Small Models

A wrong number is worse than no answer. Across factuality-critical domains -- audience metrics, scheduling and rights in media; dosages and lab values in healthcare; figures and citations in finance and legal -- a confident but fabricated value is more damaging than an honest admission of uncertainty. Yet this is the dominant failure mode we observe on small local language models: even when correct evidence is present in context, models fabricate plausible numbers and timestamps. Recent work characterizes a real limit of this regime: below 7B parameters, the bottleneck of retrieval-augmented generation (RAG) is not retrieval quality but context utilization. We present Noesis, the deterministic-first query plane of the Noesis architecture, which makes every deterministic judgment before generation. Its mechanisms follow from the ingestion architecture (subject of a separate patent application): (a) a producer-side fact layer rendering precomputed metric facts verbatim without ranking; (b) positional addressing with deterministic cross-source alignment, resolved ahead of query time at zero LLM cost; (c) provenance scoping as an attribution constraint with multi-tier named-reference routing; and (d) two-tier context with model-triggered verbatim hydration. Across four ablations, a 2B model reaches parity with a 35B model on factual integrity (exact values in all runs; zero confabulated numbers on absent-entity traps); structured retrieval beats flat RAG by +11.4 points at 2B; skeleton-only context preserves quantitative answers at 20-30% smaller prompts; and hydration recovers verbatim narrative in ~8s versus ~29s. Two properties matter for regulated domains: each query resolves in a single generation call, and every reported value is traceable to its exact source and position by construction.

General-purpose code embeddings power tools for code search, classification, and retrieval. Compact transformer encoders for code typically rely on either human-written docstrings (labor-intensive and inconsistent) or mined structural signals such as execution traces (setting-specific and costly to collect). We empirically study an alternative: contrastive pretraining of small encoders with synthetically generated natural-language descriptions emphasizing code functionality and intent, paired with code in a dual-encoder framework at training and discarded at inference. We benchmark this approach against pretraining-based baselines, generalist LLMs, and embedding-specific models on eight retrieval, classification, and generation tasks across C, C++, and Java. Synthetic semantic supervision yields statistically significant gains over pretraining baselines of the same inference-time size on five of eight tasks, with parity on two more; once fine-tuned, it matches or exceeds zero-shot models two orders of magnitude larger on classification, and it stays on par with execution-aware supervision at matched pretraining data, suggesting a scalable, effective alternative to existing code-representation paradigms.

Multimodal Large Language Models (MLLMs) are strong perceivers of images and video. We ask how far that reach extends into acting: dropping an MLLM directly into a drone's control loop, with its entire action space declared solely in the prompt. Recent systems approach this setting but increasingly narrow the model's decision-making. We widen it back. We introduce DroneCATS-Agent, an architecture where the MLLM is a swappable component, and DroneCATS, a benchmark treating the model as the independent variable. Beyond merely flying toward a pixel, our agent entrusts the model to yaw and search, deliberate when unsure, and self-declare arrival---all without fine-tuning or function-calling schemas. Evaluating frontier and open models across four core capabilities---approaching a visible target, tracking a moving one, searching outside the initial view, and commanding a multi-drone fleet---reveals that even the simplest embodied settings are far from solved. Crucially, to identify what breaks first at the edge, our roster scales down to 2B parameters. The findings expose a stark paradox: it is not the flying that fails. Small open models often navigate into the success radius more reliably than frontier models, yet lose the episode by declaring arrival prematurely or not at all. Multi-drone commanding amplifies this divide, with small models failing by blindly copying a single coordinate across distinct views. Viewed as vision-language-action agents, the models' spatial perception holds up, but their action protocol does not. What separates a deployable edge model from a frontier model is not navigation, but the discipline to sustain a declared protocol and emit the correct terminating action. The open problem is closing this gap at onboard compute costs---yielding a fast model that plans persistently and knows exactly when it is done---and DroneCATS is built to measure that distance.

Accurate molecular property prediction requires both statistical reliability and chemical reasoning. Graph neural networks can be calibrated directly on labeled assays but remain limited by the coverage of their training data. Large language models (LLMs) can compare molecular evidence and articulate chemical rationales, yet are unreliable as standalone quantitative predictors. The central challenge is therefore to determine when an LLM should influence a calibrated model and by how much. Here we present CoMPASS, a retrieval-calibrated framework for small-large model collaboration. CoMPASS retains a graph attention network (GAT) as the predictive anchor, retrieves locally relevant training molecules, provides attention-grounded evidence to an LLM, and converts its proposal into a bounded correction through an agreement-aware gate. Across six classification and two regression benchmarks, CoMPASS improves the GAT anchor in regions of correctable uncertainty while limiting LLM intervention in high-confidence regimes. Ablations show that the gains arise from validation-calibrated retrieval and bounded fusion rather than prompting alone. These results suggest that generative reasoning should augment calibrated prediction through evidence-grounded, controlled corrections rather than direct output replacement. Code is available at https://github.com/littlepeachs/CoMPASS.

Fine-tuning can improve legal question-answering accuracy without improving how models use law supplied in context. We study this distinction in bilingual Bangladeshi legal QA, where observed errors can arise from answer scoring, retrieval, or failure to use relevant law. We construct a hierarchy-preserving statutory corpus, 2,165 reviewed bilingual fine-tuning examples, and a 150-item supplied-law control. We evaluate six instruction-tuned models: Llama-3.2-1B, Llama-3.2-3B, Qwen3.5-0.8B, Qwen3.5-2B, Qwen3.5-4B, and Gemma-4-E2B, with three LoRA seeds per model. To separate effects, we combine constrained option-letter scoring, cyclic option rotation, and controlled removal of the governing provision. On 398 Bar Council outputs, an exact-line parser attributes an accuracy gain of 50.0% to the Qwen3.5-2B seed-42 adapter, whereas option scoring yields only $3.0\%$. For Gemma-4-E2B, the two scoring methods favor different systems. When the governing provision is guaranteed to be present, five of six reference models improve by $14.7\%-19.3\%$ under the four-order criterion. Removing that provision reduces accuracy by $8.0\%-15.3\%$ for models and by $13.8\%-14.9\%$ points for their adapters. However, difference-in differences estimates show no increase in reliance on the governing provision after fine-tuning. Results show that legal adaptation claims require separating scorer, retriever, and model effects. Our Code and data are available at https://anonymous.4open.science/r/bangladesh-legal-qa-11E3

Large language model (LLM) agents can retrieve memory, call tools, ask clarifying questions, and vary response style, yet adapting these execution decisions to an individual user remains difficult. Fine-tuning a separate LLM is costly or impossible for proprietary systems, while prompts and memory primarily expose user information to the agent rather than adapt its execution decisions from feedback. We formulate personalization of a frozen agent as online learning of a per-user execution policy from scalar feedback observed only for the executed action. We propose FABLE (Factorized Adaptive Bandit Layer for Execution), a lightweight policy layer outside a potentially black-box host agent. FABLE factorizes memory, information-acquisition, and response decisions so feedback updates related choices; filters actions through an externally specified feasible set before exploration; and learns user-specific residual preferences relative to a fixed default-and-cost score via Bayesian contextual Thompson sampling. Under a linear residual-reward model, a calibrated variant inherits an expected-regret bound against the best feasible action. We also characterize preferences unidentifiable under persistent feasibility constraints and provide anytime-valid false-promotion control. Across personalized-reasoning, controlled-feedback, and executable tool-use evaluations, FABLE improves several preference-sensitive behaviors relative to rule-only control while remaining competitive on end-to-end task performance.

Vision-language models (VLMs) deployed on consumer hardware must decide when to answer and when to defer, and that decision depends on having a confidence signal that tracks correctness. A practitioner with a fixed memory budget faces a choice between a small model at full precision, the same small model quantized, and a larger model quantized into the same footprint -- three configurations that push the confidence signal in opposing directions. We measure, on identical inputs, how model scale and 4-bit quantization affect two confidence signals in the Qwen2-VL family: the confidence a model states in natural language, and its own mean token probability over the answer it generates. Across 5,700 predictions spanning six realistic photographic degradations at three severities, we find that scale sharply improves the model's internal uncertainty signal (mean error-detection AUROC 0.80 to 0.98 from 2B to 7B) while its verbalized confidence stays weak and often at chance (mean 0.61 to 0.69): the gap between what the model knows and what it says widens rather than closes with size. We find that 4-bit quantization is nearly free for accuracy (-1.6 points) but expensive for the confidence signal (internal AUROC 0.95 to 0.80, and the verbalized-confidence parse rate collapses from 99% to 64%). For a fixed memory budget the recommendation is therefore to prefer a larger quantized model over a smaller full-precision one: 7B-4bit gives both the best accuracy and the best uncertainty signal (internal AUROC 0.98) of the three configurations that fit. We frame the results as selective-prediction operating points so they translate directly into a deployment recommendation, and we argue that error-detection AUROC, not calibration error, is the metric that exposes the difference between the two signals.

A small language model can receive the governing statutory provision and still answer incorrectly. We test whether fine-tuning on examples containing relevant law improves later use of retrieved law. We curate 2{,}165 bilingual QA records from six Bangladeshi acts and three schedules, then fine-tune Qwen3.5 at 0.8B, 2B, and 4B. Evaluation uses the 2022 and 2023 Bangladesh Bar Council exams in Bangla and machine-translated English, with no retrieval, BM25, or FAISS, scored by strict consistency over three seeded runs. At 0.8B, fine-tuning raises the 2022 English FAISS score from 2 to 34 of 100. Gains at 0.8B and 2B survive paired testing, but the 4B model has no detectable net gain: Bangla improves while several English conditions regress. Fine-tuning also reduces answers that drift from Bangla into mostly English from 44.0--53.2% to 0.2--0.7%, with adjusted $p<.001$ at every scale. Retrieval quality is therefore not the only bottleneck. Small bilingual legal models also differ in how they use supplied law and whether they answer in the requested language. The dataset is publicly available at https://huggingface.co/datasets/momahadi/bangladesh-legal-qa-dataset.

Transformers have demonstrated a remarkable ability to learn algorithmic reasoning, yet mechanistic analyses have mostly focused on globally invertible operations such as cyclic addition and group composition. In this work, we investigate how small transformers learn modular integer multiplication over composite moduli, a fundamentally non-invertible operation due to the presence of zero-divisors. We propose the monoid extension: a localized generalization of Group Composition via Representation (GCR) that suggests the learned computation does not rely on a single global representation space. Instead, the model partitions the input space into local hierarchical algebraic regions, where group-like structure survives and Fourier mechanisms can be applied. In transformers trained on square-free modular multiplication, we find that embeddings organize around these regions, attention exhibits class-sensitive routing and low-rank write directions, and local character features explain a large fraction of the model's output logits. Our results suggest that representation-theoretic mechanisms previously identified for group operations can extend beyond groups to more general structures.

Modern sequence models are increasingly deployed as agents that interleave token generation with calls to external tools. We give an exact, architecture-level account of when such tool access increases computational expressivity. We model any fixed finite-precision recurrent sequence model, including finite-precision state-space models (SSMs) with $B$ bits of internal state, as a deterministic finite-state controller interacting with an oracle through a finite command/observation interface. Our results form a sharp dichotomy. First, tools that are themselves finite-state add essentially nothing: a product-state simulation internalizes any finite-state bounded-interface oracle with finite memory set $M$ at a cost of only $\log_2 |M| + O(1)$ additional bits, so the augmented system remains finite-state. Second, a single minimal infinite-state tool, namely a tape supporting only local $\mathtt{read}$, $\mathtt{write}$, and $\mathtt{move}$ commands, makes the system Turing complete: for every single-tape Turing machine with state set $Q$ and tape alphabet $Γ$, a controller with $O(\log |Q| + \log |Γ|)$ bits of internal memory simulates it, and we exhibit a concrete exponential separation: $\mathrm{EQ}_n$ requires $2^n$ states without tools but a single constant-size controller with the tape tool. Third, we show that this construction is realized exactly by a natural one-layer finite-precision selective affine SSM controller with binary one-hot hidden states, $\{0,1\}$ transition matrices, and zero biases. Selectivity is essential to the construction. In the supplementary material, we make all constants explicit, prove a logarithmic oracle-assisted universal simulation, where $O(\log B)$ recurrent bits suffice to simulate any $B$-state Turing machine, and prove a matching impossibility result.

Small ($\sim$2B) GUI-grounding agents are attractive for on-device deployment, accessibility tooling, and low-cost iteration, but at this scale they face two open recipe questions: how to obtain bounding-box training data without expensive human annotation, and how to combine supervised fine-tuning with reinforcement learning. We address both, with the explicit goal of pushing small-model performance rather than scaling up. WinDOM is a $54{,}425$-record grounding corpus harvested by driving an open-source Windows 11 web reimplementation under headless Playwright, with bounding boxes read directly off the DOM and no OCR or human annotation. Self-Family Distillation (SFD) is a single rejection-sampling cold-start parameterised only by the teacher choice: either an EMA of the student (no external model) or a frozen larger same-family teacher. We then treat the saturation depth of the SFD cold-start as an explicit GRPO hyperparameter. On a Qwen3.5-2B student, the under-saturated cold-start is a better GRPO initialiser than the converged one: SFD-4B with Early-init RL gains $+5.4$ OOD-mean ($+3.5$ ScreenSpot-Pro, $+7.0$ OSWorld-G, $+5.8$ ScreenSpot-V2) over the base. The same-size EMA mode lands within roughly one OOD-mean point of the cross-size $4$B variant ($65.2$ vs $66.3$) without an external teacher.

Small open-weight models struggle at long-form creative writing: their generated stories either fall far short of the requested length, or their quality significantly degrades as length increases, especially when compared to frontier models. We present POLARIS (Policy Optimization with LLM-as-a-judge rewards and Anchored-Reference Injection for Storywriting), a lower-compute GRPO recipe with two key ingredients: a frontier LLM judge with a structured Story Quality rubric as the online reward, and human-reference injection (HRI), where a teacher-forced human-written story serves as a high-reward anchor within each GRPO group. By applying our training recipe to Qwen3.5-9B, using a dataset of approximately 1.4K prompt-story pairs derived from 100 short-story anthologies and 4 A100 GPUs, we obtain POLARIS-9B. Across five benchmarks spanning in-distribution and out-of-distribution prompts and rubrics, POLARIS-9B is competitive with much larger open-weight models while following length instructions more closely. A blinded human evaluation confirms that POLARIS-9B is preferred to the base Qwen3.5-9B and on par with Qwen3.5-27B. Despite training only on stories up to 4k words, POLARIS-9B preserves quality on prompts requesting stories up to 3 times the training length, a regime where most open-weight models degrade substantially in quality, length adherence, or both. More broadly, our results suggest that length generalization is a meaningful stress test for creative-writing models and a useful lens for distinguishing otherwise close models.

We identify a new dimension for enhancing rollout diversity in Group Relative Policy Optimization (GRPO) for LLMs. While GRPO relies on diverse rollouts, prevailing strategies primarily increase diversity by injecting more token-level randomness, which may introduce step-wise noise and lead to incoherent trajectories. We uncover that smaller models within the same model family inherently exhibit higher policy-level diversity, indicated by their superior pass@k relative to larger counterparts as sample counts increase. Unlike token-level noise, this diversity is temporally correlated, preserves logical consistency, and provides structured exploration signals for gradient estimation. We thus propose S2L-PO (Small-to-Large Policy Optimization), a framework that leverages fixed small models as natural explorers to train larger models. To balance exploration and exploitation, we design a progressive annealing strategy that transitions from offline small-model rollouts to the large learner's own sampling. This shift elegantly avoids mid-training performance drops caused by the small model's capacity limits, achieving faster convergence and unlocking a higher performance ceiling. S2L-PO improves accuracy on diverse mathematical reasoning benchmarks (e.g., +8.8% on AIME 24 using a 1.7B explorer to guide the 8B model) while reducing rollout compute.

Childhood anemia affects around 40% of children aged 6-59 months globally and arises from heterogeneous factors, limiting model generalizability. We evaluate a transformer-based tabular foundation model against classical supervised methods under cross-country and data-scarce settings. We used DHS data from 16 countries across Africa, Asia, Latin America, the Caucasus, and the Middle East (n=68,856). We compared Logistic Regression, XGBoost, LightGBM, and TabPFN v2.6. Performance was assessed using AUC-ROC, Brier score, and ECE. Generalization was evaluated using leave-one-country-out (LOCO), reverse-LOCO, and few-shot settings. Subgroup analyses included sex, age, residence, maternal education, and wealth. Feature importance was estimated using SHAP. TabPFN outperformed classical models in low-data regimes (<200 samples), showing higher discrimination and better calibration. Across countries, it achieved the lowest Brier score (0.042) and ECE (0.203). Under full-data settings, AUC-ROC ranged from 0.59-0.76 with small between-model differences ($\leq 0.05$). LOCO performance was stable (0.58-0.69), driven by country context. Reverse-LOCO showed asymmetric transferability. Subgroup performance was consistent with no systematic demographic bias. SHAP identified child age, altitude, and height-for-age z-score as dominant predictors, followed by wealth and maternal education. Performance in childhood anemia prediction is driven more by population variation than model choice. TabPFN provides advantages in low-resource settings through improved discrimination and calibration, highlighting foundation models as promising tools for data-scarce global health prediction.

Neural PDE solvers have followed the scaling trajectory of vision and language, with recent foundation models reaching billions of parameters. We argue that scale is a poor substitute for architectural inductive bias in this domain: structured priors deliver outsized parameter efficiency, and the pattern of where they succeed and fail is itself informative about what they capture. We instantiate this argument in WaveLiT, an architecture combining a discrete wavelet transform for lossless multi-resolution tokenization, an augmented linear attention block, a shared-weight multiscale feature pyramid, and a wavelet-domain auxiliary loss. Bespoke 1-10M-parameter WaveLiT models compete with foundation models of 100-1000$\times$ their size across eight TheWell benchmarks, with the largest gains on wave and acoustic-dominated benchmarks where the wavelet-multiscale prior fits the dominant dynamical structure and small per-step errors do not compound geometrically under rollout. Trained jointly across all eight benchmarks, a 10M-parameter foundation variant exhibits a structured, physically interpretable transfer pattern -- strongest where the wavelet-multiscale prior matches the dynamics, weakest on chaotic advection-dominated flows. The entire pipeline trains on a single GPU. The results suggest that small-model PDE performance is shaped by architectural inductive bias rather than scale, and that the structure of a prior's failures is a useful empirical signal about its content.

Cardiovascular disease remains the leading cause of death worldwide, and early detection of arrhythmias through continuous ECG monitoring on wearable devices can prevent life-threatening events. Federated Learning (FL) enables privacy-preserving collaborative training by keeping raw ECG data on device, yet standard FL incurs prohibitive communication overhead and standard deep learning models cannot fit on ultra-low-power microcontrollers. We propose Family-Grouped Hierarchical Federated Learning (Family-FL), a three-tier architecture that uses the family as a natural privacy boundary for intra-family aggregation before global synchronization. We further design a hardware-constrained Tiny CNN-LSTM architecture with only 669 parameters, INT8-quantized to occupy merely 4.65KB Flash and 2.95KB RAM, meeting the constraints of STC32G12K128-class microcontrollers. Experiments on the MIT-BIH Arrhythmia Database (mean of 5 independent runs with different seeds) demonstrate that Family-FL reduces communication volume by 76.7% compared to FedAvg while maintaining comparable accuracy. Family-FL-Tiny achieves 91.9 +/- 1.2% accuracy with macro-F1 of 0.483 +/- 0.031, reducing total communication to 0.31% of FedAvg. The model achieves reliable ventricular arrhythmia detection (per-class F1 = 0.80), the most clinically critical abnormality for home-based preliminary screening. These results demonstrate the technical feasibility of privacy-preserving federated learning on ultra-resource-constrained microcontrollers through simulation-based evaluation. We honestly discuss limitations: no hardware deployment, single-dataset validation (MIT-BIH, 47 subjects), reduced rare-class sensitivity, and absence of formal differential privacy guarantees.

Critical-data-size accounts of grokking suggest a natural post-threshold intuition: once training data is sufficient to identify the underlying rule, additional data should accelerate validation convergence. We show that this intuition can fail in a controlled structured-output task. In Needleman--Wunsch (NW) matrix generation, small Transformers reach high validation exact-match accuracy fastest at an intermediate dataset size, not at the largest one. Past this dataset-size sweet spot, generalization remains achievable but requires more gradient updates. Conversely, in the regime where partial validation competence first appears, larger datasets can require fewer updates to reach high training accuracy, suggesting that emerging rule structure can accelerate fitting beyond example-wise memorization. A multiplication baseline does not show the same post-threshold slowdown. These results separate the critical data size for the onset of generalization from the dataset size that optimizes update-based convergence, and identify structured-output tasks where learning the rule and completing exact-fitting can diverge.

Recent studies on long-term time series forecasting have shown that simple linear models and MLP-based predictors can achieve strong performance without increasingly complex architectures. However, many competitive baselines still rely on structural priors such as frequency-domain modeling, explicit decomposition, multi-scale mixing, or sophisticated cross-variable interaction modules, while paying less attention to how simple temporal mappings should be trained and organized. In this paper, we propose STAIR, short for Stagewise Temporal Adaptation via Individualization and Residual Learning, a training paradigm for long-term time series forecasting that aims to unlock the capacity of simple temporal mapping models without introducing complex architectural modules. STAIR decomposes forecasting ability into three progressive stages: it first learns common temporal dynamics across variables through a shared temporal mapping, then adapts the shared model to each variable via channel-wise fine-tuning to capture variable-specific patterns, and finally complements the backbone with cross-variable information through residual learning. We further introduce Shared-to-Individual Fine-tuning and alpha-RevIN to mitigate the limitations of strict channel independence and the overly strong normalization prior induced by standard RevIN. This design gradually increases modeling flexibility while keeping the core temporal predictor as a shallow MLP in the main experiments, with linear variants analyzed separately. Experiments on nine long-term forecasting benchmarks show that STAIR matches or outperforms recent strong baselines while preserving a simple temporal backbone, providing a concise and effective modeling perspective for long-term time series forecasting.

Monitoring the chain-of-thought (CoT) of reasoning models is a promising approach for detecting covert misbehavior (i.e., hidden objectives) in code generation tasks. While large models (GPT-5, Gemini-3-Flash) can serve as effective CoT monitors, they are expensive to deploy due to the lengthy reasoning traces and high API cost, emphasizing the need for smaller, cheaper alternatives. Nevertheless, we find that current small models (4B--8B) struggle to detect hidden objectives despite access to the CoT, frequently misattributing them as part of the user query. To address this, we propose a post-training pipeline combining supervised fine-tuning (SFT) and reinforcement learning (RL), where SFT narrows the gap for in-domain tasks by distilling detection behavior from stronger monitors, and RL on hard and subtly crafted hidden objectives helps the model generalize to out-of-domain monitoring tasks. To validate this generalization, we evaluate under a realistic threat model motivated by practical supply-chain attacks, where the adversary is a third-party LLM router injecting hidden objectives into code-generation requests through either prompt manipulation or code manipulation attacks. To push beyond objectives that large monitors already saturate, we also introduce four new challenging tasks even for strong monitors. Finally, we introduce CoT-Guard, a 4B-parameter monitor that demonstrates superior generalization performance under both prompt and code manipulation attacks, achieving a G-mean^2 (i.e., TNR x TPR) of 75% and outperforming GPT-5.4 (56%), GPT-5-mini (41%), and Qwen3-32B (54%), while closing the gap to Gemini-3-Flash (83%). These results demonstrate that CoT-Guard provides a practical and cost-effective user-side defense, substantially improving hidden-objective detection while avoiding the deployment cost of large monitors.

Architectural choices inside the Transformer feedforward network (FFN) block do not merely affect the block itself; they reshape the computations learned by the rest of the model. We study this effect in one-layer Transformers trained on digit addition with carry, modular arithmetic, and histogram counting. Comparing dense FFNs, gated linear units (GLUs), mixture-of-experts (MoE), and MoE-GLUs, we find that sparse MoE routing can shift computation from FFN to attention, with the strongest ablation-visible effect on carry-based addition. We decompose this redistribution into reduced per-token FFN capacity and sparse partitioning across experts. Critically, frozen random routing nearly matches learned routing, suggesting that redistribution is driven largely by architectural sparsity rather than router-learned specialization. As a secondary finding, GLU-style multiplicative gating rotates task-relevant Fourier structure out of the per-neuron basis and into distributed subspaces, making neuron-level interpretability less informative while preserving structured computation. We validate these conclusions with random-routing, narrow-FFN, and top-2 MoE controls, plus parameter-matching, activation-function, and width-scaling analyses. Together, these results show that local FFN design choices can have nonlocal consequences for Transformer computation.

Speculative decoding accelerates large language model (LLM) inference by using a small draft model to propose candidate tokens that a larger target model verifies. A critical hyperparameter in this process is the speculation length $γ$, which determines how many tokens the draft model proposes per step. Nearly all existing systems use a fixed $γ$ (typically 4), yet empirical evidence suggests that the optimal value varies across task types and, crucially, depends on the compression level applied to the target model. In this paper, we present SpecKV, a lightweight adaptive controller that selects $γ$ per speculation step using signals extracted from the draft model itself. We profile speculative decoding across 4 task categories, 4 speculation lengths, and 3 compression levels (FP16, INT8, NF4), collecting 5,112 step-level records with per-step acceptance rates, draft entropy, and draft confidence. We demonstrate that the optimal $γ$ shifts across compression regimes and that draft model confidence and entropy are strong predictors of acceptance rate (correlation $\approx$ 0.56). SpecKV uses a small MLP trained on these signals to maximize expected tokens per speculation step, achieving a 56.0% improvement over the fixed-$γ=4$ baseline with only 0.34 ms overhead per decision (<0.5% of step time). The improvement is statistically significant (p < 0.001, paired bootstrap test). We release all profiling data, trained models, and notebooks as open-source artifacts.

Production agentic systems make many model calls per user request, and most of those calls are short, structured, and routine. This raises a practical routing question that existing evaluations do not directly answer: which parts of an agent workflow truly require large frontier intelligence, and which can be handled by smaller models? We introduce AgentFloor, a deterministic 30-task benchmark organized as a six-tier capability ladder, spanning instruction following, tool use, multi-step coordination, and long-horizon planning under persistent constraints. We evaluate 16 open-weight models, from 0.27B to 32B parameters, alongside GPT-5 across 16,542 scored runs. Our results reveal a clear boundary of model necessity. Small and mid-sized open-weight models are already sufficient for much of the short-horizon, structured tool use work that dominates real agent pipelines, and in aggregate, the strongest open-weight model matches GPT-5 on our benchmark while being substantially cheaper and faster to run. The gap appears most clearly on long-horizon planning tasks that require sustained coordination and reliable constraint tracking over many steps, where frontier models still hold an advantage, though neither side reaches strong reliability. We also find that this boundary is not explained by scale alone: some failures respond to targeted interventions, but the effects are model-specific rather than universal. These findings suggest a practical design principle for agentic systems: use smaller open-weight models for the broad base of routine actions, and reserve large frontier models for the narrower class of tasks that truly demand deeper planning and control. We release the benchmark, harness, sweep configurations, and full run corpus.

We demonstrate that a small transformer can learn to execute programs in MicroPy, a simplified yet computationally universal programming language. Given procedure definitions together with an expression to evaluate, the transformer predicts small-step execution using PENCIL scaffolding for space-efficient execution within a bounded context window. After training on randomly generated, meaningless MicroPy programs, the learned transformer generalizes to various human-written programs including bit copying and flipping, binary addition and multiplication, and SAT verification and solving. We note that the trained model can achieve out-of-distribution generalization; i.e., evaluate novel programs from distribution on programs. Since MicroPy can express any computation, our results provide empirical evidence that a standard transformer can be trained to act as a universal computer.

We present Nautile-370M, a 371-million-parameter small language model designed for efficient reasoning under strict parameter and inference budgets. Nautile-370M uses a hybrid backbone in which two SeqCond Attention (SCA) layers, a linear-time spectral sequence operator inspired by SeqCondenser, alternate with one transformer layer. This design aims to retain the long-context efficiency and state-tracking benefits of structured sequential models while preserving the expressive token-to-token routing of attention. The model was trained on a single Cloud TPU v4-64 pod slice provided through the Google TPU Research Cloud (TRC) program; the subsequent reinforcement learning stage was carried out on a single NVIDIA DGX Spark. We prove that the SCA readout mechanism can exactly retrieve any individual token from the prefix summary and can reproduce any output of softmax attention as a special case, establishing that SCA is at least as expressive as full self-attention in the continuous limit. We also describe the training data pipeline and outline a reinforcement learning stage specialized for reasoning, verification, and response quality.

Multi-agent systems (MAS) demonstrate clear advantages in tackling complex problems by coordinating diverse agents and external tools. However, most existing orchestration methods rely on static workflows or serial agent scheduling, and are further constrained by heterogeneous interface protocols between tools and agents. This leads to high system complexity and poor extensibility. To mitigate these issues, we propose Agent-as-Tool, a unified parallel orchestration paradigm that abstracts both agents and tools into a standardized, learnable action space with protocol normalization and explicit state feedback. Building on this paradigm, we train a lightweight orchestrator, ParaManager, which decouples planning decisions from subtask solving, enabling state-aware parallel subtask decomposition, delegation, and asynchronous execution. For training, we adopt a two-stage ParaManager training pipeline. It improves robustness by incorporating supervised fine-tuning (SFT) trajectories equipped with recovery mechanisms, and further applies reinforcement learning (RL) to achieve an optimal balance among task success, protocol compliance, diversity, and reasoning efficiency. Experiments show that ParaManager achieves strong performance across multiple benchmarks and exhibits robust generalization under unseen model pools.

The significant computational demands of large language models have increased interest in distilling reasoning abilities into smaller models via Chain-of-Thought (CoT) distillation. Current CoT distillation methods mainly focus on transferring teacher-generated rationales for complex reasoning to student models. However, they do not adequately explore teachers' dynamic attention toward critical information during reasoning. We find that language models exhibit progressive attention shifts towards key information during reasoning, which implies essential clues for drawing conclusions. Building on this observation and analysis, we introduce a novel CoT distillation framework that transfers the teacher's stepwise attention on key information to the student model. This establishes structured guidance for the student's progressive concentration on key information during reasoning. More importantly, we develop a Mixture of Layers module enabling dynamic alignment that adapts to different layers between the teacher and student. Our method achieves consistent performance improvements across multiple mathematical and commonsense reasoning datasets. To our knowledge, it is the first method to leverage stepwise attention within CoT distillation to improve small model reasoning.

When do transformers commit to a decision, and what prevents them from correcting it? We introduce \textbf{prolepsis}: a transformer commits early, task-specific attention heads sustain the commitment, and no layer corrects it. Replicating \citeauthor{lindsey2025biology}'s (\citeyear{lindsey2025biology}) planning-site finding on open models (Gemma2 2B, Llama3.2 1B), we ask five questions. (Q1)~Planning is invisible to six residual-stream methods; CLTs are necessary. (Q2)~The planning-site spike replicates with identical geometry. (Q3)~Specific attention heads route the decision to the output, filling a gap flagged as invisible to attribution graphs. (Q4)~Search requires ${\leq}16$ layers; commitment requires more. (Q5)~Factual recall shows the same motif at a different network depth, with zero overlap between recurring planning heads and the factual top-10. Prolepsis is architectural: the template is shared, the routing substrates differ. All experiments run on a single consumer GPU (16,GB VRAM).

Reward models (RMs) play a central role throughout the language model (LM) pipeline, particularly in non-verifiable domains. However, the dominant LLM-as-a-Judge paradigm relies on the strong reasoning capabilities of large models, while alternative approaches require reference responses or explicit rubrics, limiting flexibility and broader accessibility. In this work, we propose FLIP (FLipped Inference for Prompt reconstruction), a reference-free and rubric-free reward modeling approach that reformulates reward modeling through backward inference: inferring the instruction that would most plausibly produce a given response. The similarity between the inferred and the original instructions is then used as the reward signal. Evaluations across four domains using 13 small language models show that FLIP outperforms LLM-as-a-Judge baselines by an average of 79.6%. Moreover, FLIP substantially improves downstream performance in extrinsic evaluations under test-time scaling via parallel sampling and GRPO training. We further find that FLIP is particularly effective for longer outputs and robust to common forms of reward hacking. By explicitly exploiting the validation-generation gap, FLIP enables reliable reward modeling in downscaled regimes where judgment methods fail. Code available at https://github.com/yikee/FLIP.

Modern large-scale neural networks are often trained and released in multiple sizes to accommodate diverse inference budgets. To improve efficiency, recent work has explored model upscaling: initializing larger models from trained smaller ones to accelerate convergence. However, this method can be sensitive to hyperparameters that need to be tuned at the target upscaled model size, which is prohibitively costly to do directly. It remains unclear whether tuning hyperparameters on smaller models and extrapolating via scaling laws is sound in this setting. We address this with principled approaches to width-based upscaling and efficient hyperparameter tuning in this setting. Motivated by $μ$P and any-dimensional architectures, we introduce a general upscaling method that, like Net2Net, copies and perturbs weights, but uses theoretically grounded, width-dependent scalings for the perturbation noise and optimizer hyperparameters. First, we prove that under zero perturbation, the upscaled model is functionally equivalent to the base model throughout training. Second, we extend the $μ$P theory to enable infinite-width limit analysis and establish hyperparameter transfer for upscaled models, greatly reducing the tuning cost. We empirically demonstrate that this method is effective on realistic datasets and architectures.

Limited-Area Models (LAMs) enable weather forecasting over regional domains at higher resolutions than what is computationally feasible for global models. At such high resolutions, machine learning approaches for weather prediction increasingly rely on ensemble methods to produce probabilistic forecasts. However, existing machine learning LAMs are not scalable due to relying on computationally costly diffusion models or inefficient graph neural networks. We tackle this by introducing a new hybrid CNN/GNN architecture, tailored to the LAM weather forecasting problem. Using this architecture, we construct the DET-LAM deterministic model, producing LAM forecasts both more efficiently and accurately than its graph-based competitor. We then tackle the ensemble forecasting problem, by using this architecture as a backbone for the generative model CRPS-LAM. CRPS-LAM is trained using a Continuous Ranked Probability Score (CRPS) objective, enabling efficient training and sampling in a single forward pass. This yields a speedup of $\approx \times 39$ compared to diffusion-based baselines. We evaluate our approach on regional domains in northern Europe, demonstrating that CRPS-LAM produces skillful and well-calibrated forecasts across a range of atmospheric variables.

Large language models (LMs) offer broad generalization capabilities but require vast amounts of data and computational resources for domain-specific tasks; small models (SMs), in contrast, are more efficient and tailored to specific domains yet lack general-purpose coverage. Taking a collaborative approach, where large and small models work synergistically, can accelerate the adaptation of LLMs to private domains and unlock new potential in AI. This survey presents a comprehensive overview of recent advances and challenges in harnessing the collaborative power of large and small models for private-domain adaptation. It specifically focuses on the unique constraints of cross-boundary environments, where models belong to distinct parties, and examines the resulting tensions among data privacy, model security, integrity, and resource limitations. By analyzing the information flow between distinct model and data stakeholders, we propose a unified taxonomy that classifies research into three primary directions: downward knowledge transfer (LM to SM), upward knowledge transfer (SM to LM), and inference-time collaboration across parties. Drawing on this taxonomy, we analyze the core challenges inherent to cross-boundary information exchange, including data-privacy, model-security, and integrity threats as well as efficiency constraints, and synthesize these into a multi-objective optimization problem that governs practical deployment. Finally, we review key open challenges inherent to such hybrid approaches and outline promising directions for future research. By offering a principled, boundary-centric view of this rapidly evolving landscape, this survey aims to serve as a structured resource for researchers and practitioners advancing privacy-aware, resource-efficient AI deployment.