Local language models can save memory by storing weights in fewer bits, and they can reduce sensitivity to the order of answer options by averaging a few reorderings. We ask whether the first makes the second worse, and which is the cheaper way to spend a fixed budget. On Qwen3-0.6B and Qwen3-1.7B with simulated round-to-nearest weight quantization (8, 6, 4 and 3 bits, fp16 reference), read from the logits of the letters A to D on ARC-Easy, ARC-Challenge, MMLU and HellaSwag (300 items each), 6 bits lose no accuracy against fp16 (they score +1.8 (+1.0 to +2.5) and +0.9 (+0.3 to +1.5) points higher, which we do not explain), and 8 bits change it by -0.7 (-1.1 to -0.3) and -0.4 (-0.7 to -0.2) points. At 4 bits the answer-slot skew rises from 0.15 to 0.53 (0.6B) and from 0.07 to 0.12 (1.7B); Gaussian noise at the same weight SNR gives 0.15 and 0.37 (two draws) and 0.19 (one draw), so the control is inconclusive on whether rounding adds slot skew beyond equal-power noise. Flip rates rise but stay below an accuracy-matched baseline of independent uniform errors. Averaging four cyclic orders recovers +12.6 points at 4 bits for 0.6B (+5.3 at fp16). We price bits and passes in one machine-independent unit, weight bytes read times passes. In cost-matched pairs, extra orders beat extra bits only when the weights have 6 or more bits, and the decision table reduces to: use 6-bit weights and spend the rest of the budget on up to four orders, one pass each. Everything is simulated on two small models with the simplest quantizer; a title-level search found no earlier study of quantization together with option order.
Concept-based explanations in neural models are important for the development, certification, and adoption of autonomous systems. Yet, interpretability remains correlational. Without methods to guide and quantify causal relationships between explanations, a concept explanation can be provided without being necessary or sufficient for the observed output thereby appearing correct while misleading. We introduce Causal Concept Wrapper Network (CCW-Net), the first use of interventional mediation analysis from fields such as economics, medicine, and epidemiology as a training objective to align causal effects of concept-based explanations with a model's output. We further apply mediation analysis as an evaluation tool to quantify, per concept label combination, how much of any model's output is causally attributable to each concept explanation. As a training objective, CCW-Net shapes a model's concept representations such that they are both necessary and sufficient for the outputs they are credited with and independent of other concepts' contributions. Counterfactual concept representations anchor a set of causal alignment criteria: necessity, sufficiency, and independence. Necessity drives irrelevant concept effects toward zero, sufficiency ensures relevant concept representations fully reconstruct a model's output, and independence enforces the total effect decomposes linearly across concepts. We evaluate CCW-Net across aircraft control, driving, and image classification. CCW-Net provides a step toward concept-based explanations that are not merely interpretable, but quantifiably causal, supporting safer evaluation, certification, and deployment of transparent and trustworthy neural models.
South Asian monsoon precipitation depends on the thermodynamic structure of the lower troposphere, but how its ingredients combine to shape precipitation remains unclear, partly because this dependence is nonlocal in the vertical. We use symbolic distillation, which combines neural networks and symbolic regression, to discover interpretable precipitation equations from 21 years of June-August reanalysis data. Neural networks first learn the vertical thermodynamic structure most relevant to precipitation, condensing each variable into a single interpretable feature not tied to a particular vertical grid. Then, symbolic regression derives closed-form equations from these features and the surface state. Our best equation explains 1.5 times as much precipitation variance as a widely used buoyancy-based diagnostic and recovers about 85% of neural network skill while using about a thousand times fewer parameters. Unlike the buoyancy diagnostic, which adds moisture and instability, the equations treat them as competing controls, with the heaviest precipitation occurring when the lower free troposphere is humid. Over ocean, sensible heating is associated with additional precipitation mainly when near-surface air is warm and moist, improving predicted precipitation and its diurnal cycle. Although the learned features were optimized only for precipitation prediction, the equations largely respond to each feature as expected from convective physics, consistent with the reanalysis. These responses reduce the dependence of monsoon precipitation on the entire vertical column to a few simple, testable constraints that could guide data-driven precipitation models to make them more physically consistent.
Speculative decoding accelerates LLM inference by having a draft model propose multiple tokens for the target model to verify in one pass. The speedup is governed by the expected acceptance length (EAL), yet standard training objectives such as cross-entropy, KL divergence, and total variation focus only on distribution matching or single-step overlap. Recent studies instead propose block-level training objectives, aiming to optimize EAL directly. However, these objectives remain surrogates rather than exact formulations of EAL. For example, some average later acceptance over target-distributed continuations rather than the verification branches selected by accepted tokens, introducing a covariance gap that can reverse the gradient direction. We analyze this gap and derive REAL, a real EAL training loss that uses path weights for unbiased estimation and Rao-Blackwellization for variance reduction. In a controlled loss-level comparison with a Qwen3.6-35B-A3B target and a Domino drafter, REAL outperforms WTV, VSD, DBLast, and the static LK baseline on probabilistic serving acceptance length across 9 task subsets and 5 evaluation seeds. These gains require only a change to the training loss, with no modification to the drafter architecture or inference procedure, demonstrating the practicality of real EAL training. A reference implementation of the REAL loss is available at https://github.com/King-ty/speculators/tree/real-loss.
Low-rank adaptation (LoRA) is the de facto method for parameter-efficient fine-tuning of large neural networks, achieving strong task adaptation with minimal trainable parameters and low optimization cost. Beyond this efficiency, practitioners have observed two striking properties: \textit{robustness to catastrophic forgetting} and the ability to \textit{merge independently trained adapters} into a single model that performs competitively across multiple tasks. These properties are central to continual and multi-task learning, yet remain poorly understood. In this work, we provide a theoretical explanation through a unified \textit{margin-based perspective}. We analyze LoRA in multiclass linear classification under a near-orthogonal task regime with $\ell_2$-regularization, and characterize the optimal LoRA adapter across regularization regimes. Our analysis shows that, in an intermediate regime of regularization parameters, the optimal adapter aligns with the \emph{max-margin solution} on the fine-tuning data. Building on this characterization, we derive two key consequences. First, we obtain closed-form expressions for the margins on pre-training and fine-tuning data, revealing a precise margin trade-off: the regularization parameter controls the balance between retention and adaptation. Second, we analyze adapter merging, proving that merged models achieve positive margin on each task and deriving optimal mixing coefficients that balance margins across tasks and maximize the margin over their union. These results lead to a simple, training-free merging rule, which we term \textit{margin-based LoRA merging} (MaLoRA). Experiments on modern architectures and real datasets validate our theoretical predictions, showing that MaLoRA matches or outperforms several adapter-merging baselines across a range of vision and language classification tasks.
Diffusion Language Models (DLMs) have emerged as effective alternatives to autoregressive models for latency-critical tasks, enabling faster token generation. Recent advances employ block diffusion with self-editing capabilities, where each block refines tokens iteratively before advancing to the next one. Yet each block is only as fast as its slowest token, and the next block cannot start before it. We suggest Blurring the Block Boundaries (B3), a novel method with which two blocks are refined in parallel within a shared forward pass. B3 calibrates itself on the fly, opening the next block only once the model is ready to start generating it. It is a training-free and hyperparameter-free acceleration method which seamlessly integrates within any self-editing block diffusion model. Evaluated on state-of-the-art DLMs across 10 diverse datasets spanning varied tasks and difficulty levels, B3 results in on par generation quality and 11.5-14.5% speedup compared to the base model, pushing the Pareto frontier for latency-quality trade-offs in DLM inference.
Energy-based models (EBMs) define a distribution through a scalar energy, $p_\theta(x)\propto\exp(-E_\theta(x))$, and are usually presented as one model family among many. This tutorial has a broader goal: beyond introducing EBMs, it uses them as a single lens that connects the different aspects of machine learning. Any EBM is specified by three choices, the energy, which determines what the model represents; the inference procedure, which samples from or optimizes the energy; and the learning method, which fits the energy to data. We first develop these components, including sampling, variational inference, and learning by maximum likelihood, noise contrastive estimation, pseudolikelihood, score matching, and adversarial training, and show that the learning methods share one contrastive structure. We then show that discriminative models, generative models such as GANs, normalizing flows, diffusion and autoregressive models, representation learning methods such as contrastive learning, word2vec, BERT, denoising auto-encoders, and diffusion representations, reinforcement learning with EBM policies and world models, inverse reinforcement learning, and planning, as well as inference-time computation with verifiers, are all EBMs that differ in these three choices, mostly in how they handle the partition function and sampling. This view explains existing methods, organizes them into a common framework, and suggests new algorithms by replacing one component while keeping the others, with composition of energies as a principled way to combine learned models at inference time.
Large language models (LLMs) increasingly exhibit capabilities that resemble human Learning by Thinking (LbT), the acquisition of new insight through internal processes such as explanation and reasoning, without new external input. A systematic account of the cognitive mechanisms behind these capabilities, and of their limits, remains underdeveloped. This paper proposes a Cognitive LbT Framework that maps four mechanisms (Reflective Reasoning, Synthetic Memory Formation, Internal Simulation, and Conceptual Self-Organization) onto computational processes in LLMs. Drawing on representative methods, including chain-of-thought prompting, self-consistency, self-reflection, Tree-of-Thoughts search, Self-Instruct, and self-taught reasoning, the analysis argues that selected methods can functionally approximate aspects of these mechanisms and can improve performance on some tasks without additional labeled examples, although this does not always constitute durable learning in model parameters. For example, Self-Instruct generated approximately 52,000 instructions from 175 seed tasks and reported a 33% absolute improvement on Super-NaturalInstructions (Y. Wang et al., 2023). The same mechanisms also produce characteristic failures, including hallucination diffusion, model collapse, concept drift, and high computational cost. These failures are not isolated technical defects but follow from four structural asymmetries between LLMs and human cognition: the identity of generator and monitor, closed learning loops without external grounding, symbolic rather than sensorimotor simulation, and the cost of inference-time computation. The framework yields testable propositions about when LbT in LLMs should succeed or fail and offers a structured basis for designing more reliable autonomous learning systems.
Many scientific and policy decisions require estimating the total number of target objects across a large image collection. Exhaustive human labeling is often infeasible, while raw model predictions can be biased, especially under domain shift. Recent human-AI estimation methods combine imperfect model predictions with limited human verification to obtain reliable collection-level estimates. However, existing work provides little guidance on how practitioners should choose among models, estimators, and adaptation strategies for a new deployment. We study these choices within a common human-in-the-loop counting framework across 12 tasks spanning biomedical microscopy, ground-level and aerial wildlife surveys, and overhead imagery of urban structures. Our results show that modern multimodal models provide a strong, training-free starting point for many tasks, while fine-tuned detectors become more advantageous at larger labeling budgets and across repeated deployments. Estimation efficiency improves further when model predictions are used to prioritize which units humans label, rather than sampling uniformly. We also find that each model's ranking quality can be estimated with a small pilot set, yielding practical guidance for selecting strategies on new counting tasks. In a field deployment on aerial pronghorn surveys, a task-specific adaptation of these methods meets the survey's $\pm 5\%$ tolerance after reviewing a median of only $3.2\%$ of collected images.
Start-ups and grade changes remain largely operator-driven, while developing and maintaining dynamic models for model-based control requires substantial engineering effort. Hybrid Offline Learning and Online Optimisation (HOFLON) uses one-step offline reinforcement learning to select actions from historical transition data. In its original implementation, the online objective combines the predicted return from a tree-ensemble critic with an adversarial-autoencoder (AAE) penalty for unfamiliar state-action pairs. We examine the effect of this penalty and replace the critic and AAE with a sparse variational Gaussian-process (sGP) critic. Actions maximise $\mu - \kappa\sigma$, an uncertainty-adjusted estimate of the return from one action followed by the historical operating practice. On the HOFLON polymerisation-reactor benchmark, removing the AAE from HOFLON led to 4 failed start-ups in 15 runs over five critics. At its published weight, the variation in the AAE penalty is four orders of magnitude smaller than the variation in the tree-ensemble critic. Nevertheless, it changes actions when the predicted returns of competing candidates are nearly equal, and paired runs with and without the penalty diverged from the first move. With a Mat\'ern-5/2 sGP critic, $\kappa = 0$ led to failure in 13 of 15 runs, whereas with $\kappa = 3$ none of 15 runs (five critics, three noise realisations) failed. The median and lowest returns were 13,464 and 13,398 (HOFLON: 13,393 and 12,411). A squared-exponential sGP critic with learned inducing inputs needed $\kappa \geq 3$ to remain within the empirical data support in steady state, and then settled with a temperature offset of 2.0 K.
Pretrained video diffusion models (VDMs) offer a powerful foundation for World Action Models (WAMs), which jointly predict robot actions and future observations. Building such models requires injecting low-dimensional robot signals (e.g. proprioception, actions, and values) into a latent space shaped by natural videos, and a key design choice is the interface through which these signals enter: existing schemes treat it as a formatting detail, spending a separate latent frame on each signal or bolting on dedicated encoders and decoders. We argue that signal injection is instead a distribution-alignment problem: how well a representation matches the prior of the pretrained VDMs during denoising. Using the relative reconstruction error of a frozen VDM denoiser as a probe, we find that conventional spatial encodings sit far off this prior, whereas frequency-domain packing yields latents that the denoiser corrects far less. Motivated by this diagnostic result, we propose MetaWAM, which folds all robot signals into a single unified meta-frame: each signal is assigned to a disjoint frequency-domain channel band and lifted to the spatial latent by an inverse DCT, yielding a decodable interface aligned with the pretrained VDM prior without any added module or architectural change. A fixed band layout together with slot-aware routing lets the same meta-frame support policy, world-model, and value prediction over a shared set of robot variables. On RoboCasa, MetaWAM reaches 68.9% success while collapsing the per-signal frames into one and cutting inference latency by 39%, and attains 85.0% success on three real-world Franka tasks. Code and models are available at https://github.com/showlab/MetaWAM
Research agents achieve increasingly strong benchmark scores, yet these scores do not necessarily translate into reliable delivery in ongoing research projects. Many research-agent benchmarks emphasize standalone tasks and relatively self-contained workflows in clean or curated workspaces. However, researchers ask agents to complete one step within an ongoing project, where requests depend on earlier session decisions and workspaces contain unfinished files, outdated versions, and failed attempts. We introduce ARE-bench to evaluate whether research agents advance ongoing projects while respecting prior decisions and existing workspace state, grounded in real session context and in-progress workspaces. ARE-bench includes 75 tasks built on agent sessions and full workspaces from recruited researchers' own projects, spanning seven research stages, 33 requirements, and 150 sub-requirements, with 1,173 binary checklist items in total. Each task is anchored at a real user request, preserves its requirements, and summarizes the earlier requests still in effect as background. The complete workspace is replayed to that boundary without cleanup, preserving artifacts from earlier turns. Each checklist is compiled from an intent-aligned rubric and credits any valid solution. Across seven frontier models, the strongest passes every checklist item on only 66.2% of tasks, and Design and Execution rank among the hardest stages for every model. Three models pass about 69% of tasks when only delivery contracts and general norms are checked, but just 28.4% to 40.0% once historical and material or workspace requirements are also checked. Evaluating delivery without these contextual requirements can therefore overestimate research agents and mask differences between them.
Time series foundation models (TSFMs) forecast unseen series without task-specific training. Increasingly they are built on recurrent backbones, whose constant per-step cost makes long-context inference cheap, and which, unlike transformers, carry a state. A recurrent backbone computes its prediction from two inputs: the observed context window and the initial recurrent state. In practice only the first is utilised and the initial state is unused by convention. We ask whether the initial state can be utilised in recurrent TSFMs. Using the xLSTM-based zero-shot forecaster TiRex as a case study, we freeze the backbone and train only a small head that writes a context vector into the initial state, so that any change in accuracy is attributable to the injected state rather than to adapted weights. A spectral summary of the input window, learned once and applied unchanged to every series, lowers mean GiftEval MASE by 0.175%, with every seed improving, which neither a parameter-matched LoRA adapter nor a learnable static state reproduces. An embedding of text attached to each series, evaluated on six MoTime datasets, reveals two roles for the state. As a per-dataset adaptation surface, a static state lowers MASE by 0.04-6.96% of the zero-shot baseline (8-70% of the gain of full fine-tuning). As a channel for exogenous information, the text-conditioned state lowers it by a further 0.15-0.41% of the zero-shot baseline on four of six datasets and, combined with a LoRA adapter, exceeds full fine-tuning, which has no access to the text, on one of them. In TiRex, the effect concentrates in the mLSTM blocks and is strongest at short context. The results are preliminary and the improvements small, but they point to the initial state as a direction for further research.
Rapidly evolving LLM architectures and heterogeneous mobile hardware make supporting new models in on-device inference engines a recurring and costly system challenge. Each model may require coordinated changes across conversion, graph construction, operators, and hardware backends. Existing coding agents often miss these cross-layer dependencies and struggle to diagnose failures across host--device execution paths. We present EngineAgent, an agentic system for automating on-device LLM model enablement. Its key insight is to separate reusable engine and hardware knowledge from model-specific implementation. Before enablement, EngineAgent constructs a hierarchy-aware engine index that captures engine abstraction layers and cross-layer dependencies, and pre-validates backend-specific build, execution, and debugging workflows using already supported models. For each new model, a builder uses this knowledge to coordinate cross-layer changes, while an auditor checks implementation completeness and drives iterative debugging through device-side validation. Across 12 tasks spanning three mobile LLMs and four engine--backend pairs on CPUs, GPUs, and NPUs, EngineAgent achieves 75% end-to-end success with either of two underlying LLMs, versus 33.3% for the strongest coding-agent baseline. Success requires compilation, on-device execution, and output-logit agreement with reference implementations. EngineAgent also reduces average per-task time and model-API cost relative to baselines using the same underlying LLMs.
Retrieval-augmented systems often retrieve passages that are accurate and relevant but state the answer at a coarser level of granularity than the question requires. For example, a question may ask which specific disease a drug treats (familial hyperlipidemia), while the retrieved passage mentions the broader disease family (hyperlipidemia). Prior work on context sufficiency treats such coarse passages simply as insufficient. We argue that they form a distinct case, which we call resolvable insufficiency: text alone cannot settle the fine-grained answer, but the broader category it mentions maps to a short list of candidate answers in a domain ontology. When language models read these coarse, family-only passages, they exhibit a phenomenon we call Clade (Category-Level Answering from Deficient Evidence): they answer instead of abstaining or searching further. On biomedical questions from PrimeKG, family-only passages reduce abstention by 13.8 points compared to uninformative passages, and the resulting guesses are wrong about half the time. A counterfactual test confirms the cause: replacing an exact disease name with its broader family name lowers abstention by 17.4 points compared to deleting the sentence. Simple fixes like confidence filtering, scaling to a larger model, and explicitly instructing the model to abstain cannot resolve Clade, and this replicates on ecological species interactions in GloBI. Finally, we test whether an ontology can turn coarse passages into exact answers. Given the correct family, an ontology list includes the true answer in 98.5% of cases, but string matching finds that family on the page for only 19.8%. Letting the model choose the family raises this to 30.0% and end-to-end recovery from 10.0% to 26.8%. Overall, our results show when language models should stop guessing and consult an ontology, and highlight which steps of that process remain unsolved.
Users often complain that AI assistants such as ChatGPT are getting worse. Some versions do change, but human evaluation leaves a second possibility untested: the people judging may change. We separate the two on Chatbot Arena, a public platform where people compare answers from two anonymous large language models (LLMs) and pick the better one or say both are bad. We study 1,260,967 blind votes cast from May 2023 to August 2024 on pairs of models whose versions are fixed: dated vendor snapshots and models with fixed open weights. For the same pair of unchanged models, the share of votes saying "both are bad" rises by 1.71 percentage points a month on a mean of 17.1%, and by 1.91 points a month in 2024. In fourteen pairs observed often in both 2023 and 2024, it rose in every one, from 15.5% to 23.3%. The ranking does not trend: in 2024, the stronger model's share of decisive votes is flat. The rise holds within the same voter and on the same pair, and across arrival cohorts from their first recorded vote. It holds with prompt controls, in every major language, for dated vendor snapshots, open-weight models and models with stable answer length, and between interface changes. The drift is shared across models: month effects estimated from other models' pairs remove it from each model's own pairs. Platform-wide, the average model would be expected to score 78% against the 2023 average, yet the share of "both are bad" votes ended near where it began. The pattern fits a standard that follows the category's progress, so a product that stands still is judged as declining. Absolute verdicts on AI drifted while relative comparisons held steady, an evaluation drift that matters for LLM leaderboards and for the human feedback used to train models.
Stochastic diffusion samplers often require substantially more neural function evaluations (NFEs) than deterministic samplers to achieve comparable sample quality. Although higher-order discretization accelerates deterministic sampling, its benefits for stochastic sampling remains unclear. We establish matching upper and worst-case lower bounds showing that stochastic samplers with higher-order approximations remain first-order convergent, even with exact score functions. Thus, increasing the approximation order does not generally improve the convergence order. Motivated by this limitation, we introduce forward-value discretization, a training-free framework that evaluates the data/noise prediction model at a lookahead estimate constructed from previous predictions. For deterministic sampling, the resulting Forward DPMSolver (F-DPMSolver), improves sample quality at low NFE budgets through its DDIM-based variant, despite remaining first-order convergent. For stochastic sampling, we develop Forward DDPM (F-DDPM), which evaluates the model at a cleaner lookahead point adjusted for subsequent noise injection. This estimate reuses both the previous model prediction and the previously injected noise. Both variants admit first-order convergence guarantees with one model evaluation per iteration. Across CIFAR-10, ImageNet, FFHQ, and LSUN, F-DDPM consistently achieves lower FID than the evaluated stochastic baselines at matched NFE budgets. With 1.5 times as many NFEs as deterministic baselines, it achieves competitive sample quality, narrowing the practical efficiency gap between stochastic and deterministic sampling. Code is available at https://github.com/Na-Li66/F-DDPM.
LLM agents increasingly operate in long-horizon, verifier-rich environments, but current self-improvement mechanisms often write episode-derived lessons directly into memory, making local success an unreliable signal for reusable and safe knowledge. This paper addresses the core question of when an experience-derived behavioral update should be admitted into persistent agent state. We propose Falsification-Gated Self-Evolution (FGSE), a verifier-grounded framework that represents each candidate lesson as a structured hypothesis with an explicit precondition, behavior change, expected effect, and verifier. FGSE applies the hypothesis only in a temporary state, tests it on target-transfer, falsification, and archived-regression probes, and then commits, refines, or rejects it according to measured gain and risk. Across web, app, tool-use, and long-term memory benchmarks, FGSE achieves strong task performance, including 39.8 WebArena success rate, 78.8 AppWorld average completion, and 0.724/0.487 Pass$^1$ on $\tau$-Retail/$\tau$-Airline, while reducing falsification failures to 5.0%, regression damage to 1.3%, and harmful committed updates to 3.6%. These results suggest that testing memory updates as hypotheses supports more reliable long-term adaptation.
Text encoders turn earnings calls into vectors for search, for language models that answer from retrieved passages, and for forecasting models. A live retriever can use an encoder trained once on all available text, as finance retrievers are, but a backtest needs an encoder that read no text written after the date of each vector. To address this look-ahead bias, from 412,248 calls we train vintages under four pretraining recipes, with yearly cutoffs from 2010 to 2024, and each vintage reads only text dated on or before its cutoff. We then fine-tune each vintage on the question-answer pairs of the calls themselves. The vintages we train on calls alone, with 33.8M parameters, retrieve answers on held-out calls better than the general retrievers we tested in every test quarter. Their vectors also raise by about half the rank correlation between next-day returns and forecasts built from past returns and firm size. When each vintage continues the previous one, a later vintage retrieves better on earlier quarters than the vintage available at the time, and a point-in-time language model that answers from the passages the later vintage retrieves inherits the gain. The 2024 vintage of our families also forecasts earnings surprises better, although the interval of this gain includes zero on calls that no vintage read. It forecasts returns no better, so point-in-time vintages matter most for retrieval and little for forecasts of returns.
Language-model systems use self-judgment to decide which answers to release, review, or recompute. Because intervention is costly, an assurance score must rank errors well at the available budget, not merely over the full ranking. We compare the judge's continuous verdict-token score with a supervised correctness probe trained on attention-head activations from the same prompt. Exact identities link fixed-budget error recall to residual risk and show that AUROC averages recall over all budgets. Across three open 7-12B models on GSM8K and MATH, the probe finds 1.3-22.5 additional percentage points of errors at 5% review in all six model-dataset settings (median 7.4 points). At 10% routing, probe-guided allocation improves accuracy over equal-cost random allocation in both GSM8K experiments, but beats judge-guided routing only for the weaker judge. Cross-dataset transfer lowers probe AUROC for every model in both directions relative to target-domain training. Because dataset, weight precision, and generation budget change together, this measures a joint protocol shift rather than a dataset-only effect. Internal signals can improve supervised, in-domain budget allocation, but are not domain-general confidence estimators.