Inference-Time

Compute governance today is a governance of training: the thresholds, reporting requirements, and frontier-AI regimes now in force attach to training compute and treat the trained model as the regulatory unit. That picture is incomplete: capability increasingly migrates to the deployment stage through inference-time scaling, agentic scaffolding, and compression onto consumer hardware. This paper asks which mechanisms are available once the regulatory object shifts from the training run to the inference call. We develop a feasibility taxonomy of twenty inference-time mechanisms across monitoring, verification, and enforcement, each rated on a four-point readiness scale against a documented four-vendor evidence base. We then stress the taxonomy against a two-dimensional adversary model (three capability tiers crossed with four adversary roles) and map each mechanism to four governance scenarios (domestic regulation, bilateral or multilateral coordination, industry self-regulation, and compute-marketplace governance). Fifteen of the twenty mechanisms have commercial technical substrates in production today, although governance-grade assurance and adversarial robustness vary substantially. The adversary analysis shows that this readiness holds only against a cooperative deployer and a low-to-medium-capability user: no mechanism rates adequate against a high-capability state-level deployer, and fine-tuning removes the model-internal components of the enforcement cluster, although platform-external controls can persist. A substitution analysis connects the taxonomy to a companion hardware paper as a conditional substitution principle describing when inference-stage and hardware-stage mechanisms provide comparable regulatory coverage under stated conditions. A second-rater reliability check on a random subset of the readiness ratings returned a quadratic-weighted Cohen's kappa of 0.74.

Generalist robot policies carry broad manipulation priors from large-scale data, but specializing them to a new task remains the deployment bottleneck. This requires eliciting task-specific behavior from limited demonstrations without degrading their broad capabilities. We introduce Proxy Policy Steering (PPS), an inference-time adaptation method that resolves this challenge by training two lightweight proxy policies whose calibrated velocity-space difference steers the frozen base sampler. A reference proxy models the frozen base's behavior on target-task observations, and a task proxy, initialized from the reference, captures how this behavior changes under task supervision. Their difference forms a calibrated velocity-space residual that steers the frozen base sampler at every denoising step. We identify the conditions under which this residual isolates the change induced by task supervision, and validate them empirically. Because the base is never directly modified, its broad capabilities remain available at inference, including behaviors such as recovery from failure that the demonstrations themselves do not exercise. Adaptation requires only forward velocity predictions from the base, making PPS lightweight to train and applicable even without access to the base's parameters. On 8 real-world and 4 simulation manipulation tasks, PPS lifts the state-of-the-art pi 0.5 base policy by 53% absolute success rate on average, with zero-to-one gains on tasks the base never solves, while preserving the base's broad capabilities. PPS outperforms LoRA fine-tuning, from-scratch specialists, residual policies, and prior inference-time steering methods.

Large Language Models (LLMs) are increasingly being investigated for physiological time-series prediction, yet their effectiveness may depend not only on the model itself, but also on how physiological information is represented and presented at inference time. This study investigates prompt-based general-purpose LLMs for postprandial hyperglycemia and hypoglycemia prediction in individuals with type 1 diabetes. Using the OhioT1DM dataset, we evaluate multiple open-weight LLMs under zero-shot and few-shot inference across prediction horizons of 30, 60, and 90 minutes. The analysis varies both the textual representation of the available physiological information and the amount of information exposed to the model, ranging from glucose observations alone to derived descriptors and additional contextual variables related to insulin, meals, carbohydrates, and physical activity. Performance is compared with conventional patient-specific supervised models and with Gluco-LLM, a language-model-based architecture explicitly adapted to glucose time-series forecasting. Results show a marked task-dependent behavior. Conventional supervised models achieve the strongest performance for hyperglycemia prediction, whereas the best observed prompt-based LLM configurations improve performance for hypoglycemia across all investigated horizons. The effectiveness of prompt-based inference is also strongly influenced by how physiological information is represented, while providing additional contextual information does not lead to a systematic improvement. Overall, these findings highlight physiological information representation as a central design factor in prompt-based LLM approaches to glycemic-event prediction.

Quantization is widely used to reduce the computational and memory demands of neural-network inference. In recurrent networks, however, the quantized state is stored and returned at the next time step, so the rule used to store that state can alter subsequent computations. Here, we introduce recurrent-state write-back to denote this rule and isolate its effect in a compact GRU encoder--decoder for fluorescence lifetime imaging, a molecular imaging modality used in quantitative biological imaging. A central task is estimating two lifetime parameters, the short-lived component {\tau}1 and the long-lived component {\tau}2, from high-noise time-resolved fluorescence signals. Holding the trained model fixed, replacing continuous state propagation with deterministic 4-bit state storage increases estimation errors for {\tau}1 and {\tau}2 by approximately 70x and 300x, respectively. Failure occurs when repeated small updates remain below the write threshold, leaving the stored state nearly fixed while the network continues to propose change. Error feedback, residual memory, and direction memory carry information from these suppressed updates across time and recover accuracy without retraining. Precision sweeps show that increasing state precision can worsen a fixed recurrent solution, while matched training shows that compatibility with the state interface can be learned. To test whether this behavior extends beyond the GRU, we repeat the post-training intervention in an independently trained LSTM, where coarse write-back reproduces the failure, error feedback restores accuracy, and state-specific interventions reveal greater sensitivity of the cell state than the hidden state. Our results establish recurrent-state write-back as a key determinant of low-precision recurrent dynamics and identify the state-storage interface as a central design consideration for quantized recurrent inference.

We present Iris-mini and Iris-pro, two search agents trained at the 35B-A3B and 397B-A17B scales, together with the data pipeline and training recipe behind them. Tasks are reverse-constructed from the hyperlink structure of a web corpus: we author multi-hop chains over an entity graph distilled from a seed page and its out-links, rewrite every non-answer entity into a descriptive reference so that no clue can be resolved by string matching, and admit only questions that a reference model fails closed-book yet solves once the supporting evidence is supplied. These questions are then turned into trajectories, which are filtered at both the trajectory and the turn level before SFT. The policy is then optimized by RL against live search, with the reward judge and the observation summarizer served inside the training cluster, and with over-long rollouts interrupted at the request level and resumed from their committed prefix at the next step. We alternate the two stages in a procedure we call SFT-RL climbing, returning the hardest solved and most efficient rollouts of each RL round to the next supervised pass. Because inference-time context management is worth more on these benchmarks than most reported differences between systems, we evaluate every benchmark both with and without it, holding the tool set, the context limit, and the judge fixed. All results come from a single ReAct agent, with no sub-agents and no test-time verification. With management enabled, on BrowseComp, BrowseComp-ZH, DeepSearchQA, and HLE the two models reach $82.2/84.8/86.9/52.3$ and $88.6/85.1/92.9/56.4$, the strongest overall results among open-source search agents in their respective parameter ranges. We plan to release the model weights together with the complete recipe for data construction, training, and evaluation.

Training-free, camera-controlled novel view synthesis from a single image using pre-trained video diffusion models often becomes unstable under large camera motion and long generation horizons. Existing approaches commonly combine several inference-time components, making it unclear which design choices are most important for stability. We show that the main source of stability is simple. Decomposing camera motion into small autoregressive steps limits per-step geometric distortion and reduces error accumulation. A controlled camera-step study shows that performance remains stable for small motions and degrades more strongly as the per-step motion approaches $18$-$20^\circ$. We further evaluate geometry-constrained spatial attention and low-frequency appearance anchoring as supporting refinements, together with an efficient registration-free warping pipeline. Across RealEstate10K and MegaScene, CamTrol++ improves temporal and geometric consistency, downstream 3D reconstruction quality, and generation efficiency over training-free baselines. The method remains effective for 56-frame generation and under substantial controlled depth corruption. These results show that careful control of camera motion at inference time can substantially improve the stability of camera-controlled novel view synthesis without retraining or modifying the diffusion backbone.

Diffusion models and recursive reasoners are both iterative, but they carry information across iterations differently. We add a persistent hidden state to a diffusion denoiser and remove its timestep conditioning, leaving a single shared update that can be run to arbitrary depth. The result is an anytime solver: accuracy keeps improving with inference depth far beyond the rollout lengths and backpropagation window used in training, reaching 99.90% exact solve on Sudoku-Extreme. We also obtain 98.93% solve rate on Maze-Unique. Surprisingly, progressive denoising is unnecessary at inference: holding corruption at its maximum by replacing every non-clue variable with fresh Gaussian noise at each step retains near-perfect solving and converges to stable solutions. This simple noise-injection mechanism enables a single trajectory to efficiently explore the solution space and settle on the correct answer without parallel rollouts, candidate selection, or external verifiers required by prior reasoning models. Nonetheless, ordered annealed corruption remains critical during training, which suggests that diffusion's primary contribution to our anytime solver is not a sampling procedure at inference, but a denoising training curriculum.

Although pretrained joint audio-visual diffusion models offer rich control over \emph{what} to generate, they provide no explicit control over \emph{when} an utterance should occur. To address this, we study \emph{inference-time speech scheduling}, a novel task that places coupled speech and visual articulation within user-specified begin--end intervals without finetuning the backbone model. We uncover two intrinsic properties of the denoising process that enable this task. First, a timing-sensitive text-to-audio cross-attention head exposes each utterance's model-implied source span along the latent timeline. Second, the predicted clean latent already organizes coupled speech and visual articulation, allowing their temporal placement to be edited without regenerating the content. Building on these discoveries, we propose \textbf{TimeSteer}, a training-free framework that localizes each utterance's source span through \textbf{Source Span Localization} and transfers the associated audio-visual latent content from the source interval to the specified target interval through \textbf{Region-Aware Latent Remapping}. We further introduce \textbf{SpeechShift}, the first benchmark for interval-level speech scheduling in joint audio-visual generation. Experiments across two representative backbones show that TimeSteer substantially improves interval controllability over training-free baselines while maintaining competitive overall generation quality.

The ability of AI systems to improve their behavior during deployment is becoming increasingly important. As inference moves beyond the static execution of a fixed trained model, a growing body of work studies how models can refine their behavior on the fly by exploiting test-time information and additional computation. These developments have largely evolved along two directions: methods that modify the model's state using test-time signals, and methods that improve predictions through extra inference-time resources such as more sampling and tool use. However, these directions are often studied in separate communities with different terminology, making their connections harder to see. In this survey, we present feedback-driven Test-Time Intelligence (TTI) as a unified perspective for understanding such deployment-time improvement. We use this view to relate test-time adaptation, test-time learning, and test-time scaling, highlighting both their distinctions and their growing overlap in hybrid systems. This unified framework helps connect previously fragmented ideas and provides a clearer conceptual foundation for studying inference-time self-improvement. We review major methodological paradigms, representative applications, and open challenges across vision, language, multimodal learning, generative models, robotics, and healthcare. Our goal is to provide a coherent foundation and research roadmap for the study of self-improving AI systems at test time.

Automated mining of formal specifications is vital for verifying real-time systems. However, existing passive learning approaches remain restricted to deterministic specifications or limited fragments of Timed Regular Expressions (TRE). To our knowledge, this paper presents the first framework to tackle \emph{precise} passive learning for an expressive timed logic, \emph{Metric Interval Temporal Logic} (MITL) without relying on predefined templates or restricted logic fragments. Our approach formally reduces the timed learning problem into a scalable untimed one. By identifying quantitative timing differences between positive and negative traces, we synthesise precise timed constraints and inject them as new Boolean atomic propositions. This embeds timing into the alphabet, delegating the complex formula evaluation to highly optimised, off-the-shelf untimed LTL tools. Crucially, our framework is complete, guaranteeing a separating specification can always be found. We evaluate our implementation across several benchmarks, demonstrating the effectiveness of our approach.

Inference-time search with large language models (LLMs) often concentrates on a small set of structurally or semantically similar trajectories, leaving alternatives underexplored---a failure mode we call \textit{reasoning basin collapse}. We introduce BASIN, a training-free, structure-aware selection method that groups reasoning states into basins and penalizes repeated visits to the same strategy, thereby reallocating search across genuinely distinct reasoning paths under a fixed compute budget. Under matched inference budgets, BASIN improves over Tree of Thoughts (ToT) by up to $+22$pp on Game of 24 and $+6.7$pp on MuSR. A quality-aware variant, QA-BASIN, further improves robustness by preserving high-quality basins when unconditional diversification over-explores. To explain when basin-aware selection helps, we introduce the redundancy gap $Δ$, which measures how differently search concentrates for correct versus incorrect predictions: standard ToT often operates near $Δ\approx 0$, while BASIN consistently shifts $Δ$ positive. More broadly, BASIN suggests structure-aware selection as a simple and general approach to improving inference-time reasoning. Code can be found at https://github.com/GitHubLuCheng/basin.

Local learning methods offer an alternative to end-to-end backpropagation, but their unstructured local objectives can produce weak feature learning in deep networks. We study whether the order of hidden-state optimization can address this limitation. We propose a boundary-first inference schedule that partitions a model into chunks, first coordinates hidden states at chunk boundaries, and then refines representations within each chunk. We instantiate this schedule in predictive coding networks (PCNs), a local-learning framework in which hidden activities and prediction errors are explicitly exposed during inference. On CIFAR-10, the resulting boundary-first predictive-coding instantiation improves accuracy over standard predictive coding by $9.77\%$ under a standard parametrization and by $5.51\%$ under a $μ$-parametrization. Diagnostic analyses further show more non-trivial early-layer updates, lower initial-to-final CKA, and more diverse layerwise gradients, consistent with stronger feature learning. These results support boundary-first, chunk-based inference as a practical design principle for predictive-coding training and motivate its study in broader local-learning systems.

A prominent paradigm in inference-time alignment employs lightweight supervisors to steer Large Language Models (LLMs). Through empirical analysis, we identify a structural mismatch in this paradigm: weak supervisors exhibit pervasive high entropy across the vast majority of tokens, yet prevailing dense intervention approaches mandate supervision at every decoding step. This leads to frequent low-confidence interventions that can disrupt valid base-model reasoning and incur substantial utility costs. To resolve this, we propose TUSA (Trust-based Uncertainty Sparse Alignment). Moving away from continuous oversight, TUSA reframes alignment as a dynamic arbitration process, introducing an uncertainty-aware arbiter that authorizes intervention only when two conditions are met: the supervisor is confident and the token is semantically salient. This mechanism effectively filters out uncertainty-driven noise and redundant supervision. Extensive experiments across multiple models and benchmarks show that TUSA consistently improves both safety alignment and general helpfulness. By bypassing approximately 50% of alignment steps, it not only enhances safety preference by up to 15.6%, but also boosts general preference rates by up to 12.0% compared to the dense baseline, demonstrating that selective, high-precision alignment can outperform continuous supervision.

As pretraining scaling laws approach saturation, Test-Time Scaling (TTS) has emerged as an important direction for improving reasoning by allocating inference-time compute to a fixed model prior. Viewed at a high level, TTS reframes inference as search over a space of partial reasoning states. While Chain-of-Thought (CoT) exposes intermediate steps, common instantiations rely on single-trajectory decoding, limiting recovery from early errors and exploration. This survey systematizes recent progress in tree-search-based reasoning, viewing inference as instance-specific optimization rather than decoding. We trace the evolution from uninformed search to Monte Carlo Tree Search (MCTS), highlighting how sampling-based control supports principled exploration-exploitation trade-offs. To unify a fragmented literature, we introduce a Unified Design Space spanning search topology, evaluation signals, and control dynamics, and advocate a standardized compute-reporting abstraction to make compute-accuracy trade-offs explicit and comparable.

Speech enhancement models are widely deployed as frontend modules in real-time speech services, yet their vulnerability to backdoor attacks remains unexplored. Existing backdoor methods are confined to classification tasks and rely on active trigger injection, an assumption incompatible with the passive processing nature of speech enhancement models. In this paper, we propose Ouroboros, a novel backdoor attack framework that leverages the ideal clean outputs of speech enhancement models as natural triggers, enabling inference-time activation without any external trigger injection. Extensive evaluations show Ouroboros achieves near-perfect attack success rates with minimal performance degradation on diverse models and datasets. Physical-world validations confirm that naturally recorded, unaltered clean audio can reliably activate the backdoor. Moreover, Ouroboros generalizes to targeted content-tampering attacks and remains effective against common filtering and finetuning defenses.

Dense voxel-level annotation remains a major bottleneck in 3D medical image segmentation. Single-slice propagation methods such as Sli2Vol reduce this burden by propagating one annotated seed slice through a volume using label-free registration. However, axial-only propagation accumulates errors with distance from the seed, especially in surface-distance metrics, because it ignores coronal and sagittal evidence and therefore underuses the 3D information available in CT/MRI volumes. To better leverage volumetric geometry, we study how key training and inference choices affect slice-propagation models, including single-axis versus multi-axis label-free registration, single-seed versus multi-seed propagation, and orthogonal seed configurations. Instead of propagating from a single axial seed, we use three orthogonal seeds---one axial, one coronal, and one sagittal---and fuse their propagated labels with a simple label-free rule. Our results show that the training paradigm has limited impact: an axially trained network applied to off-axis seeds captures nearly all the improvement, while explicit three-axis training adds little. Instead, performance is driven by inference-time seed geometry, especially orthogonality rather than the number of annotated slices, as a budget-matched three-axial control provides no benefit and can even degrade performance. On a multi-organ CT cohort, orthogonal seeding with the axial Sli2Vol backbone improves Dice by 21.9%, Normalized Surface Dice by 25.5%, and reduces Average Hausdorff Distance by 53.5% over the single-axis baseline.

Recent work on distillation transfers the capabilities of large models to smaller ones often by updating the latter's parameters, through teacher forcing, on-policy distillation, and related training-time methods. In this paper, we ask whether such transfer can instead occur at test time. We study strong-to-weak scaffolding: whether a stronger builder model can construct inference-time harnesses that help a weaker target model solve tasks more reliably without any parameter updates. Using four representative Theory-of-Mind benchmarks, each builder model uses 5% of the data as a validation set to iteratively refine its harness over multiple rounds, after which the finalized harness is evaluated on the full test set. Empirically, this form of test-time capability transfer is highly effective, nearly doubling average target-model performance from 0.49 to 0.91. Our analysis shows that the gains come primarily from offloading unstable model reasoning into deterministic code, benchmark-specific routing, and strict answer-format enforcement, rather than from encouraging the target model to reason more extensively or sample more broadly. We further find that builder-model reasoning effort improves harness quality monotonically, platform effects are modest relative to the builder model's own capability, and weaker target models receive the largest gains. These results suggest that inference-time harness design is an important complement to conventional training-time distillation, enabling strong models to transfer cognitive structure to weaker models without retraining.

Prior white-box studies show that large language models can retain latent traces of target knowledge after unlearning, even when the knowledge is no longer expressed in their outputs. However, existing audits remain limited to one-off diagnostics: it is unclear whether these residual signals can predict future recovery under continued training or serve as reliable optimization targets. Resolving this gap is essential to determine whether internal auditing can move beyond post-hoc evaluation toward proactive risk monitoring and safer unlearning. We propose J-Access, an inference-time audit that uses the Jacobian lens to map intermediate representations into vocabulary space and measures how often target concepts remain accessible along the model's output pathway. We hypothesize that residual accessibility reflects recovery susceptibility: knowledge that remains closer to the output pathway requires less fine-tuning to restore, leading to faster recovery. We audit 398 public unlearned models spanning eight unlearning methods. We find that: (1) most unlearned models retain access above the retain-only gold level; (2) pre-attack accessibility predicts recovery speed and extent at the model level, but cannot identify which specific facts will be recovered; and (3) directly minimizing J-Access does not promote genuine deletion. Instead, the model learns to hide knowledge from the audit, producing lower audit scores but greater post-attack recovery. These findings position J-Access as a model-level diagnostic for assessing residual susceptibility in unlearned models. We argue internal audits should serve as an independent diagnostic dimension in unlearning evaluation, and should not be converted into optimization targets without validation.

Self-consistency (SC) via majority vote is a widely used way to spend inference-time compute: sample N chains of thought, return the plurality answer. On the full GPQA Diamond benchmark (198 graduate-level science questions), majority voting reduces per-problem accuracy on a majority of problems for two instruction-tuned models from different families: 56.6% of problems for Qwen2.5-7B and 65.7% for Llama-3-8B, with Qwen the primary demonstration and Llama corroborating the direction from a near-chance baseline. The effect was pre-registered on a 151-problem confirmatory split after being observed on 47 exploratory problems, and all four confirmatory hypotheses passed. A grid oracle that routes each problem to the best N across {1, 2, 4, 8, 16, 32, 64} marks a theoretical upper bound 14 accuracy points above N = 1 for Qwen and 17 for Llama, an oracle bound requiring ground truth rather than a deployable method. No verifier-free gate reaches it: neither a plurality-agreement gate nor a token-entropy gate moves accuracy more than 0.002 from fixed-budget voting at N = 64. The mechanism is direct: confidence does not track correctness on these problems. In the highest-agreement bin the plurality answer is correct about half the time for Qwen, and for Llama that bin is less accurate than its lowest-agreement bin. We pre-register and confirm these findings on small instruction-tuned models; we do not test reasoning-native models, which we flag as the central open question.

Large Reasoning Models (LRMs) improve performance by allocating additional inference-time compute to generate extended chain-of-thought reasoning. However, recent studies reveal that sequential test-time scaling often yields diminishing or even negative returns, as longer traces exhibit increased uncertainty, error compounding, and drift from the original problem. We propose ThinkRetrieve, a test-time scaling framework that augments the reasoning traces of LRMs with dynamically retrieved solved examples at each reasoning step. Given an external corpus of problems paired with step-by-step solutions, ThinkRetrieve retrieves relevant exemplars at each intermediate step and injects them directly into the thinking trace, providing the model with guidance on how to reason rather than merely what facts are relevant. Experiments across five reasoning models (1.5B--8B parameters) on GSM-8K, MATH-500, AIME 2025, and SciQ demonstrate that ThinkRetrieve consistently improves accuracy over standard test-time scaling, with relative gains of up to $60\%$ on AIME 2025.

Hospital quality improvement (QI) programs routinely face multiple candidate interventions to optimize hospital flow, but existing methods struggle to estimate and rank the causal effects of such interventions. This work focuses on one of the most standard hospital metrics, the average length of stay (LOS), and its causal estimand, the average time saved. To characterize this causal effect, qualitative approaches rely on expert judgment to map patient trajectories, making them susceptible to cognitive biases; quantitative approaches rely on data-driven models, which fail when interventions are hypothetical with no historical data or have complex causal mechanisms that require clinical reasoning rather than data alone. We propose expert-guided g-computation, or egg-computation, which combines the complementary strengths of both approaches by connecting the Gantt charts commonly used to map patient trajectories with the causal DAG literature. We introduce a causal model over Gantt charts and establish identification using a variant of g-computation that seeks expert input only for components unidentifiable from data. To make egg-computation practical, we develop an LLM-assisted pipeline that reliably scales up expert reasoning. In simulations, egg-computation outperforms conventional causal inference methods when patients have diverse causal structures and intervention mechanisms. In a study of eleven candidate QI interventions at an urban safety-net hospital, the LLM pipeline generated graphs and time-saving estimates highly concordant with those of human experts. Beyond healthcare, egg-computation is a broadly applicable framework for estimating the average time saved for candidate interventions whose causal mechanisms can be represented using Gantt charts.

Due to the diversity of real-world time series, no single forecasting model consistently dominates across all samples. Ensemble learning addresses this by combining complementary model strengths, yet existing methods rely on fixed rules or black-box models based solely on numerical inputs, failing to leverage LLM reasoning for interpretable weighting decisions. We propose REATS, which leverages LLM reasoning capabilities as an intelligent ensemble router that jointly processes textual temporal pattern descriptions and numerical features to produce interpretable, sample-adaptive ensemble weights through chain-of-thought reasoning. To enable effective LLM-based ensembling, we study its key design choices and propose: (i) a structured input pipeline that transforms raw time series into hybrid textual--numerical representations with fixed token cost, enabling rule-based chain-of-thought construction without API dependency, augmented with retrieved similar-sample priors; (ii) a diverse multi-row weight supervision scheme coupled with a token-efficient percentage-table format that reduces numerical complexity and mitigates LLM hallucinations; and (iii) a two-stage fine-tuning framework combining SFT with GRPO, where a reciprocal reward mapping transforms the continuous unbounded MSE gap into bounded signals with amplified near-oracle sensitivity, addressing the uniform sensitivity and outlier-dominated advantage compression inherent in naive reward designs for regression-based GRPO. Experiments on eight benchmarks demonstrate that REATS outperforms competitive ensemble baselines while providing natural language explanations and demonstrating strong transfer learning and out-of-domain generalization to unseen candidate models.

Foundation world models have made remarkable progress in planning, simulation, and embodied intelligence. However, optimizing a pretrained world model toward a user-specified objective remains difficult: each campaign typically rediscovers optimization strategies from scratch, and the resulting knowledge rarely transfers to the next model. Existing research agents automate the optimization loop but treat successful strategies as directly reusable recipes, without principled safeguards for when transfer is appropriate. We argue instead that retrieval is not transfer: a strategy validated on one model is at best an optimization hypothesis for another, and becomes transferable knowledge only after target-side experimental valida- tion. Guided by this principle, we propose VERDI , a continual framework for evidence-licensed world model optimization. VERDI characterizes each world model through shared inference-time probes to construct an Optimization Fin- gerprint, retrieves relevant prior experience as ranked hypotheses, and validates every candidate under a frozen target-side verifier before admitting it as reusable evidence; contradictions among nearby fingerprints further trigger probe evolution, continually refining the diagnostic representation itself. Experiments on Ctrl-World, the Cosmos family, and RoboCoin show that VERDI reduces search cost by 68%, GPU cost by 69%, and negative transfer from 0.34 to 0.06, while predicting transfer outcomes with 83% sign accuracy.

Aligning large language models (LLMs) with human privacy preferences requires capturing individuals' disclosure boundaries beyond general privacy norms. However, a gap remains in eliciting such nuanced preferences to evaluate alignment in realistic settings. We introduce CIDER, a dataset of 14,850 human annotations from 169 users, forming 1,650 contextual disclosure boundary sets across 60 interpersonal communication scenarios involving information sharing that violates privacy norms. Each boundary represents a real user's disclosure decisions over 9 sharing variants in a scenario, for a given communication role and AI-mediated condition. We formulate a task in which models predict a user's disclosure decision from historical boundaries, with varying levels of contextual information. Across 12 open and proprietary models, in-context personalization improves prediction accuracy by up to 11.41 percentage points using only 6 historical examples. Larger models such as GPT-5.4 (with medium reasoning effort) and Claude Sonnet 4.6 are better at leveraging semantic context to understand user-specific, context-dependent disclosure preferences for more accurate predictions, while smaller models tend to rely on structured heuristics based on disclosure granularity and identifiability. Personalization generally improves prediction accuracy, but the improvement is often accompanied by imbalanced shifts in false-positive and false-negative rates across models, with only Claude Sonnet 4.6 achieving balanced improvements in both. Our findings reveal both the promise and limitations of inference-time personalization for privacy preference modeling and position CIDER as a resource for advancing personalized privacy alignment.

Temporal decisions in language-model systems often depend on both symbolic task state and elapsed wall-clock time, such as cache expiration, job completion, quota resets, deadlines, or stale sessions. We study whether elapsed time can be supplied as a non-token, system-side scalar and composed with visible symbolic state by a frozen-backbone language model. We introduce ChronoState, a compositional temporal-state benchmark in which symbolic state appears in the prompt, elapsed seconds tau are supplied through a hidden chronometric-injection channel, and the model selects a forced-choice temporal action. Here, "hidden" means hidden from the user-visible token sequence, not from model computation. Using Qwen2.5-3B-Instruct as a frozen bf16 backbone with a 31-dimensional sinusoidal-plus-log time encoding, gated FiLM residual modulation, and a rank-8 LoRA action surface, hidden-time CI reaches 0.9305 +/- 0.0134 accuracy and 0.9410 +/- 0.0103 balanced accuracy. No-time and shuffled-time controls fall to 0.5511 +/- 0.0042 and 0.3323 +/- 0.0097, respectively, with high shuffled-time wrong-state consistency supporting causal dependence on the injected scalar within the trained distribution. Generalization remains strong for held-out templates, durations, and multi-constraint compositions, but held-out quota-family transfer is weak at 0.5065 +/- 0.0559, while a fair prompt+LoRA timestamp baseline reaches 0.9893 +/- 0.0052. Thus, ChronoState supports a narrow conclusion: hidden elapsed time can be composed with symbolic task state under direct supervision, but does not establish autonomous time tracking, broad unseen-family abstraction, or superiority over prompt-injected timestamps.

When videos extend from hours to days, directly processing them end-to-end becomes impractical for current Multi-modal Large Language Models (MLLMs). This ultra-long setting necessitates a two-stage paradigm: query-agnostic memory construction followed by retrieval-based inference. Prior work invests in complex memory construction to pre-model high-level relations in videos, despite not knowing the downstream query at build time. We instead prioritize high-recall retrievability during memory building, and defer query-specific, high-level relation composition to inference time. To this end, we propose MERIT(Multi-key Episodic Retrieval with Inference-time Temporal expansion), a simple yet effective agentic framework for ultra-long video understanding. First, we formulate an episodic multi-key representation that enables precise retrieval of fine-grained memories through a simple key-matching mechanism. Second, we introduce a neighbor filtering mechanism to capture broader semantic context without the massive computational overhead of global memory construction. This is achieved by expanding the temporal scope exclusively around the retrieved segments at inference time. By leveraging simple key-matching with this on-demand temporal expansion, MERIT achieves state-of-the-art performance across three long-video benchmarks: EgoLifeQA, LVBench, and Video-MME (Long).

We describe DAEP, team BIGC's submission to NLPCC 2026 Shared Task 1 Track 3: Difficulty-Aware Temporal Answer Grounding in Video Corpus (DA-TAGVC). The task requires retrieving the target video from 50 candidates and localizing the answer-supporting span. DAEP ranks videos with subtitle, visual, and procedural-context evidence, expands high-scoring anchors into temporal spans, and reranks spans for final output. Its main design is to convert the task-provided simple/complex input label into an inference-time evidence plan controlling modality weights, Top-K aggregation, boundary threshold, expansion length, and reranking strength. In the official evaluation, BIGC ranks first among ten systems with an Average score of 0.2728. Validation ablations show that visual evidence, procedural context, and difficulty-aware planning improve ranking quality, with the largest gain on complex questions.

Large language models can solve harder reasoning problems with more inference-time compute. The term "test-time scaling," however, covers several inference algorithms: extending deliberation along one trajectory, sampling completed candidates and aggregating them by voting or verification, and searching over partial states. These algorithms differ in statistical structure, compute requirements, and failure modes. Treating them as interchangeable under a scalar "budget," or reporting accuracy without specifying the inference protocol, makes results difficult to compare across studies. We study test-time scaling along three axes. First, we formalize it as budgeted inference over the implicit prefix tree of an autoregressive model and distinguish single-trajectory sequential scaling, leaf-level scaling with terminal reduction, and prefix-level scaling. Second, we treat the full inference system as the evaluated object and separate end-to-end performance from candidate-bank diagnostics. We introduce an evaluation profile whose coordinates and simple functionals recover or bound common repeated-sampling metrics, and require compute accounting and uncertainty estimates that match the protocol. Third, we distinguish exact replay from distributional reproducibility and state the requirements for each. We also organize open-weight reasoning models by model-side and interface mechanisms. Our empirical study covers broad knowledge, symbolic reasoning, and competition mathematics, and we publicly release 1,403,520 sampled model attempts. The project website is available at https://mohsenhariri.github.io/scorio/tts. The released datasets are Trace (https://huggingface.co/datasets/harimo/scorio-trace), Lite (https://huggingface.co/datasets/harimo/scorio-lite), Math (https://huggingface.co/buckets/harimo/scorio-math), and SuperGPQA (https://huggingface.co/buckets/harimo/scorio-gpqa).

Recent advances in agentic workflow optimization automate workflow design through task-specific workflow search or input-conditioned architecture selection. However, they determine the workflow before execution and cannot adapt failed workflow regions using execution-time label-free quality signals. Naively enabling such inference-time adaptation through whole-workflow re-optimization would be computationally prohibitive. To tackle this challenge, we introduce GRAFT, which preserves a globally optimized workflow while locally replacing only selected regions for each input. Without parameter training, GRAFT evaluates region-level alternatives using label-free execution-quality signals and accepts only replacements that improve local quality while preserving workflow-level consistency, thereby enabling instance-wise adaptation without whole-workflow re-optimization. GRAFT applies without modification across a range of tasks spanning mathematical reasoning, code generation, and multi-hop and knowledge-intensive question answering. Under matched optimizer and executor settings, it improves over the strongest prior workflow-optimization method, MaAS, by 3.85 points on average. Replacing only the executor with a stronger model yields further gains without re-optimizing the global workflow. This suggests that an optimized workflow is not merely a static optimization artifact, but an adaptable execution policy that can evolve with inference-time feedback and stronger executors.

Human mobility prediction models, which forecast the next location in a user's trajectory, are increasingly deployed in urban analytics, navigation, and personalized services. Yet, little is known about their potential to memorize and expose sensitive user trajectories from training data. While memorization has been extensively studied in language models, mobility prediction poses unique challenges: training sequences encode human behavior at various spatial and temporal scales, creating privacy risks at different granularities. In this paper, we conduct the first systematic audit of memorization in mobility prediction models. While prior work has shown that privacy leaks can arise from such models, we systematically assess and quantify memorization risks at scale. We identify key challenges, including the lack of a randomness space, the multi-scale structure of trajectories, and user-specific behavioral diversity. To address these challenges, we introduce a framework to quantify mobility memorization at different levels of granularity: individual locations, anchor pairs, and subtrajectory segments. We also develop user-grounded reference sets to assess how likely a model is to prefer training data over realistic alternatives. Our evaluation across multiple models and datasets reveals pervasive memorization patterns that correlate with user regularity and increase the risk of data extraction at inference time. Our findings call for mandatory privacy auditing in mobility prediction models.

Multi-agent debate commonly exchanges complete rationales even when disagreements concern only a few intermediate claims. We introduce Localized Multi-Agent Debate (LMAD), an inference-time protocol that represents agent rationales as nodes, locates their earliest conflict, and restricts debate to the corresponding local segments. Guarded resolution extends a shared committed state so that later conflicts can be addressed without reopening accepted steps. We evaluate LMAD on four multi-hop question-answering benchmarks using ten backbones from four model families. Our method achieves the highest macro-averaged judge accuracy across all ten backbones, outperforming the strongest conventional baseline by up to 7.20 percentage points.

Deep search has become a fundamental capability of large language models (LLMs) for solving open-domain complex tasks. However, existing approaches typically rely on linear sequential reasoning for both trajectory generation and inference, making it difficult to consistently preserve intermediate states and constraints throughout long-horizon multi-hop search. Consequently, they often suffer from context forgetting, search drift, and inefficient exploration. To address these limitations, we propose $\textbf{G-ReAct}$, a reasoning framework for deep search that organizes reasoning as $\textbf{state evolution over a fixed-topology query graph}$. The evolving graph state explicitly tracks search progress and guides subsequent decisions, transforming exploratory search driven by textual history into graph-guided reasoning under explicit constraints. G-ReAct supports both training and inference: it generates high-quality deep-search trajectories for supervised fine-tuning and provides structured guidance for inference-time search without additional fine-tuning. Experiments demonstrate that with only 1.9K generated trajectories for fine-tuning, Qwen3-30B-A3B-Thinking-2507 achieves $52.6\%$ accuracy on BrowseComp-ZH and $79.0\%$ on XBench, outperforming comparable open-source methods trained on substantially larger datasets, including RL-enhanced methods. Furthermore, when applied at inference time, G-ReAct consistently improves the performance of existing strong LLMs on deep-search tasks. We will publicly release all code and model weights.

Hierarchical Graph Retrieval-Augmented Generation (GraphRAG) organizes corpus knowledge at multiple levels of granularity, yet fixed context construction may fail to translate these multi-resolution representations into a context suited to the current query. We identify this mismatch as the representation--inference gap. We propose Agentic Context Engineering for Hierarchical GraphRAG (ACE-GraphRAG), an inference-time context policy layer that supplements and adapts the initial context for generation. ACE-GraphRAG formulates context construction as a policy over gap-aware refinement, retrieval branches, and task-conditioned adaptation. Parallel Differential Retrieval acquires supplementary evidence from depth-oriented factual and breadth-oriented semantic branches. These evidence increments are consolidated with the initial context while preserving provenance and abstraction levels. Full-ACE applies the full policy uniformly within each task family, whereas Adaptive-ACE selects task- and topology-specific policies for individual queries. We evaluate ACE-GraphRAG on HotpotQA, 2WikiMultiHopQA, and four UltraDomain subsets across multi-hop QA and query-focused summarization. Full-ACE outperforms the evaluated RAG and GraphRAG baselines across both task families, while Adaptive-ACE further improves multi-hop QA and is preferred over Full-ACE on all four UltraDomain subsets. Ablation and topology analyses support treating context construction as a query- and task-dependent inference policy rather than a fixed procedure.

High-resolution 3D segmentation of hip and shoulder anatomy from CT and MRI is essential for surgical planning, yet frozen segmentation models often fail under domain shift. CNN-based expert models are fully automatic but lack adaptability, whereas promptable foundation models generalize better but require manual prompting. We present MedSAM2-Anatomy, a training-free inference-time optimization framework that improves frozen segmentation models without retraining or human interaction. A frozen expert model generates anatomical priors that are automatically converted into multiple prompt hypotheses for a frozen 3D foundation model. Candidate masks are fused while anatomically implausible priors are rejected. No model weights are updated and no manual prompts are required. TotalSegmentator and MedSAM2 are used as representative expert and foundation models, allowing the contribution of the inference policy to be isolated. Evaluation on the independent Balgrist-V0 CT and MRI cohorts shows that inference-time optimization increases median Dice from 0.71 to 0.92 on hip MRI and from 0.89 to 0.92 on shoulder CT, while reducing median HD95 on hip MRI from 22.0 mm to 5.0 mm. On public TotalSegmentator benchmarks, the expert model remains strongest, indicating that the optimal fusion strategy depends on the reliability of the expert prior. These results demonstrate that training-free inference-time optimization provides a practical strategy for improving frozen segmentation models without manual prompting.

Deploying autonomous computer-use agents (CUAs) locally is increasingly important for privacy, cost efficiency, and practical usability, yet improving their performance under strict hardware constraints remains challenging. While recent studies show that inference-time scaling can improve frontier computer-use agents through additional computation during execution, its effectiveness for resource-constrained local models remains poorly understood. We present a systematic empirical study of inference-time scaling in local CUAs across contextual, temporal, structural, and parallel dimensions. We evaluate Qwen3-VL-8B/30B-A3B, UI-TARS-1.5-7B, and OpenCUA-7B on the OSWorld benchmark. Our results show that additional computation often yields diminishing returns while changing failure modes. Contextual scaling provides historical grounding that improves trajectory stability and task accuracy, but its gains saturate as token cost increases and failures shift from repetitive or stalled trajectories toward premature false successes. Temporal scaling similarly reduces max-step stalls, yet does not substantially improve task success, indicating that longer horizons often extend erroneous trajectories rather than correct them. We further find that structural decomposition can introduce planning and formatting overhead in local two-stage agents, while parallel scaling partially mitigates these failures at a substantial computational cost. Overall, our findings suggest that efficient local CUAs require selective compute allocation, failure-aware control mechanisms, and agentic frameworks designed around the capabilities and limitations of local models.

Recent advancements in inference-time scaling have significantly unlocked the complex reasoning capabilities of Large Language Models~(LLMs). However, for agents, these approaches suffer from a critical inefficiency, operating in a stateless manner and engaging in redundant search processes. Existing memory mechanisms largely rely on the reasoning capabilities of LLMs, leading to prohibitive computational costs. In this paper, we propose a novel framework, \textit{GAMER}(Graph-based Action-centric Memory with Episodic Reasoning), that bridges the gap between inference scaling and episodic memory. Our approach models historical reasoning as a dynamic \textit{Action-Centric Graph}. By decoupling the memory mechanism from LLMs, our method can save token/money usage by providing less memory context than memory mechanism baselines. To extract knowledge from the graph effectively, we use a dual-stream Temporal Difference learning mechanism to estimate the positive(suggestion) and negative~(avoidance) value of action nodes based on past successes and failures. During the inference phase, this learned value function optimizes decision-making bi-directionally, so that positive values provide action suggestions, while negative values indicate high-risk actions. By performing efficient searches on the graph, our method significantly improves the efficiency of inference scaling. Experiments on multiple benchmarks demonstrate that \textit{GAMER} achieves superior performance by \textbf{20.81%/6.17%} for success/progress rate compared to vanilla baselines.

Instruction hierarchies are a core safety assumption of language model deployment: higher priority inputs, such as system prompts, should override conflicting lower priority inputs from users or tools. Yet frontier LLMs often violate this hierarchy. We introduce V-Steer, a training-free inference time method that restores privileged influence by editing cached value vectors at prompt positions. Using direct logit attribution on the first next token prediction, V-Steer identifies heads where lower priority spans dominate privileged ones, then boosts privileged spans and suppresses conflicting lower priority spans through in-place multiplicative edits to cached V tensors. Since the method acts only on cached values, it remains compatible with fused attention backends and adds only a one time prefill overhead. Across models from 7B to 70B, this attribution guided intervention raises primary constraint accuracy from under 18% up to 92% on controlled role conflict benchmarks, and on broader instruction hierarchy evaluations substantially outperforms prompt only baselines while matching or exceeding SoTA training based methods on 3 of 4 scales of LLMs, with negligible decoding-speed overhead. The code is available at https://github.com/cindy2000sh/v-steer.

LLM provenance testing asks whether a suspect LLM belongs to the same lineage as a source. Existing black-box methods largely infer this relationship from response-level characteristics, but these characteristics may shift under adaptation or deployment even when the underlying meaning remains unchanged, weakening the reliability of provenance evidence. To address this limitation, we introduce induced decision regions by mapping open-ended outputs into a finite decision space, thereby abstracting away surface-form variation and reframing provenance testing as measuring the inheritance of decision regions. Empirical analysis shows that the source's induced regions are preserved more strongly in related models than in unrelated models. Building on this signal, we propose Stemma, a practical black-box LLM fingerprinting method that operationalises stability, robustness, and specificity as complementary probe-selection principles for reliably estimating induced decision region inheritance. Across 770 source-suspect pairs drawn from 56 public checkpoints and spanning diverse model-weight transformations, Stemma achieves 0.967 AUC and 87.8% TPR at 1% FPR, substantially outperforming four representative baselines. It further achieves 0.995 AUC and 93.5% TPR at 1% FPR on 1,260 pairs covering 91 deployment instances, demonstrating robustness to diverse inference-time deployment settings.

Future event prediction carries broad social impact yet remains challenging. SOTA approaches augment LLMs with external agent frameworks whose predictive capability vanishes once the harness is removed. While recent Tool-Integrated Reasoning (TIR) internalizes deep search for multi-hop retrieval of facts, forecasting further demands temporal search and reasoning over historical trends and dynamic shifts. The key obstacle is data: historical queries induce temporal leakage that degrades forecasting into retrieval. Prior works either freeze information gathering with static observations, or rely on rejection sampling or unresolved fresh queries that discard vast amounts of data, degrading synthesis efficiency. We propose a time-truncation harness that enforces a temporal cut-off at every turn, enabling TIR-style sampling from historical events, reducing temporal leakage and reliance of rejection sampling or unsolved queries, increasing the sampling efficiency. We further build a large-scale corpus and a process-based metric and show that our harness naturally induces a broader temporal breadth of search and raises the proportion of high-quality data, further increasing the efficiency and reducing the reliance on complex rubrics. Distillation experiments show that students trained on harness-intervened data achieve the best performance, demonstrating harness-assisted model evolving that turns higher quality temporal search and reasoning data into a parametric advancement of the students.

Swarm Learning is a decentralized collaborative learning mechanism that allows multiple organizations to train a shared model without central coordination or direct data sharing. In typical horizontal Swarm Learning, datasets across sites are usually assumed to share the same feature set. However, in real-world applications, sites often have partially overlapping features because measurements, protocols, and available covariates differ across sites. This feature heterogeneity creates a practical issue for machine learning algorithms such as Random Forests. Specifically, when decision trees are pooled into a global Random Forest, inference at a given site can become ill-defined if a traversal encounters a split on a feature that is not available locally, often forcing organizations to discard site-specific variables upfront. In this paper, we address feature heterogeneity in Swarm Learning with Random Forests under partially overlapping feature spaces. We propose several deterministic and probabilistic inference-time strategies that resolve such missing splits without restricting training to the intersection of features. We evaluate the methods on nine datasets and demonstrate that they outperform both the intersection baseline and locally trained models across a broad range of scenarios.

Recent work shows that fine-tuning language models on even a small amount of poisoned data can install targeted misbehavior, and ostensibly benign data can transmit hidden preferences that generalize broadly. Standard defenses, such as data filtering, mixing in harmless data, and regularization, attenuate these effects but do not eliminate them. We instead pursue robustness through redundancy: collecting multiple datasets from different sources and only learning what is common between them. Thus, if only a subset of sources are malicious, the misbehavior will be blocked. In order to implement this defense strategy, we fine-tune a separate reference model on each source's dataset and aggregate their next-token distributions at decoding time. We introduce two consensus decoders: a token-wise minimum, which caps each token at the lowest probability any source assigns, and a base-relative variant, which reverts to the base probability on any token the sources move in opposing directions. We further relax exact agreement to tolerate partial support across sources and different surface expressions of the same intention. Across controlled poisoning tasks, subliminal learning, and emergent misalignment, consensus decoding suppresses source-specific misbehavior while preserving shared desirable behavior, including cases where union training and weight averaging retain the unwanted behavior.

Trustworthy deployment of AI-based diagnosis in safety-critical mechanical systems hinges on validation: whether a prediction can be checked against physical reality before it is acted upon. Current intelligent fault diagnosers fail this standard in two ways. Their standard output, a class label with a softmax confidence score, is an internal statistic of the classifier, offering nothing checkable against independent physical knowledge; and the growing use of generative language models in maintenance reporting adds a second risk: hallucinated content entering reports on which decisions rest. Taking bearing fault diagnosis as the testbed, this work addresses both problems from the output side. The proposed Diagnostic Evidence Network (DENet) is an encoder-agnostic multi-task framework extending the output to a structured evidence record: the classification, a predicted characteristic frequency comparable against the theoretical value determined by bearing geometry and shaft speed, and a temporal localization of transient impulses inspectable on the raw waveform. Across four encoders and three public datasets, this evidence incurs no statistically significant accuracy cost, with a frequency error of about 6 Hz on 1,024-point segments where spectral estimation is structurally inapplicable. Centrally, the deviation between predicted and theoretical frequency constitutes a label-free, inference-time validation signal: it detects misclassifications with AUROC values of 0.970 and 0.871, and remains discriminative in the high-confidence regime where confidence-derived detectors are blind. Finally, a QLoRA-adapted language model is constrained to translate, but never generate, diagnostic content, reducing unsupported-claim rates from 10-12% to 2% and eliminating fabricated quantities.

In this work, we explore how the inference time of a Transformer Neural Network can be efficiently optimized with applications to real-time anomaly detection in financial time series. The financial time series are price series such as asset prices. Unfortunately, the data is often with errors or outliers that make the downstream data processing tasks useless, unstable or even harmful. Moreover, the amount of financial time-series data has been significantly increasing. Hence, there is a need for better data-cleaning methods in terms of accuracy and in terms of processing speed. Transformers as a neural network architecture have achieved superior performances in many tasks such as Natural Language Processing and Computer Vision. Time series modelling and especially anomaly detection tasks can benefit from the features of transformers architecture in multiple ways, including the capacity to capture long-range dependencies and interactions. Increasingly powerful hardware, such as field-programmable gate arrays (FPGAs), have seen increasing usage in recent years due to their reconfigurability and high performance. They can be efficiently utilized to speed up the computations of the Transformer architecture. We explore different Transformer architectures for time series modelling and how they can be efficiently implemented on an FPGA board (PYNQ-Z2). In particular, we examine the application of Transformers to detect anomalies in time series and we show how they can be efficiently implemented on an FPGA board to minimize latency. The code is available at https://github.com/thxi/icl_thesis

Diffusion and flow-matching models dominate conditional image generation, yet inference-time scaling for these models is far less developed than for autoregressive language models. Because final quality is highly sensitive to the initial noise seed, many approaches spend extra compute on seed search or resampling under a black-box reward, but typically maintaining a constant memory footprint throughout inference. We show that relaxing this constraint enables an underexplored inference-time scaling axis: by front-loading exploration, evaluating many seeds early, and pruning aggressively, we can use a fixed compute budget more effectively. \emph{Progressive Seed Pruning} (\PSP) scores intermediate denoised estimates and progressively narrows the candidate set so that only promising trajectories are fully denoised, while keeping the total number of model evaluations fixed. Across diffusion and flow-matching backbones, \PSP \ consistently improves reward-guided selection and achieves higher GenEval scores (automated) and better human evaluation on prompt-alignment than best-of-$N$, importance-sampling, and tree-search baselines at matched compute. Project page: https://www.vision.caltech.edu/psp. Code: https://github.com/rogerioagjr/psp.

Faithful explanations of time-series classifiers should identify subsequences that are not only sufficient to preserve a black-box model's prediction, but also necessary for maintaining it. However, existing sufficiency-oriented methods can assign high importance to spurious subsequences that support the prediction without being essential to the model's decision. We introduce \textbf{TimePNS}, a necessity-aware framework for time-series explanation. Inspired by Pearl's counterfactual notion of necessity, TimePNS assesses whether a temporal factor is necessary by intervening on it and measuring whether the original prediction is disrupted. The framework adopts a two-stage design. Stage I learns an identifiable causal generative process together with a sufficiency-oriented explanation mask. Stage II performs counterfactual interventions on temporal factors to derive necessity signals, which supervise a temporal gate that refines the initial explanation by suppressing non-essential components and emphasizing counterfactually necessary ones. Experiments on synthetic and real-world time-series benchmarks show that TimePNS more accurately identifies decision-critical subsequences and consistently improves sufficiency-necessity trade-offs over strong baselines.

Modern AI agents rely on elaborate inference harnesses such as Claude Code, Codex, and OpenClaw to drive multi-turn reasoning, tool use, and access to external systems. While powerful, these complex harnesses also make agents hard to train end-to-end with open infrastructure, whose SFT/RL stacks cannot natively express stateful, multi-process harness inference. To address this, we present OpenForgeRL, an open-source framework for training harness-based agents end-to-end in diverse environments. OpenForgeRL achieves this with a lightweight proxy that serves the harness's model calls while recording them as training data for a standard RL codebase (e.g., veRL), and a Kubernetes orchestrator that runs each rollout in its own remote container, together enabling training on any harness in any environment at scale. By decoupling training and inference, OpenForgeRL allows researchers to easily train, study, and improve agents directly in the real harnesses and environments they are deployed with. We validate our framework across diverse, complex harnesses and environments, spanning tool/claw-based agents and multimodal GUI browser- and computer-use agents. Using only hundreds to a few thousand tasks, OpenForgeClaw reaches 31.7 pass^3 and 55.9 pass@3 on ClawEval and 33.7 on QwenClawBench. OpenForgeGUI reaches 37.7 on OSWorld-Verified, 63.0 on Online-Mind2Web, and 72.3 on WebVoyager. Both outperform open baselines of similar size on nearly all benchmarks, and in the GUI setting match or surpass models several times larger. Beyond benchmarks, we analyze how harness choice (e.g., ZeroClaw, OpenClaw, Codex) and RL shape agent behavior. We find that some harnesses are substantially harder to learn than others, and that RL improves agentic reliability, such as self-verification, tool coverage, and completing multi-step plans, though critical abilities such as error recovery remain weak.

Scaling inference-time computation has emerged as a reliable method to improve the performance of large language models on complex reasoning and programming tasks. However, standard approaches such as independent sampling and sequential multi-turn refinement operate without token-level credit assignment, resulting in computational inefficiency, since valid reasoning prefixes are frequently discarded. In this work, we introduce Test-Time Scaling via Error Localization (TTEL), an inference-time algorithm that utilizes fixed or environment feedback to perform token-level error localization. By comparing conditional probabilities under informed feedback against a null-context baseline, TTEL isolates the step at which an error occurred. The algorithm then truncates the trajectory and branches a new generation, maximally reusing the valid prefix. Extensive evaluations demonstrate that TTEL establishes strictly dominating Pareto frontiers across sequential reasoning domains, measured by pass-at-k vs. generated-token cost. With Qwen3-8B on LiveCodeBench, TTEL attains a pass@64 of 71.0% while generating approximately half as many tokens as independent sampling (360.4k vs. 735.0k). Generalizing to math benchmarks AIME-2025 and HMMT-2025, TTEL cleanly outperforms competing test-time baselines across both Qwen3-8B and Qwen3-4B-Thinking-2507.

BoN improves model outputs by sampling several candidates and selecting one with a proxy score, but it assumes that complete candidates can be evaluated reliably. Many vision-language tasks instead provide only partial verification: a finding, span, value, region, or relation may be checkable even when no dependable whole-response verifier exists. Moreover, the same claim may recur across candidates with opposing stances, allowing one observation to support part of the pool and contradict another. We introduce Best-of-Evidence (BoE), an inference-time selection framework that keeps the BoN candidate pool fixed, represents reusable claims with a signed candidate--factor graph, and allocates a limited budget to evidence actions that can change the final choice. BoE formalizes selection under partial verification and provides a practical score-based controller, with the zero-budget case recovering the underlying BoN decision. Theoretically, we show that residual evidence capacity limits any evidence-driven improvement and that shared factor queries can achieve an O(log K) versus Θ(K) query separation in a factor-code model. Common-ledger experiments on four medical VQA settings show that BoE can improve fixed-pool selection and rescue some BoN failures when evidence is reliable, contrastive, and decision-relevant, while also revealing the channel-quality and candidate-generation limits that prevent universal gains.

Quantized small reasoning models can enter repetitive or otherwise unproductive trajectories, yet standard decoding does not adapt to the trajectory as it unfolds. We study MGT-B, a fixed, weight-preserving controller that converts overlapping windows of uncertainty, repetition, and local-change features into position-conditional empirical tail probabilities. It accumulates mixture betting factors with a CUSUM-shaped reset, and, after an alarm, restores a coherent earlier token and key-value-cache state before constrained re-decoding. On MATH-500, a paired three-seed evaluation over 1,500 generations per method raises exact-normalized accuracy from 54.73% for vanilla decoding to 56.40% (+1.67 percentage points; problem-clustered bootstrap 95% CI [+0.47, +2.80]), while a prospectively profiled random-intervention control reaches 54.60%. The gain is positive in all three seeds and costs 5.14% more sampled tokens. Seed-0 ablations show that rollback alone does not explain the result and that an isolated repetition penalty is harmful. Five-sample self-consistency reaches 70.0% but uses about 4.84x as many tokens as MGT-B. On the harder, non-overlapping Omni-MATH evaluation, however, MGT-B obtains 16.60% versus 16.67% for vanilla (-0.07 points; clustered 95% CI [-0.33, +0.20]) with 2.10% more sampled tokens. Thus, MGT-B provides a modest, reproducible local improvement on MATH-500 in the studied configuration, but the effect does not transfer to Omni-MATH and should not be interpreted as a general improvement in mathematical reasoning.

Large Reasoning Models (LRMs) often suffer from overthinking due to redundant verification steps. Existing approaches for mitigating overthinking, such as fast-slow thinking switching and reasoning trajectory compression, fail to make a fine-grained distinction between beneficial and redundant steps within the LRM's reasoning process, and may thus impair reasoning capability in their pursuit of efficiency. To simultaneously improve reasoning efficiency and capability, we propose EvoThink, a framework that reduces redundant verification and encourages the exploration of new reasoning paths. EvoThink comprises two key components: Self-Pruning Training (SPT), an unsupervised method that iteratively prunes redundant reasoning steps and self-trains on the concise trajectories; and Aha-Moment Preference Optimization (AMPO), which, inspired by genetic algorithms, identifies valuable failed reasoning attempts, synthesizes from-wrong-to-right aha-moment data, and optimizes the model to internalize this reasoning pattern. Extensive evaluations across mathematical reasoning and code generation benchmarks demonstrate that EvoThink not only substantially reduces inference-time token usage but also improves the reasoning capability of LRMs.

Time-series foundation models can forecast across heterogeneous domains without task-specific training, but their forecasts are fixed once produced and cannot directly incorporate task-specific expert feedback. We study expert-guided forecast editing: a frozen foundation model generates candidate future trajectories, and an expensive expert evaluator scores them to guide forecast revision. Under a tight query budget, two natural strategies sit at opposite ends: best-of-$N$ purely exploits the foundation model's predictive distribution, while optimization approaches mostly explore the forecast horizon as an unstructured high-dimensional vector. Each extreme is individually sub-optimal. We introduce \textbf{DEFT}, an expert-guided forecast editing framework that balances the two by first exploiting the foundation model's predictive samples in a decomposed trend--seasonal space, then exploring around them via component-wise refinement. DEFT queries the expert only on complete trajectories, then reuses scores for the trend and seasonal components that appeared in the queried recombinations. This lets each expert query provide structured component-level feedback while keeping the foundation model frozen. We compare DEFT against direct search approaches, including best-of-$N$, cross-entropy methods, and Bayesian optimization, under matched expert-query budgets. Across two forecasting benchmarks consisting of 78 datasets, three time-series foundation models, four feedback types, and seven query budgets, DEFT consistently improves the effectiveness of expert guidance. A molecular-dynamics case study further suggests that the same principle extends to more physically grounded feedback, supporting the hypothesis that sparse test-time guidance should be spent balancing prior exploitation with structured exploration.

Although Large Language Models (LLMs) demonstrate remarkable multilingual fluency, their internal knowledge representations remain disproportionately biased toward high-resource languages. This leads to cross-lingual factual inconsistency, where they shift their empirical answer distributions based solely on the prompt language. We investigate whether these biases can be mitigated at inference time, forcing an English-prompted model to answer as if it were queried in target languages (German, Spanish, Bulgarian), and evaluate four intervention strategies: zero-shot contextual steering (persona prompting), internal representation manipulation via Contrastive Activation Addition (CAA), and lightweight weight modification via Direct Preference Optimization (DPO) trained on benchmark-derived factual data as well as conceptual generalization data. To assess alignment, we curate a multilingual factual dataset alongside a novel generalization benchmark comprising culturally rooted queries to determine whether factual interventions transfer to broader target-centric preferences. Experiments on Gemma 3 12B Instruct reveal persona prompting to be the strongest overall intervention, balancing efficacy, safety, and out-of-domain generalization. While CAA yields sharp inconsistency benchmark shifts, it is configuration-sensitive and risks knowledge degradation. DPO-based adapters offer permanent, yet narrower and less transferable gains. These findings suggest that cross-lingual inconsistency is at least partly a selection problem, and that simple contextual interventions may outperform more invasive methods for robust, transferable alignment.

We study the inference-time behavior of deep linear encoder-only transformers through the lens of interacting particle systems. In this perspective, tokens are modeled as particles that interact dynamically through successive linear self-attention layers. We show that in embedding dimension two, for any key, query, and value matrices, the dynamics can be reformulated as a generalized Kuramoto-type model with pure second-harmonic coupling. This formulation is amenable to Watanabe--Strogatz theory which reveals the dynamics are intrinsically low-dimensional regardless of the parameter matrices. For a class of token initializations associated with the Ott--Antonsen (OA) manifold, we show that the parameter matrices induce a diverse variety of long-time behaviors in linear transformers, including clustering, oscillations, and bifurcations. The oscillations and bifurcations are characterized by uncovering a hidden Hamiltonian structure in the dynamics. By establishing a structural stability result, we further show that dynamics initialized near the OA manifold exhibit the same long-time behavior as those initialized exactly on the manifold. Motivated by our theory in dimension two, we conduct numerical experiments for analogous parameter regimes in higher-dimensional transformers. Our numerical experiments suggest that the long-time behaviors characterized in our theoretical results persist in higher dimensions.

The prevailing inference framework for diffusion models formulates generation fundamentally as a problem of numerical integration. This perspective casts the model as an exact estimator, neglecting the inherent statistical uncertainty of the denoising process. In this work, we propose Forward-Process Aligned Diffusion prediction (\textbf{DiFA}), a training-free framework that reframes inference-time data prediction refinement as a sequential state estimation problem. Rather than reusing past outputs solely for numerical integration, DiFA treats iterative data predictions along the reverse trajectory as correlated observations to build a forward-aligned temporal consensus. Inspired by Kalman filtering, this consensus aggregates historical predictions according to structural consistency and noise-level compatibility. To counteract the over-smoothing tendency of temporal consensus, we introduce a deviation guidance mechanism to adaptively preserve residual details. Empirically, DiFA yields significant improvements on CIFAR-10 and ImageNet across the evaluated metrics, including FID, IS, and FD-DINOv2, demonstrating that aligning inference with the forward statistical structure substantially improves generative fidelity.

Regional climate prediction presents unique challenges for time series foundation models, which typically process temporal patterns through single-pass inference. Expert climatologists, in contrast, employ multi-scale temporal analysis and iterative refinement based on systematic error diagnosis. We present RGMR (Residual-Guided Multi-Resolution Refinement), an inference-time framework that adapts pre-trained foundation models to perform structured coarse-to-fine refinement for climate forecasting without updating backbone parameters. Applied to drought forecasting using the Standardized Precipitation Evapotranspiration Index (SPEI), RGMR is architecture-agnostic across the three TSFM backbones evaluated per site (TimesFM, TimeGPT, TabPFN) and consistently lowers test-set MSE on three South Australian sites and three additional regions outside South Australia. Applied to TimesFM, the wrapper reduces one-month-ahead SPEI MSE by up to 18.9% across the three South Australian sites (mean reduction $\approx$18.7%). Overall, RGMR provides a practical route for deploying frozen TSFMs in regional climate forecasting workflows.

Retrieval-Augmented Generation (RAG) systems retrieve and integrate external knowledge to ground large language model (LLM) outputs. However, current RAG architectures treat all retrieved facts as equally valid regardless of temporal provenance, leading to temporal hallucination, where plausible but obsolete facts corrupt the output. A clinical lab reading from yesterday is actionable; the same reading from six months ago is noise. We present Chronofy, a three-layer neuro-symbolic framework implementing the Temporal-Logical Decay Architecture (TLDA) that embeds temporal validity directly into the representation, retrieval, and reasoning layers of RAG systems. Layer 1 reserves a dedicated temporal subspace within Matryoshka embeddings to make fact age structurally irremovable from the representation. Layer 2 integrates learnable exponential decay functions into graph-based retrieval, where the decay coefficient $β_j$ is grounded in Bayesian decision theory as an approximation of twice the latent process mean-reversion rate. Layer 3 applies Signal Temporal Logic (STL) robustness functions to evaluate the temporal validity of retrieved knowledge, not LLM output confidence, and enforces the possibilistic weakest-link principle to bound output confidence by the most decayed evidence in the reasoning chain. We evaluate Chronofy on temporal knowledge graph forecasting benchmarks, the TimE temporal QA benchmark, and a domain-specific sensitivity analysis, demonstrating that explicit temporal decay modeling improves retrieval precision, reduces temporal hallucination, and enables principled data re-acquisition triggers when temporal context is insufficient.

Under model--harness co-evolution, harnesses are not merely inference-time scaffolds but data-generating components whose execution traces can shape future foundation models. This motivates harness-in-the-loop learning: optimizing harnesses for both immediate agent performance and the quality of traces used for future model training. However, continually updating provider-built scaffolds is costly and labor-intensive. We therefore investigate whether optimizing user-constructed harnesses in a task-specific manner can improve execution-trace quality while remaining computationally lightweight and requiring only a few update iterations. To this end, we introduce Recursive Harness Self-Improvement (RHI), which represents the harness as a prompt-level specification of the agent loop and iteratively refines it using pairwise feedback over its own revision history. Across 30 synthetic machine-learning research tasks spanning quantitative finance, robotics, and pharmacy, a few RHI iterations suffice to substantially raise the performance ceiling of low-reasoning-effort agents, exceeding the corresponding maximum-reasoning-effort setting while reducing inference cost by up to 60%. We show that these gains arise primarily from improved task-specific context management through more effective inter-agent information flow rather than longer reasoning traces. Finally, we formalize this behavior as an information-theoretic hypothesis for RHI's implicit optimization objective, suggesting RHI as a practical algorithm for continual learning within the paradigm of model--harness co-evolution.

In recent years Flow Matching has become a prominent method for generative modeling robot motion generation. In its generic form Flow Matching is an ODE-based neural sampler that is trained by regressing empirical flow fields associated with motion samples as data. However, in robot motion generation we often have additional constraints that might not be present in the collected data. The majority of current approaches train the flow on the available data and use inference-time guidance to enforce task-specific constraints. To address this mismatch, we propose \textbf{ConFlow}, a constraint-guided flow matching framework that incorporates constraint information directly into the training objective via differentiable barrier or cost functions. To address design specifications such as smoothness and boundary conditions, we propose replacing the standard Gaussian source distribution used in flow matching training with a conditional Gaussian Process. Our approach also uses infeasible demonstrations as negative supervision, improving constraint satisfaction without requiring additional expert data. Experiments on a two-robot navigation task demonstrate that ConFlow achieves lower collision rates and higher trajectory quality than standard flow matching baselines, with or without inference-time guidance. These results validate training-time constraint integration as an effective approach to closing the training--inference gap in generative motion models.

Large language models are increasingly deployed as agents, but reliable agentic behavior requires more than next-token prediction. At inference time, it is preferred that an agent can decide whether to proceed with its current reasoning, defer to a stronger model, request additional information, invoke external tools, or abstain under the given setup. Existing approaches address these decisions through prompt-level routing, external orchestration, or task-specific fine-tuning, which primarily rely on input-side signals, and are often costly and difficult to maintain as model backbones evolve. We ask whether such control decisions can be inferred directly from a model's latent generation process. We introduce Multi-Head Latent Control, a lightweight layer that reads hidden-state trajectories from a frozen LLM or VLM to produce deployment-time control signals. A Capability Head predicts whether the current model can solve the instance or should defer to a stronger collaborator, while a Resolution Head predicts appropriate resolution decision Clarification, Tool Use, Abstention, or Direct Answering. Both heads are trained only on latent traces from the same frozen LLM backbone, enabling post hoc adaptation without modifying the model. Across language and vision-language settings, Multi-Head Latent Control consistently improves the quality-cost tradeoff of multi-model systems, enabling early handoff from partial generations and more accurate intervention decisions. In routed execution (small + large model), it reduces large-model usage by up to 90.7 percent on AndroidWorld and 27-53 percent on average across benchmarks, while retaining most of large-model performance. Additionally, the learned control signals improve tool-use decision quality, yielding up to +158 percent relative score gain and 65.5 percent fewer missed-required tool calls.

Text-to-video (T2V) generators can synthesize realistic and temporally coherent videos, but controllably removing a target concept from a generator remains difficult. Unlike text-to-image concept erasure, T2V unlearning must suppress a target concept that may persist across frames while preserving non-target subjects, actions, scenes, and temporal structure. We propose \textbf{SIRUS}, a training-free inference-time framework for concept-level T2V unlearning. Given textual aliases of a target concept, SIRUS localizes target-related prompt evidence and suppresses target expression during sampling, without updating the text encoder or denoising network. We further introduce a video-oriented evaluation framework for T2V unlearning that separately measures target forgetting, non-target preservation, video quality, jailbreak robustness, and efficiency, using video-level failure criteria, frame-level residue statistics, paired preservation analysis, VBench-based quality diagnostics, and deployment overhead measurement. Across five safety, object, and style concepts on CogVideoX, SIRUS reaches 70.4% average forgetting success and 25.7% average frame hit, compared with 44.4% / 47.2% for VideoEraser, while reducing the average VBench quality drop from -0.043 to -0.016, yielding the strongest forgetting-quality trade-off among fully evaluated baselines. Transfer experiments on Wan2.2 further suggest that SIRUS generalizes across modern T2V backbones.