Multimodal memes are dynamic and often require up to date background knowledge for interpretation. Existing methods often overlook such knowledge or rely on fixed parametric knowledge of pretrained models that may be incomplete, outdated, or unavailable for emerging memes. We introduce Query Retrieve Conclude, a zero shot framework that identifies missing knowledge, retrieves open web evidence, and synthesizes evidence grounded background knowledge for meme understanding and detection. We also introduce a curated meme understanding benchmark of recent memes from 2024 to 2026 with external background knowledge annotations. Experiments on three meme understanding datasets and five meme detection tasks show that our framework improves knowledge recovery, meme understanding and downstream detection over zero shot baselines.

Implicit chain-of-thought (iCoT) methods aim to internalize reasoning in large language models, but often underperform explicit CoT prompting. We empirically find that hidden-state reasoning trajectories exhibit low-rank structure. Motivated by this observation, we propose a low-rank distillation framework that transfers reasoning by aligning teacher and student trajectories in a shared low-rank tensor subspace using first- and second-order statistics. The resulting formulation captures the global structure of reasoning while supporting a compact latent reasoning process. We evaluate the method across multiple model families, including LLaMA and Qwen, at different scales on mathematical reasoning benchmarks. Our approach consistently improves performance, especially on challenging multi-step tasks, approaching explicit CoT accuracy and outperforming prior iCoT distillation methods.

With PRECISE, we extended Prediction-Powered Inference to produce bias-corrected estimates of ranking evaluation metrics by combining a small human-labeled set with a large LLM-judged set. PPI is provably unbiased regardless of the LLM judge's error profile. We make it applicable to hierarchical metrics like Precision@K, where annotations are per-document but the metric is per-query, by reducing the output-space computation from O(2^|C|) to O(2^K). On the ESCI benchmark, augmenting 30 human annotations with Claude 3 Sonnet judgments reduces the standard error of Precision@4 estimates from 4.45 to 3.50 (a 21% relative reduction). In a production system, our framework correctly identified the best of three system variants from 100 human labels and 2 hours of domain-expert annotation; A/B testing confirmed this ranking with +407 bps in daily sales.

Multi-agent systems (MAS) built on large language models are typically organized around roles, pipelines, and turn schedules, while the content that agents pass to one another is often left as unconstrained natural language. However, this free-form communication can rapidly inflate token usage, consume the shared context window, and ultimately affect both system performance and inference cost. We analyze five common inter-agent communication strategies across two MAS topologies, finding that no fixed strategy is universally optimal. Instead, effective inter-agent messages consistently preserve action-centered information needed by downstream agents. Building on this, we propose the PACT (Protocolized Action-state Communication and Transmission), which treats inter-agent communication as a public state-update problem and projects each raw agent output into a compact action-state record before it enters shared history. Across different MAS topologies, PACT consistently improves the performance-cost trade-off, achieving comparable or stronger task performance with substantially fewer tokens. The gains extend to production coding harnesses: PACT lifts OpenHands' resolve rate at -10% tokens-per-resolved, and is resolve-neutral on SWE-agent while halving input tokens. Our code is publicly available at https://github.com/iNLP-Lab/PACT.

LLM agents operate in two distinct regimes: open-weight agents amenable to reinforcement learning (RL) and black-box agents whose behaviour must be controlled purely at test time. Although black-box agents are often backed by state-of-the-art proprietary LLMs, API-only access precludes parameter-level optimization, rendering most RL methods inapplicable. To address this limitation, we turn to a known equivalence between RL and Bayesian inference. We propose Agentic Monte Carlo (AMC) to directly sample from the optimal policy of a black-box agent rather than training it through RL. The optimal policy is a posterior over trajectories whose prior we define as the fixed black-box LLM agent. We employ Sequential Monte Carlo to sample from this posterior by learning a value function to steer the agent while leaving the underlying black-box model unchanged. We validate AMC on three diverse environments from the AgentGym benchmark, demonstrating significant improvements over prompting baselines and even outperforming Group Relative Policy Optimization (GRPO) as we scale the test-time compute of our method. AMC demonstrates the feasibility of performing principled RL-style optimization of black-box LLM agents. Code is available at https://github.com/layer6ai-labs/Agentic-Monte-Carlo

Recent progress in generative modeling has made safety control a central challenge, yet existing approaches remain largely model-specific, requiring retraining or tailored interventions for each new architecture. In this work, we ask whether safety can be represented as a portable latent direction, learned once and reused across heterogeneous generators. We introduce the first framework for cross-model safety steering, in which a safety direction is estimated in a source LLM from paired safe-unsafe prompts, transported to a target generator through a lightweight alignment fitted on benign data alone, and applied at inference time. Crucially, our pipeline never accesses unsafe data on the target side, isolating whether safety can be transferred through shared representation geometry. Beyond a single global direction, we also identify a multi-vector extension that captures category-specific safety behaviors, enabling more selective control. We evaluate our approach in text-to-image and text-to-video generation across diverse source-target model pairs. Across models, transferred safety directions achieve ASR reduction and CLIP-Score/FID trade-offs comparable to directions learned natively on the target model using unsafe data, while requiring no target-side unsafe data. This indicates that safety improvements do not come at the expense of generation quality. Our results point to a modular view of safety: safety-relevant behavior is not purely model-local, but can be controlled through latent directions that persist across models. This suggests a new path toward lightweight, reusable safety mechanisms that do not require target-side unsafe data.

We study the personal camera roll visual question answering setting. In this setting, a conversational AI assistant can access a user's personal camera roll and retrieve relevant photos to answer queries, ranging from simple factual questions (e.g., ``Name of the food I tried yesterday?'') to more open-ended ones (e.g., ``Recommend some dishes I have never eaten before''). Given the vast nature of the personal camera roll (i.e., multiple years, hundreds to thousands of photos), a successful AI assistant needs to understand a long-horizon, highly personalized visual content stream in order to navigate and locate the correct and/or relevant information. To support this, we collect and manually annotate questions that mimic real-world usage. The final dataset, camroll, contains 50 users, 31,476 images, and 2,500 QA pairs. We further design camroll-agent, a conversational AI agent equipped with hierarchical memory and a minimal set of tools for efficient navigation over large, personalized visual memory. Experimental results show that camroll-agent outperforms numerous baselines and methods for long-context understanding AI agents system. Together, the camroll dataset and camroll-agent highlight the gap in AI agents' long-context reasoning: personalized visual memory requires different approaches from standard long-context textual memory, especially when consistency, visual details, and user-specific context are present.

Electro-Hydrostatic Actuators (EHAs) are widely used in aerospace and industrial systems, where timely detection of sensor anomalies is essential to ensure safe and reliable operation. However, the large volume and high sampling frequency of EHA sensor data pose challenges for accurate and efficient anomaly detection. Conventional statistical and classical machine-learning methods such as Z-score, Interquartile Range (IQR), Median Absolute Deviation (MAD), Isolation Forest, Gaussian Mixture, and k-means often fail to capture the temporal dependencies inherent in EHA signals, resulting in limited detection accuracy and elevated false-alarm rates. Furthermore, systematic evaluations of data-driven anomaly detection approaches for EHA systems remain scarce, particularly under varying operational conditions. This study presents an offline anomaly-detection framework for univariate EHA sensor signals, focusing on temperature and pressure data collected from a controlled test bench. The method employs a reconstruction-based Long Short-Term Memory (LSTM) autoencoder, calibrated and evaluated using validation-set reconstruction-error distributions. Performance is assessed across multiple fault-injection scenarios using accuracy, precision, recall, and F1-score, complemented by sensitivity analyses under varying operating conditions. The LSTM autoencoder achieved an average accuracy of 99.0\%, precision up to 100\%, recall between 90.2\% and 99.6\%, and F1-scores from 93.1\% to 99.8\%, demonstrating high detection sensitivity and a very low false-alarm rate across all evaluated sensors. These results highlight the feasibility of data-driven offline anomaly detection for EHAs. Future work will focus on adapting the developed framework for an online (real-time) environment.

Neural rough differential equations (NRDEs) stay accurate under irregular sampling while taking far fewer integration steps than standard neural differential equations, summarising a finely sampled driver by its log-signature and advancing the hidden state over coarse intervals using the log-ODE method. This efficiency rests on the shuffle algebra, the algebraic counterpart of Stratonovich calculus. This reliance means NRDEs cannot expose the quadratic-variation terms Itô dynamics require, nor the ordered covariant derivatives that govern Itô flows on connection-equipped manifolds. Ameliorating this, we introduce Branched Neural Rough Differential Equations (B-NRDEs), a Hopf-algebraic framework that recasts the NRDE log-ODE step as geometric numerical integration on the state-space manifold, matching the driving algebra to the governing calculus: Grossman--Larson rooted trees for Euclidean Itô dynamics, Munthe-Kaas--Wright planar rooted trees for ordered covariant derivatives on manifolds, and the shuffle algebra in the classical Stratonovich case. This yields intrinsic coarse-step dynamics that exactly preserve manifold constraints. Finally, we introduce a branched signature-kernel objective to enable Itô-consistent law matching by making quadratic-variation terms visible during training. On rough Bergomi volatility, sim-to-real $\mathrm{SO}(3)$ dynamics forecasting, and SPD covariance dynamics, B-NRDEs offer a unified, effective approach to stochastic and manifold-valued dynamics beyond the Euclidean--Stratonovich setting.

We establish the first sharp thresholds for low-degree polynomial tests in planted-vs-planted settings, where the goal is to determine with vanishing error which of two structured planted mechanisms generated the observed data. We prove matching low-degree upper and lower bounds for counting communities in the planted submatrix and planted dense subgraph models. The resulting testing threshold coincides, down to the sharp constant, with the known low-degree recovery threshold. In contrast, the task of weak testing, where the goal is to outperform random guessing, does not have a sharp threshold but rather a smooth transition, which we identify. To prove our results, we develop a framework for planted-vs-planted testing that builds on a latent-variable expansion originating in low-degree recovery and employs new methods to identify and prune non-signal contributions.

Near-real-time flood depth prediction demands surrogate models that are accurate, fast, and transferable across watersheds. Supervised surrogates can match physics-based simulators in accuracy but need millions of training rows per watershed and cannot extrapolate beyond their original mesh. We propose a domain-aware coreset construction pipeline that conditions a tabular foundation model at inference time. The pipeline stratifies storms by return period and most-affected watershed, then samples hexagons with a target-aware spatial selector. With 0.7% of the per-watershed training pool, the model attains a mean $R^2$ of 0.663 across nine Houston-area watersheds, within 98.5% of the supervised reference ($R^2$ = 0.673). It transfers to held-out watersheds without task-specific retraining, staying ahead of a coreset-trained supervised baseline. On real storms it exceeds the supervised reference on a far out-of-distribution case and trails it on a mostly in-distribution one. Domain-aware coreset construction lets tabular foundation models deliver data-efficient, watershed-transferable flood predictions without per-watershed training.

Training robust multivariate time series forecasting models requires large, diverse corpora, yet many real-world domains provide only a handful of observed sequences. Existing generators fail to resolve this mismatch: prior-based approaches (e.g., CauKer, TimePFN) produce domain-agnostic samples, while data-driven methods (e.g., TimeGAN) treat references as black-box supervision, forfeiting explicit control over periodic structure, local variability, and cross-variable dynamics. We propose ReGeN, a reference-guided generative pipeline that treats observed sequences not as examples to imitate, but as structural scaffolds for controllable synthesis. ReGeN decomposes each reference into three interpretable components: a phase-aligned periodic backbone capturing dominant domain morphology; per-variable stochastic residuals modeled with a deep-kernel Gaussian process; and lag-aware cross-variable dependencies injected through a structural causal model with fitted coupling coefficients. Sampling these components at controllable temperature broadens distributional coverage while preserving domain-grounded structure. We show that ReGeN-generated data consistently substitutes for real sibling data with minimal forecasting degradation, and in strongly periodic domains such as traffic, can outperform the real source itself. We further show that a foundation model pretrained on ReGeN corpora outperforms those pretrained on prior-based and data-driven synthetic alternatives. This suggests that in low-data regimes, how reference data is structurally exploited can matter as much as how much data is available.

Reinforcement learning with verifiable rewards improves reasoning and tool use, yet long-horizon language agents still learn unsupported evidence chains, belief drift, and shortcut actions that satisfy terminal checks. Existing process rewards are mostly correlational: they reward retrieval-, reflection-, or verification-like steps without estimating whether the step contributes to final verified success under a specified intervention. We propose CVT-RL, a constrained policy-gradient algorithm with dense verifiable rewards, intervention-validity gating, and a policy-conditioned counterfactual contribution (PCCC) estimator. Deletion, semantic substitution, evidence substitution, and tool-output perturbation define separate controlled interventions; continuations are sampled from a frozen reference policy, and a selection-adjusted doubly robust estimator augments the advantage. Belief control uses only prefix-observable labels, while an augmented Lagrangian constrains unsupported claims, skipped verification, tool tampering, and unsafe calls. On long-context QA, ALFWorld, ScienceWorld, and web/tool tasks, CVT-RL improves average task success from 71.8% for compute-matched non-causal RL and 75.4% for an information-matched counterfactual-process baseline to 78.9%, improves evidence F1 from 78.9 to 82.8 over the information-matched baseline, and reduces measured hacking from 7.2% to 3.9%. Independent human audit estimates 4.6% hacking for CVT-RL versus 8.1% for the information-matched baseline, and adaptive detector-evasion attacks raise hacking only to 7.1%. Stratified bootstrap and mixed-effects tests give p<0.01 after Holm correction for all primary metrics. Carefully scoped counterfactual credit, paired with validity gating, diagnostics, and verifiable constraints, provides a reproducible route toward more reliable long-horizon RL for language agents.

We introduce VideoKR, the first large-scale training corpus specifically designed to strengthen knowledge- and reasoning-intensive video understanding. It comprises 315K video reasoning examples over 145K newly collected, CC-licensed, expert-domain videos. We develop a human-in-the-loop, skill-oriented example generation pipeline that targets progressively deeper video reasoning capabilities while ensuring the difficulty, diversity, and reliability of both the examples and their CoT rationales. We also curate VideoKR-Eval, a new expert-annotated benchmark where questions require genuine video understanding and knowledge-intensive reasoning rather than textual shortcuts. Our experiments show that, under a standard SFT$\rightarrow$GRPO pipeline, models post-trained on VideoKR outperform prior post-training approaches on knowledge-intensive video reasoning while remaining competitive on general video reasoning, highlighting data design as a key driver of progress in video reasoning. We further conduct comprehensive ablations to isolate the contributions of VideoKR, providing actionable insights for future work.

Foundation models are increasingly trained on sequences of user actions in recommendation, payments, fraud, and commerce, but these models still lack the kind of compute calibration that scaling laws provide for language models. We study a common two-part behavioral-model architecture: a feature-based event embedder maps each multi-modal item to a vector, and a decoder-only transformer predicts the next event from the resulting sequence. Across roughly 600 runs on real interaction data, spanning $10^{15}$-$10^{19}$ training FLOPs, we jointly vary four deployment-relevant axes: the two-part parameter split, critical batch size, model/data allocation, and the number of sampled negatives used after freezing the embedder. A small embedder ($s^{\star}\!\approx\!2\%$ of parameters) is compute-optimal at every budget we test because embedder parameters are both more expensive per step and exposed to far more repeated items than contextualizer parameters. Compute-optimal training is data-heavy relative to text at low compute, but its $D/N$ ratio moves toward the Chinchilla heuristic as compute increases. The sampled training objective and deployed ranking metrics disagree in ways that themselves scale: critical batch size, optimal negative count after freezing, and the agreement between loss and ranking quality all shift with compute and with the chosen evaluation metric. For negative sampling, larger budgets increasingly prefer more negatives; by $10^{19}$ FLOPs the active constraint is candidate-axis memory rather than FLOPs. In behavioral foundation models, the evaluation metric is therefore part of the scaling law: changing it can change the compute-optimal recipe.

World-action models (WAMs) jointly generate future video and robot actions through iterative diffusion, achieving strong performance on manipulation benchmarks but requiring tens of denoising steps, a cost that precludes real-time control. Step distillation has emerged as the natural remedy, but off-the-shelf methods break down in the joint video-action setting because video and action streams use different SNR-shifted noise schedules and reach training with substantially different marginal noise distributions, an asymmetry that single-modality distillation methods cannot accommodate. We introduce \textbf{Flash-WAM}, a modality-aware step-distillation framework inspired by consistency distillation that selects the consistency function for each modality to match its noise regime: a linear-gradient-scaling parametrization for the action stream's low-noise regime, paired with a variance-preserving parametrization for the video stream's high-noise regime, grounded in a structural analysis of the consistency-function family that characterizes the achievable gradient scaling under the consistency boundary condition. Instantiated on LingBot-VA, Flash-WAM compresses inference to a single step in each modality. On RoboTwin 2.0, this reduces per-chunk latency from $8.1$ seconds to $348$ ms on NVIDIA L40S, a $23{\times}$ speedup that enables real-time inference. Flash-WAM preserves task success on simulation benchmarks ($85.5\%$ RoboTwin 2.0, $95.7\%$ LIBERO) and substantially recovers real-world performance ($60\%$ average on a Unitree G1 humanoid robot), while naive consistency distillation drops to $24\%$ at the same step budget.

Large language models (LLMs) have shown increasing promise in generating functionally correct register-transfer-level (RTL) hardware designs. Recent systems improve further through EDA-integrated reinforcement learning with syntax, simulation, and PPA rewards, but train a general RTL generator before deployment while test-time approaches search with a frozen policy. We instead perform reinforcement learning at test time, allowing the LLM policy to adapt to executable EDA feedback for the specific RTL problem at hand. We propose TTT-RTL, to our knowledge the first per-design test-time training framework that closes the loop between an LLM policy and an EDA pipeline for RTL optimization. TTT-RTL samples candidate implementations, verifies them through syntax checking and simulation, scores valid designs using synthesis-derived PPA product, reuses high-reward variants through a PUCT-indexed design-state pool, and updates the policy with an entropic policy-gradient objective. To stabilize policy updates under sparse or plateaued rewards, we introduce an adaptive KL-budget controller that adjusts the entropy constraint using reference KL, effective sample size, and reward saturation signals. On RTLLM v2.0 under Nangate 45nm, TTT-RTL reduces the geometric-mean PPA product by 65.1% over the reference, outperforming the strongest published frozen-policy agent baseline at 26.1%. On an industrial XuanTie C910 FPU leading-zero-anticipation unit under Sky130, TTT-RTL achieves a 59.4% ADP reduction, and ablations confirm that policy adaptation, state reuse, and KL-budget control each contribute. These results suggest that test-time training with executable EDA feedback can move LLM-based RTL generation beyond functional correctness toward physically optimized hardware.

Inverting a robotic task requires more than reversing symbolic state transitions or rewinding motor trajectories. In robot manipulation tasks, symbolic inverse plans often fail to fully restore the effects of forward executions under continuous interaction dynamics. We present a hybrid framework for inverse manipulation that derives inverse-skill objectives from STRIPS-like operators automatically extracted from demonstrations through soft geometric predicates. For each extracted operator, we construct an inverse restoration objective that preserves preconditions, restores delete effects, and negates add effects. A task planner first attempts to satisfy this objective using available action primitives. Unresolved symbolic predicates then induce a residual operator learning problem solved through Reinforcement Learning (RL). We evaluate the framework on the ManiSkill3 PushCube task. For a forward pushing skill, the symbolic inverse performs a coarse pick-and-place restoration, while a residual Soft Actor-Critic policy refines the cube pose to satisfy the remaining inverse predicates. Our results show that predicate-derived residual control can turn an approximate symbolic inverse into a physically grounded inverse skill.

Enforcing nonlinear inequality constraints in neural networks remains challenging, especially when the output is subject to many coupled constraints. Existing hard constraint methods often impose structural restrictions on the constraint set or introduce substantial computational overhead for large-scale nonlinear problems. Here, we propose DiffSlack, a differentiable projection layer for nonlinear inequality-constrained neural prediction. DiffSlack reformulates inequalities as equalities with learnable slack variables, which are predicted as part of the augmented network output and provide a data-driven warm start for damped Gauss-Newton projection. The projection layer maps raw predictions onto the augmented feasible manifold while preserving end-to-end differentiability. A two-stage curriculum further stabilizes training and improves constraint satisfaction. We evaluate DiffSlack on vehicle path planning with 200 nonlinear inequality constraints from collision avoidance, curvature limits, and waypoint spacing. Compared with existing learning-based baselines, DiffSlack achieves a higher planning success rate and stronger geometric constraint satisfaction under a comparable inference budget. Ablation studies further show that the hard projection layer reduces sensitivity to supervision quality. Closed-loop tracking in CARLA and real-world vehicle experiments confirms the executability of the generated trajectories. These results demonstrate that DiffSlack provides a practical and scalable approach to embedding hard inequality constraints into neural networks for engineering applications.

Robotic arms capable of traversing arbitrary spatial paths, especially in highly obstructed workspaces, are highly desired across several industries. Quaternion-joints have recently empowered a specific class of robotic arms -- cable-driven redundant manipulators -- beyond its prior capabilities. Specifically, quaternion-joints reduce the number of required motors per degree of freedom, paving the way for more compact solutions.An ongoing challenge is that the complexity of the kinematic model of quaternion joints challenges a priori decisions on manipulator configurations and imposes higher computational demands on the control system and its non-linearities amplify all discrepancies between design and physical artifact arising from fabrication imprecision. Here we show a that a 4-segment, 8-joint manipulator can achieve a broader workspace than extant configurations, at lower hardware cost, and that Residual Reinforcement Learning outperforms extant state-of-the-art methods -- specifically, the FABRIK algorithm -- on the control of such manipulator. Our results show that this configuration is more workspace-effective than prior designs, and that Residual Reinforcement Learning outperforms FABRIK by three orders of magnitude on positional and orientational accuracy, effecting precise control of the novel 4-segment, 8-joint manipulator. Additionally, the control implementation is simpler: we describe the complete FABRIK process for control and corresponding learning implementation. Our methodology is applicable to the design of new systems, providing designers with further tools for the development of this class of manipulators and corresponding control systems for novel configurations.

Wearable exoskeleton systems hold promise for restoring mobility in individuals with physical impairments, yet most existing controllers rely on static gait policies that lack the ability to adapt to dynamic real-world environments or individual user characteristics. We present \olive (\underline{O}nline \underline{L}ow-rank \underline{I}ncremental Learning for Efficient Adapti\underline{ve} Exoskeletons), a parameter-efficient online adaptation framework that continuously personalizes exoskeleton control during deployment. \olive decomposes the adaptive component of the control policy into a low-rank residual form~$\dW = \At\Bt^\top$ with rank~$r!\ll!\min(d,k)$, reducing online update cost from $\mathcal{O}(dk)$ to $\mathcal{O}(r(d{+}k))$ while preserving the stability of a pretrained base controller~$\Wz$. Parameters are updated via a reward-shaped policy gradient driven purely by on-body sensor feedback (EMG, IMU, vibration), eliminating dependence on offline reference trajectories. A gating mechanism modulates the strength of personalization based on contextual state, and a dynamic rank scheduler adapts the update dimensionality to terrain complexity -- allocating minimal capacity on simple flat terrain and expanding to higher-rank updates on demanding uneven surfaces -- enabling robust performance across diverse activities: flat walking, stair navigation, slopes, and uneven terrain. Experiments on the wearable platform demonstrate that \olive achieves +13, +22, and +15 percentage-point improvements in gait smoothness, effort reduction, and motion stability over the strongest baseline, converging within $\sim$1{,}800 walking steps at 7.4,ms end-to-end latency. Our code implementation is available at https://github.com/FastLM/OLIVE.

Efficient multimodal foundation models often rely on manually designed token-reduction operators, such as pruning, merging, pooling, and adaptive reweighting. Although these operators appear different, we show that they can be interpreted as distinct regimes of a shared operator space. Based on this view, we introduce Efficient Operator Search, a differentiable framework that jointly searches where to reduce tokens, how many tokens to retain, and how reduced token information should be processed. The proposed search space parameterizes layer activation, retention budget, and operator behavior, while the search policy optimizes task performance under one-sided budget and cost constraints. This formulation recovers representative hand-designed baselines as special cases and further discovers hybrid operators beyond isolated manual designs. Experiments on multimodal benchmarks show that the searched operators achieve competitive accuracy-efficiency trade-offs, especially under aggressive visual-token reduction. These results suggest that efficient multimodal inference can be reframed from manual operator design to differentiable operator search.

Recent analyses of multi-pathway Deep Linear Networks use Gradient Flow to predict a "winner-takes-all" specialization in which path symmetry breaks and each feature concentrates in a single pathway. In this work, we show that discrete Gradient Descent (GD) with a large step size tells a different story. We prove that single-path solutions are sharp minima, whereas distributing signals across pathways reduces sharpness by a factor that decreases with both the number of pathways and depth. Consequently, while early training reproduces the depth-driven symmetry breaking predicted by GF, oscillations at the Edge of Stability subsequently override this tendency and drive the network into a re-balancing phase, where signals redistribute across pathways. Together, these results clarify how depth shapes pathway competition and explain why large-step GD favors shared representations rather than persistent single-pathway dominance.

Reasoning language models do not distinguish tokens used for computation from tokens that constitute persistent state: once generated, all hidden thoughts remain in context and influence future predictions. As a result, downstream reasoning may depend on failed attempts, dead ends, and private scratch work that should not be safely relied on later. We recast this phenomenon as a new training objective, state commitment learning: training models to explicitly distinguish information that should be committed as persistent state from temporary computation that can be discarded. We define a counterfactual criterion, persistent-state sufficiency, which makes it trainable and measurable whether an answer remains usable after hidden thoughts are erased. We then propose Counterfactual Erasure RL (CERL), which evaluates, under the same prefix, both a path that keeps hidden thoughts and a path that erases them, and gives reward only when the erasure path remains correct. We also introduce the Erasure Dependence Protocol and show across mathematics, long-chain logic, scientific QA, and multi-turn tool-use evaluation that CERL substantially reduces answer dependence on hidden thoughts without sacrificing accuracy, consistently outperforming correctness-only RL and long-answer SFT baselines.

Large Language Models (LLMs) are increasingly being deployed to make decisions that require trading off near-term gains against long-term consequences, yet little is known about how they internally represent or resolve these tradeoffs. In this work, we causally localize an underlying subgraph for temporal preference in a distilled LLM (Qwen3-4B-Instruct-2507), identifying mid-to-upper-layer nodes through converging evidence from gradient-based attribution and activation patching. We find that the geometry of time horizon is encoded in the residual stream at the expected localized layers. A behavioral analysis reveals that unintervened LLMs discount the future several times less steeply than humans, yet this preference is unstable across contexts, motivating explicit control rather than implicit reliance on training. Finally, we find suggestive evidence that steering vectors can shift temporal preference. Our work demonstrates how mechanistic interpretability can bring us closer to reliable control over how LLMs plan and reason

Budget-constrained micro-pretraining often requires triaging many candidate recipes on a shared accelerator before larger search budgets are spent. We study whether a staged fractional-factorial workflow can recover stable early effect structure in this setting. On a fixed autoresearch-derived single-GPU training loop, we run 613 experiments across pilot and follow-up screens at 2, 5, and 10 minutes; full 16-condition seeded reruns at 5 and 10 minutes; targeted seeded anchor checks; same-host greedy and matched-cost random baselines; a 60-minute bridge package; and bounded Windows A100 and Linux L40S anchor continuations through 24 hours. Main penalties from total batch, depth, and width are largest at short budgets and relax as budget increases. Within the predeclared seeded full-screen families, D, A, B, and C retain non-zero estimates at 5 and 10 minutes after within-budget Benjamini-Hochberg correction, while E does not. Random search can reach strong incumbents in this 32-condition space, but repeatedly in the same low-penalty region and without factor attribution. The 60-minute bridge anchor has the lowest mean, although that package does not separate workflow refinement from the larger bridge model's capacity advantage. In bounded 12-hour and 24-hour three-anchor continuations on both hosts, the bridge has the lowest sample mean while the non-bridge ordering stays host-sensitive. We therefore present a bounded methods result: use short designed screens to identify high-penalty directions, confirm promising anchors under repeated runs, and refine locally inside the reduced space. The evidence supports a bridge-centered recommendation through 24 hours on two hosts, not hardware-invariant ranking or general hyperparameter-optimization superiority.

Large language models are increasingly deployed as high-stakes advisors, yet standard alignment benchmarks treat sycophancy as a binary failure mode. We introduce the Granularity Gap: coarse binary metrics mask substantial social-compliance behaviors where models capitulate to user framing, validate questionable premises, or soften factual corrections without producing overtly false outputs. We evaluate six Gemini variants across generations 2.0, 2.5, and 3.0 on 73 adversarial prompts under three guardrail conditions (Control, Simple, Protocol), yielding 8,830 graded responses. Using a 0-4 Likert scale validated against a human annotator triad (Fleiss kappa = 0.71; Cohen kappa = 0.78 vs AI consensus; 95.9 percent binary accuracy, 100 percent specificity), we quantify sycophancy as continuous rather than binary. Three findings emerge. First, 27.2 percent of responses contain substantial sycophantic content (Likert >= 2.0) and 22.7 percent reach moderate or severe levels (>= 3.0), while binary win-rate framing reports only modest failure rates; coarse metrics explain just 29 percent of graded variance. Second, generational progress is non-monotonic: Gen 2.5 regresses sharply (mean Control 2.64) relative to Gen 2.0 (1.90) and Gen 3.0 (2.01), and Gen 2.5 shows inverse scaling (Pro 1.94 worse than Flash 1.71) while Gen 3.0 restores standard scaling. Third, we document an Alignment Tax: Spearman rho = -0.63 between sycophancy and truthfulness, indicating social compliance trades against factual accuracy. Egotistical Validation prompts act as a sycophancy trap (mean 3.27), nearly double Unethical Proposals (1.72). Simple guardrails outperform elaborate Protocol scaffolding on flagship models, but distilled Gen 3.0 Flash inverts this, suggesting small models may structurally require chain-of-thought scaffolding. We release the dataset and rubric to support continuous sycophancy measurement.

Large language models discard critical details when conversation history is compacted to fit within finite context windows. We present LANTERN (Layered Archival aNd Temporal Episodic Retrieval Network), a lightweight memory layer that proactively archives every conversation turn and restores relevant details after compaction via hybrid retrieval -- requiring zero LLM calls and adding fewer than 25ms of latency per turn. On 94 real multi-turn conversations (1,894 ground-truth facts, human-validated at kappa=0.81), LANTERN-Rerank recovers 78.3% of verifiable facts lost to compaction, significantly outperforming a faithful reimplementation of MemGPT's LLM-driven extraction and multi-query search pipeline (72.4%; Wilcoxon p<0.0001, 95% CI [+3.1, +8.6] pp, d=0.43) at a fraction of the inference cost. Even without the reranker, base LANTERN matches or exceeds this LLM-driven baseline (p=0.005) using zero LLM calls. When four production LLMs answer fact-bearing questions using LANTERN-restored context, accuracy improves by 8.4 percentage points on average (Wilcoxon p<0.05 for each model individually), demonstrating that the recovered context is useful across diverse model architectures. We release the full evaluation framework -- paired significance tests, failure analysis, fact-type stratification, and compaction robustness analysis -- to support reproducibility and future work.

Natural Language Inference (NLI) is a fundamental task in natural language understanding that requires determining the logical relationship between a premise and a hypothesis. Despite the remarkable success of transformer-based pre-trained models, most existing approaches primarily rely on the final-layer token representations, which are often insufficient for capturing the complex and hierarchical semantic interactions required for effective reasoning. In particular, fine-grained lexical cues, phrasal compositions, and higher-level contextual semantics are typically entangled or diluted in a single representation space. To address these limitations, we propose a novel \emph{Multi-Granularity Reasoning Network} (MGRN) that explicitly leverages hierarchical semantic features within an interactive reasoning space. The proposed framework mimics the human cognitive process of language understanding, which naturally progresses from shallow lexical matching to deeper semantic abstraction and logical reasoning. By integrating semantic information across multiple granularities in a progressive and structured manner, MGRN is able to uncover intricate semantic relationships underlying natural language expressions. Extensive experiments on multiple public benchmarks demonstrate that MGRN consistently outperforms strong baseline models, validating the effectiveness and robustness of the proposed approach.

Automated scoring models are increasingly used to assign rubric-based quality ratings to complex language performances, including classroom transcripts, yet they typically provide little insight into why a particular score is produced. We propose a general framework for sentence-level interpretability of rubric-based scoring that combines model-agnostic Shapley-value attributions with rationales generated by large language models (LLMs). Instantiated on the Quality of Feedback dimension of the CLASS framework using the NCTE corpus, the framework enables systematic comparison of fine-tuned pretrained language models (PLMs) and prompted LLMs on both scoring performance and explanation faithfulness. Across 6k annotated transcript segments, fine-tuned PLMs outperform LLMs in prediction accuracy but exhibit label compression toward mid-scale scores. Deletion-based tests show that SHAP identifies sentences that reliably drive model predictions, producing typically larger and more coherent prediction shifts than LLM-generated rationales. Cross-model analyses further reveal that SHAP attributions transfer robustly across architectures, whereas LLM rationales exert limited and inconsistent influence. Overall, the findings demonstrate that SHAP provides more faithful and transferable explanations for rubric-based scoring, and that the proposed framework offers a principled basis for evaluating both scoring models and their explanations in high-stakes educational settings and other rubric-based language assessment tasks.