolution starts from stronger open-source implementations. On measured SAC locomotion runs, evaluation uses approximately one-fifteenth the estimated compute required to fully evaluate the same candidate pool. Remarkably, independent searches repeatedly discover interpretable combinations of adaptive robust losses, progress-dependent value targets, and running statistics, with selected programs transferring to unseen tasks. These findings point toward a broader role for self-evolution in AI: discovering interpretable algorithms that improve how agents learn.

Autism spectrum disorder (ASD) is diagnosed through specialist observation of a child's social behavior, and access to that expertise is the bottleneck for early identification. Vision-language models (VLMs) describe a child's behavior from video well; the verdict drawn from the description is unstable: at temperature~0, across eight pipeline configurations on one backbone, 16–37% of clips change their predicted label between repeated runs, and the cause lies in the serving stack. We keep the VLM frozen and move the decision out of the model. Under our grounded perception constraints the VLM writes an event table of timestamped, glossary-labeled events that names the eliciting press and logs counter-evidence; a text-only stage age-calibrates the confidence of each row; a deterministic weight-of-evidence scorer sums it into an evidence total and stratifies it into a risk category, so every decision decomposes into named per-feature contributions and can be re-scored from the saved table. On 43 caregiver-recorded, protocol-free home free-play clips of preschool children, the pipeline reaches AUC $0.851 \pm 0.012$, 86.0% accuracy, and F$_1$ 71.8 over three runs. It labels 74.4% of clips correctly in every run (60.5% for the zero-shot baseline) and flags no typically developing clip in every run (9 of 31 at zero-shot). An ablation on the same backbone attributes the gain to the grounded perception constraints read through the deterministic scorer.

Given a task and an evaluator, a language model can rewrite a candidate program while a search loop decides which rewrites survive, offering a practical route to algorithm discovery. But that loop is governed by five constants set by hand: which parent to select, how hard to mutate, how to keep diversity, what to remember, and a scalar score that never says which part of the program earned it. Reinforcement learning already has an estimator for each. The obstacle is that program evolution is not usually written down as a decision process. We formalize it as a Markov decision process whose action is the modular prefix the model is conditioned on, rather than the program it emits. Credit assignment, value estimation, adaptive exploration, and experience memory can then attach to distinct components. AGAR (Algorithm Generation As RL) provides the resulting substrate: any estimator can be replaced or switched off without changing the controller, making the transfer auditable one mechanism at a time, with no gradient training of the backend model. Across 19 tasks, two backends, and three seeds under one harness, AGAR improves on the stronger of two published baselines on most tasks, with gains concentrated in the competitive-programming family. The formalization also yields a checkable reading of prior work: these systems are implicitly zero-discount, not by choice, but because fitness is exogenous to an individual rather than a return over successors, leaving a discount factor nothing to act on.

The best two-bit weight quantizers for large language models, such as QTIP and Proteus, rotate each weight matrix by a random orthogonal transform, which must be undone at every decoding step, then encode it with a trellis or lattice code under a Euclidean search; the layer Hessian enters only through error feedback between coding blocks. We show that this leaves part of the Hessian unused. Error feedback turns the loss into a weighted sum of per-coordinate rounding errors whose weights, the diagonal of the Hessian's LDL factorization, existing quantizers compute but never read. We put these weights into the Viterbi branch metric, so the search follows the curvature within each coding block. This also explains the rotation: it removes this within-block variation, so weighting in the native basis and rotating are substitutes. On three models the weighted native search matches a full-dimension randomized Hadamard to within about one point of downstream accuracy, and weighting after the rotation gains little. Around this search we build CurveTQ, a trellis codec with no rotation, which handles the weights' amplitude and marginal shape with a factored scale field and a closed-form quantile table, and stores a start state per coding block so the trellis can adapt to the residual that error feedback carries into it. At two bits CurveTQ is 1-3 points higher in mean downstream accuracy than QTIP and Proteus on three 4-8B Instruct models, even after both are given our start state, which alone lifts either baseline by 1-3 points. It also leads on a 35B mixture of experts, to our knowledge the first trellis-coded result on such a model. With no rotation to undo, our decoder is the fastest of the three at all tested batch sizes and bit widths.

Substance Use Disorder (SUD) counseling requires patient responses that reflect underlying cognitive states such as beliefs, coping strategies, and readiness for change. Although large language models (LLMs) can generate fluent text, they often fail to produce cognitively coherent and clinically realistic patient behavior, especially under ethical and data-scarce clinical settings. Moreover, deploying frontier-scale LLMs in healthcare applications presents practical challenges including high computational cost, latency, privacy concerns, and limited deployability in resource-constrained environments, motivating the need for cognitively aligned small language models (SLMs). We propose a cognitively grounded framework for SUD patient dialogue generation that explicitly models and aligns latent cognitive components with patient histories and counselor questions. Our pipeline consists of two stages: cognitive component detection and cognitive component-aligned dialogue generation. To enable effective learning with smaller models, we combine knowledge distillation from high-capacity teacher models, preference optimization from human-annotations, and attention-guided reward shaping. Extensive evaluations using automatic scores like BERTScore, ROUGE, METEOR and BLEU, and LLM-as-judge hit-metrics against both human and teacher-model references show that cognitively informed fine-tuning substantially improves cognitive realization and alignment over a generic instruction-tuned baselines and mental health domain specific SLMs, with particularly strong gains for open-ended cognitive components.

Modern image editing models can satisfy a text instruction while breaking the physics of the edited scene. A new object may cast no shadow, a mirror may fail to reflect visible geometry, or an object may float above a surface that should support it. We study physical plausibility diagnosis, detecting whether an edited image violates scene physics, naming the violation type, localizing the affected region, and explaining the failure in language. We introduce a counterfactual benchmark whose controlled synthetic component uses Mitsuba~3 to generate 5,500 images from 500 scene families. Each family contains one clean image and ten matched violations involving shadows, reflection, support, surface response, and occlusion. The renderer pipeline provides category labels, affected-region masks and boxes, scene metadata, and explanation targets. We use LLaVA-1.5-7B, Qwen2.5-VL-7B, and InternVL3.5-8B as diagnostic baselines rather than proposed methods. On a 1,650-image synthetic test set, the adapted baselines reach 64.0–67.8% category macro-F1 on standard held-out scenes. For LLaVA-1.5-7B, category macro-F1 falls from 64.0% on the standard split to 40.8% under intervention shift. This gap shows that high in-distribution accuracy partly reflects cues tied to rendering and counterfactual construction.

Extreme low-bit LLM compression is most challenging when weights, activations, and KV caches are quantized together: their distributions differ, and quantization errors interact throughout the network. We introduce CanonQ, a unified quantization-aware training framework that addresses these challenges by separating source canonicalization from task-aware adaptation. Fixed rotations and energy normalization map heterogeneous tensor sources to canonical coordinates, enabling frozen Gaussian-reference codebooks to be reused across layers and models. Joint training then adapts the network to the coupled errors of weight, activation, and cache quantization within a common scalar/vector interface. We bound frozen-codebook transfer error and local task loss, and derive an exact normalization-aware straight-through Jacobian that links quantization distortion to gradient bias. The strongest gains arise under joint W2A4KV2 compression: across LLaMA3-1B/3B/8B, CanonQ-Omni achieves up to 14.28x lower WikiText-2 perplexity and up to 57.9% higher mean zero-shot accuracy than prior state-of-the-art and representative quantization baselines. The benefits extend to Qwen3-1.7B, code generation, and mathematical reasoning: on instruction-tuned MobileLLM-Pro-1B at W2A16KV16, CanonQ achieves relative improvements of 41.7% in HumanEval pass@1 and 39.1% in GSM8K exact match over the strongest evaluated quantization baseline.

Agent evolvers automate the design of the prompts, skills and workflows around language model agents, yet the optimization process they follow is still designed by hand: a fixed search loop decides how candidates are evaluated, which are kept and when the search stops. We propose FREEEVOLVE, which automates this process as well. An environment specifies the goal, target agent, evaluator, data and resource limits; within these limits, the evolver itself decides what to test, how much evidence to collect, which candidates to pursue and when to stop. These decisions follow an editable evolution skill, which we improve through meta-evolution by scoring each candidate skill on the fresh target agent it produces. The optimization process thus becomes a capability learned from experience rather than a loop engineered in advance. On tau3-bench, ARC-AGI-2, ARC-AGI-3 and Terminal-Bench 2.1, FREEEVOLVE controls the evolution campaign by itself, yet improves the primary held-out metric by 13.6 points on average and matches or exceeds hand-designed evolvers. The learned process keeps improving with experience: meta-evolved skills add 6.9 points over the seed skill on fresh target agents, demonstrating transferability across environments.

Longitudinal models forecast how a patient's imaging state evolves, but accuracy does not show whether the patient's observed trajectory drives the prediction. A population-average forecast may be useful but cannot establish a patient-specific world-model claim. We introduce Patient, Place, Prior (P$^3$), an audit asking whether a forecast benefits from the patient's longitudinal imaging history (Patient), benefits from patient-matched externally supplied spatial support (Place), and gains predictive value beyond a population-average prediction under matched support and context (Prior). We also propose Cancer JEPA, a one-step model that forecasts frozen representations of future breast dynamic contrast-enhanced MRI examinations during neoadjuvant therapy. It adds a lesion-constrained neural correction, trained with an occlusion-based latent objective, to a patient-conditioned low-complexity reduced-rank regression baseline. This factorization permits a post-hoc P$^3$ audit of the frozen model. In a validation cohort previously used in development, forecast error is lower when the neural correction receives the patient's history rather than another patient's and patient-matched lesion occupancy maps rather than substituted maps. However, the descriptive 95% interval comparing the correction computed from patient history with the population-average neural correction includes zero. P$^3$ thus separates input use from evidence of patient-specific predictive value beyond a population-level pattern.

MemoryAgentBench's Conflict Resolution split is read as measuring "selective forgetting". We execute the benchmark's own rule - the newest statement about a fact wins - as a zero-learning resolver frozen on one of the four fact lists. Under the official metric the rule answers 80.25% of the questions (74.5% on the three held-out lists). Of the rest, 67 items have a released gold that the last-write graph cannot reach but overwritten statements would ("The capital of India is New Delhi." superseded by "The capital of India is Grosseto."; gold New Delhi); such items are a third of the multi-hop questions at 262K. Two long-context models and our pre-registered approximate re-implementation of the benchmark's BM25 agent, one retained run per item and outcomes only, score 84.7%, 82.6% and 41.6% on the items the rule solves against 10.4%, 11.9% and 6.0% on those 67. The failures are a reachability split plus a small parser-scope residual; the per-item split, not the aggregate, is the unit at which a score here can be read.

Probability-only models, which TypeSafe calls System One models, return calibrated probabilities for fixed choices in milliseconds and generate no text. We study one such model, Jev, through two tasks that require decisions under tight constraints. In bullet chess, a bot that places Jev's judgment inside Stockfish search alongside an opening book and endgame tablebases climbs above a 2200 Lichess bullet rating against other bots. Live model calls are too slow for search, so we distill pairwise judgments into a compact evaluator that runs at every position. We then ask how best to spend a fixed labeling budget when an LLM, Qwen3-32B, is available as a second teacher. In chess, averaging both judges' labels beats spending the whole budget on Qwen alone by 9.6 Elo (95% interval 4.3 to 14.9), and the gain replicates on fresh openings; a second answer from the same judge is no substitute, and Jev is the strongest partner for Qwen among the models tested. In passage reranking, Jev's labels alone train a reranker that scores as high as Qwen's, from 21 minutes of API calls instead of 5.1 GPU-hours, and adding Qwen gains at most a few thousandths in ranking quality. Search supplies the lookahead, distillation makes the judgment cheap enough to use at every position, and an LLM partner pays off in chess.

As data propagates through a Transformer, the norm of its hidden states grows by orders of magnitude with depth, a phenomenon framed as 'curse of depth' and nearly universally treated as a pathology to be suppressed. We take the opposite view. Across 16 pre-trained LLMs from 9 families, spanning dense, mixture-of-experts and hybrid architectures and Pre-, Peri- and Post-Norm designs, we find that this growth reflects an emergent depth-positional encoding, carried by the only learned per-layer gain on the residual stream, the normalization weight $\gamma$: with depth, $\gamma$ grows in magnitude and rotates in direction, jointly encoding the layer index. We make this depth-conditioned encoding explicit with LayerRoPE, an implicit analog of RoPE along the depth axis, which replaces all layerwise $\gamma$ vectors with a single shared vector and depth-conditioned scalars, at a net reduction in parameters and $<0.02\%$ change in FLOPs. Across a model ladder scaled up to $100$B+ tokens, LayerRoPE consistently outperforms Pre-, Post- and Peri-Norm and Layer-Norm Scaling, reaching Pre-Norm's 1.3B loss with $3.4\times$ less compute; LayerRoPE is the only approach that shows strong convergence and improves near monotonically as depth scales to 512 layers. It improves learning-rate sensitivity by $3$-$10\times$, and transfers naively to and consistently improves looped latent models and Vision Transformers. Inspecting its learned schedule inverts the prevailing premise: LayerRoPE does not shrink the residual stream but widens it, damping what each block reads while amplifying what it writes. Depth stability, our results suggest, calls not for suppressing the residual stream, but for depth-conditioned regulation of the computational blocks it feeds.

Standard diffusion samplers generate samples through repeated evaluations of a learned score function. Parallel sampling methods seek to accelerate generation by trading additional evaluations for fewer sequential rounds. This raises the question of how much sequential dependence is unavoidable, even when many score queries can be made simultaneously. We establish the first polynomial parallel-round lower bounds for diffusion sampling with approximate scores. Specifically, we prove (1) a $\widetilde{\Omega}(d^{1/3})$-round lower bound for sampling smooth, near-isotropic Gaussian mixtures in $R^d$, and (2) an $\Omega(d)$-round lower bound for uniform sampling from anisotropic axis-aligned boxes contained in the unit ball. Both bounds hold for arbitrary randomized algorithms making polynomially many queries per round at arbitrary locations and noise levels, with inverse-polynomial score error and constant total variation accuracy. The linear bound is tight for our box family. Our constructions use fixed approximate score oracles that enforce sequential access to hidden information while satisfying the accuracy guarantee at every noise level.