LLM 理论进展

Converting a pretrained autoregressive (AR) model to a diffusion language model (dLLM) enables parallel generation without pretraining a new model. Published conversion methods differ by roughly three orders of magnitude in training data and have not been compared under a common protocol. We compare two conversions of the same 30B Mixture-of-Experts (MoE) parent, holding the corpus, supervised-token budget, trainable parameter set and evaluation harness fixed, each under its own training recipe. The in-place model updates a subset of the parent's weights using denoising and representation-alignment losses; the frozen-tower model instead conditions through cross-attention on a frozen causal copy of the parent. With 1B training tokens, the frozen-tower model scores 71.60 on HumanEval pass@10 against 6.19 for the in-place model, an 11.6x improvement. At the same budget it also keeps 95% of the parent's GSM8K score and 99% of its MMLU-Pro score. A dense-parent experiment reproduces the HumanEval separation. Within the two-tower design at about 500M tokens, freezing the context tower retains substantially more MMLU-Pro performance than training it, while both give similar observed HumanEval scores. Our theoretical analysis establishes that both conversion classes contain an exact sampler for the AR parent under a hard attention mask and left-to-right commitment of one position per round. Under a shared loss, freezing removes the gradient contribution through the context states. Furthermore, evaluation protocol substantially affects a published 500B-token conversion's scores in both directions across tasks, while its AR parent's scores vary by less than three points, so comparing dLLMs needs a common protocol. These results show that, in the tested low-budget regime, the frozen-tower configuration retains substantially more of the parent's generation performance than in-place conversion.

Multimodal Large Language Models (MLLMs) achieve strong performance across vision-language tasks, yet the internal mechanisms by which visual and textual information are fused across layers remain insufficiently understood. We investigate representative MLLMs from two architectural paradigms: concatenation architectures and native multimodal architectures. We conduct three progressively connected analyses: alignment decoupling identifies which modality changes, attention routing and entropy characterize how cross-modal information is distributed, and intrinsic dimensionality examines how fusion reshapes feature spaces. Separately, we perform causal intervention experiments as a validation of the resulting interpretation. As a supplementary analysis, we use visual CKA to examine the Platonic Representation Hypothesis. Together, these analyses reveal two distinct fusion pathways: concatenation models follow a text-first, vision-later pathway, whereas native models exhibit earlier visual-textual co-adaptation and feature-space reorganization. This work provides a mechanistic perspective for understanding multimodal fusion and supports architecture-aware diagnostics of multimodal representations.

Ensuring the safe and reliable deployment of large language models (LLMs) remains a fundamental challenge. Existing safety alignment approaches either incur high computational cost or unintentionally disrupt the model's core knowledge, leading to degraded fluency and factual accuracy on benign tasks. This reveals a persistent trade-off between safety and utility. We propose DNAlign, a lightweight alignment framework that integrates control-theoretic optimization with null-space projection. By treating the LLM as a dynamic system, the proposed framework introduces controllable perturbations to steer generation toward safe behavior. A key component is the projection module, which restricts these perturbations to the harmful-related subspace derived from neutral hidden states, thereby preserving general knowledge and response quality. A value function trained on human preference data adaptively optimizes the control signals to align with human safety preferences. Extensive evaluations across multiple LLM backbones demonstrate that our framework consistently reduces harmful outputs while maintaining fluency, coherence, and factual utility. It achieves superior overall performance compared to prior alignment baselines without sacrificing generation diversity. These results indicate that the proposed framework provides an effective and practically deployable solution for safe LLM alignment. Code is available at https://anonymous.4open.science/r/DNAlign.

A targeted adversarial perturbation can drive a vision-language model's (VLM's) teacher-forced training loss for a fixed target caption to near zero, yet the same model, allowed to generate freely, produces the original, correct description with no trace of the target. We call this dissociation the train/inference gap, and give it a precise mechanistic account on Qwen2.5-VL-7B-Instruct using a controlled two-stage PGD attack on 200 held-out COCO images. First, we show that image-level pixel statistics, including a correctly re-implemented, texture-based attackability measure from the CNN robustness literature, have essentially no predictive power over which images are corrupted (best predictor r=-0.050, p=0.484; ridge regression R^2=0.069). Second, using the logit lens, we localise the gap to a single autoregressive step: the rank of the target token, conditioned on the correct first token already being generated, is fixed at exactly 3,488 out of 152,064 vocabulary entries for every image and every condition, with zero variance. Third, tracking target-token rank across all 28 LLM decoder layers reveals that the visual encoder corrupts every image's representation by a comparable margin regardless of eventual outcome, but the language model decoder then differentially arbitrates: amplifying the corrupted signal for susceptible images and actively suppressing it, past its clean-image baseline, for resistant ones (p<0.001, rank-biserial r=0.579). A linear probe on the merger hidden state separates these two outcomes with AUC=0.858, though we flag a circularity concern in this estimate. Together these results argue that adversarial robustness in autoregressive VLMs is substantially a property of the language decoder's prior, not the visual encoder, with direct implications for where faithfulness evaluations and defenses for deployed VLM systems should be targeted.

Designing effective heuristics for diverse combinatorial optimization problems requires substantial expertise and repeated search. Large language models (LLMs) automate heuristic generation and refinement, but heuristic search typically depends on evaluation feedback from the problem being optimized. Generalizing to new problem definitions using only source-task feedback therefore remains a central challenge. We introduce MECo, an LLM-driven multi-task evolutionary framework for zero-shot cross-problem generalization. MECo maintains task-conditioned heuristic populations and uses a transfer gap based on cross-task population performance to guide their interactions. These interactions enable the transfer and recombination of heuristics. A complementary selection criterion then constructs a compact heuristic set by rewarding each member's additional coverage of source combinations. The selected set is applied to target problems without further search or adaptation. Experiments on 32 problem variants across vehicle routing (VRP) and flexible job-shop scheduling (FJSP) show that MECo achieves the lowest mean costs compared with eight automated heuristic design (AHD) baselines under the same budgets. On out-of-domain problems, it outperforms the strongest baseline in each family. Moreover, integrating the framework of MECo with different AHD methods improves their ID and OOD performance in both families, supporting its effectiveness across different methods.

Vision language models (VLMs) incur substantial inference cost because every visual token is processed by the attention and MLP projections of every decoder layer, even when token-specific visual computation is unnecessary at many depths. We introduce Patch-to-Prune (P2P), inspired by Mechanistic Interpretability, a training-free framework that converts activation patching from a diagnostic tool into an inference-time computation bypass. P2P performs validation-guided forward and backward layer sweeps to identify decoder regions whose visual-token projection outputs can be replaced by fixed neutral proxy activation vectors within a user-specified accuracy tolerance. Unlike conventional token-pruning methods, P2P preserves the sequence length, token order, positional information, attention mask, and residual pathways, thereby pruning computation without removing tokens or modifying the pretrained model weights. We evaluate P2P on four VLMs from the Qwen2.5-VL and LLaVA families across seven multi-modal benchmarks using mutually disjoint calibration, validation, and test partitions. P2P at a 3% tolerance retains around 94% of dense accuracy while reducing FLOPs by 55%. Beyond these efficiency gains, our layer-wise analysis suggests that visual processing in VLMs is non-uniformly distributed across decoder depth: early and late layers often require little token-specific visual computation, whereas intermediate layers appear to perform most task-relevant visual integration, enabling later reasoning to rely largely on visual information already embedded in shared residual and textual representations. This makes P2P both an efficient inference framework and a causal lens into visual information processing in VLMs.

During post-training of large language models (LLMs) with Reinforcement Learning with Verifiable Rewards (RLVR), GRPO-style algorithms can exhibit severe late-stage collapse. Prompt-based probing reveals that this is not benign strategic pruning, but a harmful contraction of effective strategy capacity that makes distinct reasoning strategies increasingly inaccessible. To characterize this phenomenon, we define strategies through trajectory-level policy-update interactions and develop a unified theoretical framework combining optimization dynamics and information theory. We prove that major RLVR objectives progressively concentrate probability mass onto a single strategy, while sustaining nontrivial task accuracy requires a minimum strategy capacity. The conflict between these two results provides a mechanistic explanation for catastrophic collapse. We further derive the {Mirrored Entanglement Index (MEI)} as a lightweight online warning signal. To prevent collapse, we propose Mesh Learning, which exposes multiple reasoning strategies and prevents any single strategy from dominating optimization. Across AIME26, AIME25, MATH-500, GPQA, and LiveCodeBench, Mesh Learning consistently outperforms strong baselines across Qwen and Phi model families, with gains of up to 13.4 pp and 11.5 pp, respectively. These results establish strategy preservation as a key principle for stable RLVR. Code is available at https://github.com/Ayanami-0123/Open-Mesh-Learning.

Large language models (LLMs) can support rubric-based screening of CVs, but applying a high-capability model to every candidate and criterion is costly. We present CreateScore, a domain-theory-informed Bayesian network for criterion-level LLM routing. A hand-specified directed acyclic graph with Dirichlet-multinomial conditional probability tables converts CV evidence into posterior uncertainty; low-uncertainty decisions are resolved by a local 8B model and uncertain ones are escalated to a 120B reference model. The graph is causally motivated, but the system performs standard Bayesian conditioning, not causal inference. The escalation threshold is calibrated on a training fold (target: 70% resolved locally) and then frozen. On 200 synthetic Data Science CVs (139 training and 61 test candidates, five criteria), 77.7% of criterion decisions were resolved locally (237 of 305). Relative to a reference condition in which the 120B model adjudicated every criterion, routed escalation reduced token use by 65.2% and raised exact score agreement from 32.8% (8B alone) to 42.6% (95% CI 31.0-55.1%); at n = 61 the gain was not statistically distinguishable. The uncertainty signal did not, however, identify the decisions on which the 8B model erred: disagreement with the reference was 16.2% among escalated and 19.4% among locally resolved decisions (AUROC 0.47, 95% CI 0.39-0.56), no better than random selection. We also document how an earlier evaluation was invalidated when truncated reasoning-model outputs were silently replaced by local labels, and we recommend safeguards for cascade evaluation. CreateScore is supported as an auditable cost-reduction mechanism, not yet as a targeted error detector, and is not an autonomous hiring system.

Large Language Models (LLMs) increasingly support Uncrewed Aerial Vehicle (UAV) swarm operations such as data collection scheduling, where the model reads structured sensor reports and decides which sensors to visit. An adversary who quietly manipulates those reports can redirect the swarm without modifying the model weights or the UAV. Defenses for this interface have been proposed architecturally but rarely implemented or evaluated. We implement and evaluate defense-in-depth at the perception-reasoning interface of LLM-Centric Agentic UAV Swarms. Five layers check the provenance of a report, whether its values are physically admissible, whether they agree with what swarm geometry and service history predict, whether the resulting schedule starves any sensor, and, when these fail, hand control to a deterministic scheduler that ignores the suspect input. We test each layer against an adversary strong enough to defeat the layer before it. For each of the three input-side layers, we derive in closed form how far a report can be distorted before that layer reacts, fixing each boundary from deployment parameters before any attack data is collected; across thirty matched simulation runs, predicted and measured boundaries agree. Separating attack detection from response is a well-established principle, and we quantify the cost of neglecting this distinction at the perception-reasoning interface. When the system rejects a report, it replaces it with the most recent accepted report. This prevents the adversary from controlling the UAV schedule, but it also increases cumulative cost by 79% and 74% for the two detectors, respectively, compared with the undefended system. The safety check does not detect any attacks, but it nevertheless reduces the attack-induced cost by 37.5%.

Chain-of-thought (CoT) reasoning allows humans to inspect how large language models reach their answers, and oversee model behaviour. This reasoning comes at an increased inference cost, motivating efficient methods that train models to solve tasks using fewer tokens. However, a common concern is that such training may cause models to skip important reasoning steps, so the CoT no longer faithfully reflects the model's decision. It is unclear whether or when this occurs in practice, since different efficiency methods apply length pressure to models' CoT in distinct ways, and faithfully explaining a model's decision takes more tokens on some tasks than others. To understand these dynamics, we fine-tune a variety of models with three methods that apply length pressure differently, namely a fixed generation budget, a per-example length target, and a group-relative length reward. We evaluate how efficient reasoning affects CoT faithfulness (i.e., how well the CoT reflects model decisions on related inputs) and monitorability (i.e., whether the CoT reveals when input interventions alter the output). We find that it affects faithfulness and monitorability differently. Faithfulness falls in most settings, primarily because the trained models are less consistent. Monitorability is more robust, as models keep acknowledging the influence on their answer even when the CoT is much shorter.

Large language models (LLMs) contain broad knowledge, but they cannot access all of it reliably. We study this problem through knowledge accessibility, which describes whether the knowledge needed for a query can be recalled from the model. We find that knowledge accessibility has a simple geometric structure in the model's representation of the query alone, before any generation. More accessible queries are closer to a center in the representation space, while less accessible queries are farther away. This geometry reveals a knowledge boundary that separates more accessible queries from less accessible ones. Accessibility consistently decreases with distance from the center, and this distance-based ordering transfers across datasets even when the centers differ. Controlled experiments further show that the centered geometry is more closely related to knowledge accessibility than to reasoning difficulty. The geometry also reveals when different interventions are useful. Query rewriting helps more for accessible queries, chain-of-thought reasoning helps more near the boundary, and retrieval gives larger gains beyond the boundary. These findings not only provide a new geometric view of how knowledge is organized in language models, but also suggest a useful pre-generation signal for adaptive inference.

Reliable large language models (LLMs) must not only produce accurate answers but also express confidence that faithfully reflects their probability of being correct. Such calibration is essential for identifying uncertain predictions and supporting reliable decision-making in real-world deployment. Recent studies incorporate calibration learning into reinforcement learning (RL), jointly optimizing answer correctness and verbalized confidence using ground-truth correctness supervision. However, their reliance on labeled data limits their applicability in practical test-time settings, where ground-truth labels are unavailable and calibration may need to adapt to newly encountered target tasks. To address this challenge, we propose Test-Time Calibration Learning (TTCL), a label-free framework that jointly adapts reasoning accuracy and verbalized confidence directly on unlabeled target-task data. Specifically, TTCL derives self-supervision signals for both correctness and calibration from multiple model-generated responses, enabling calibration learning at test time without ground-truth labels. Theoretical analysis further establishes TTCL as a bounded surrogate for the ideal calibration objective. Extensive experiments on mathematical reasoning and factual question answering demonstrate that TTCL consistently improves both accuracy and calibration across diverse models and tasks. On base models, TTCL achieves an average relative accuracy improvement of +40.13% and an ECE reduction of +70.80% across eight benchmarks. Moreover, TTCL can further improve both accuracy and calibration for already calibrated models under domain shift, particularly when source-domain calibration transfers poorly to target tasks. In the math-to-factQA setting, TTCL achieves an average relative accuracy gain of +20.35% and reduces ECE by +53.83%. The source code is released at https://github.com/tmlr-group/TTCL.

Recent years have seen the employment of a plethora of machine learning (ML) models in high-stakes domains, but they remain largely opaque to the practitioners who act on their predictions. While post-hoc explanation methods offer a lens into this model behavior, wielding them effectively demands expertise most domain experts lack: navigating high-dimensional outputs, selecting the best explanations, and synthesizing evidence across disparate tools. To this end, we present MEA, a multi-agent framework that removes the explanation knowledge barrier entirely: a Proposer agent selects and configures explanation tools based on the question and modality, while an Actor agent is optimized end-to-end against faithfulness, transforming the outputs into natural language explanations grounded in model behavior across tabular, text, and vision modalities. Further, we introduce diverse question types spanning feature attribution, counterfactual reasoning, and spurious feature detection, each paired with a perturbation-based faithfulness metric. We find that frontier LLMs systematically produce unfaithful explanations. By optimizing against faithfulness rewards augmented with a modality-adaptive penalty, MEA consistently outperforms post hoc explainers, agentic, and closed-source baselines across six datasets, with reward-driven optimization yielding faithfulness gains of +28% (tabular), +21% (text), and +34% (vision) over the untrained backbone. More broadly, our findings suggest that AI agents themselves can serve as a scalable, adaptable interface to ML explainability, opening a path toward natural-language explainability that generalizes beyond the fixed, single-purpose tools that have long defined the field.

Personalized large language models often require a complete adaptation state for each user. However, this paradigm scales poorly as the user population grows. We revisit this design through the lens of personalization capacity allocation: how much adaptation capacity can be shared across users, how the shared capacity should be composed, and how much must remain user-specific. We answer them through three complementary empirical analyses. We find that independent user adapters contain substantial cross-user reusable structure, that the utility of reusable directions reflects both user relevance and variation across queries, and that user histories provide transferable signals for compact individual correction. Motivated by these findings, we propose LINEUP. It learns a bank of reusable low-rank personalization factors, composes them through user-conditioned recall and query-dependent calibration, and restricts target-user adaptation to a tiny user code over a shared correction space. This design decouples expressive personalization capacity from per-user trainable state. Each target user optimizes only eight scalars, while all shared components remain fixed. By comparison, the evaluated private-LoRA configuration uses 4.19 million per-user parameters. Our theoretical analysis gives a finite-step, finite-history risk bound and sufficient conditions for user-code refinement to improve on history initialization. Across six tasks spanning personalized classification, prediction, and generation, LINEUP leads on all 12 metrics, each averaged over three independent runs (e.g., reducing LaMP-3 RMSE by 11.4% relative to the strongest baseline). It maintains advantages under limited history. These results show that rich personalization can be supported primarily by reusable, conditionally composed shared capacity, while independent user adaptation remains confined to a tiny correction state.

On-policy distillation (OPD) trains a student on the responses it generates. Existing LLM multi-teacher OPD transfers what specialists predict through their output distributions. We introduce Latent-MOPD, to our knowledge the first representation-level multi-teacher OPD method for LLMs. It integrates existing specialists through both their predictions and the hidden states used to compute them, without additional teacher training. To coordinate representation supervision from multiple specialists, we select late-layer targets according to the teacher-student relationship, bridge unequal hidden widths with a shared projection, and group updates by domain. Each teacher's supervision gradually shifts from hidden states to token predictions, with both channels using the same routed specialist. In our main same-family setting, Latent-MOPD outperforms the token-only, representation-only and uniform-averaging baselines on all nine benchmarks across math, code and logic. With the same parameter count as each teacher, the student also surpasses the per-benchmark best teacher on a majority of these benchmarks. With larger, separately developed cross-family teachers, Latent-MOPD outperforms both single-channel baselines on all benchmarks. A same-family all-layer representation-only control remains stable with domain-pure updates but collapses when teacher domains are interleaved within an update. Our results show that a single student can integrate capabilities from several specialists through both their output distributions and internal representations.

Increasing the batch size during training is a common practice in large language model (LLM) pretraining, yet the theoretical justification behind its success is not well understood. Analyses of stochastic optimization often assume uniformly bounded stochastic gradient variance, yet recent evidence suggests that this assumption fails in many practical nonconvex problems. The Blum--Gladyshev (BG-$0$) noise model relaxes this assumption by allowing the variance to grow quadratically with the distance from initialization, suggesting that batch size schedulers can help by controlling the variance growth during training. However, this growth can be overly conservative in practice. We empirically investigate variance growth in LLM pretraining and observe that a generalized BG model with a tunable growth exponent provides a tighter description of practical noise behavior. Motivated by this observation, we introduce the generalized BG-$a$ noise model, which interpolates between bounded variance ($a=0$) and BG-$0$ noise ($a=2$). Under $L$-smoothness, we derive an information-theoretic lower bound with growth-dependent oracle complexity $Ω(ε^{-(4+a)})$ and establish a matching upper bound in $ε$-dependence by increasing the batch size as the iterates move away from initialization. Finally, we propose an adaptive batch scheduler that controls variance growth through dynamic batch size adjustments during training. In pretraining OLMo2 models of up to 1B parameters on C4, our scheduler achieves a lower validation loss than both small and large batch training under matched token budgets, while using less than 10% of the iterations of small batch training.

When a large language model solves a mathematical problem, its reasoning is largely hierarchical, and the solution often branches at a few tokens where the next-token entropy is high. Such tree-like structure embeds in hyperbolic space with far lower distortion than in Euclidean space. Activation steering, however, usually edits the hidden states of a pretrained model by adding one fixed Euclidean vector at every token, even though most tokens of a solution are already determined by the context. We propose Hyperbolic Entropy Steering (HEST), which embeds the hidden states in the Poincaré ball with a lightweight probe whose only label is the model's own next-token entropy. Where this entropy exceeds a threshold, HEST moves the embedded state along the geodesic of steepest descent of a readout of the probe and maps the change back to the hidden state. For the Busemann readout of a learned ideal point, we prove that a step of fixed length lowers it by the same amount at every state. On three instruction-tuned models from the Qwen2.5-Math and Llama-3.1 families, HEST with the Busemann readout improves greedy accuracy on MATH-500 and GSM8K in five of six settings, by up to 1.8 points, whereas a contrastive steering vector added at every token lowers accuracy. With a Euclidean probe trained in the same way, this gain disappears on Qwen2.5-Math-1.5B-Instruct. The gains are largest on problems where the model hesitates often, and accuracy on the remaining problems is almost unchanged.

Time and event expression extraction are fundamental temporal reasoning tasks, but the problem remains difficult due to annotation ambiguity, domain sensitivity, and unstable model behavior. Existing evaluations focus on in-domain performance, offering limited insight into reliability under distribution shifts. We evaluate multiple model configurations across families, architectures, and reasoning strategies over four dimensions of generalization, examining transfer from base performance, cross-dimensional correlations, and the effects of scale, architecture, and prompting. This provides a systematic study of how prompted LLMs generalize in time and event expression extraction tasks. We find that strong base-task performance generally predicts better generalization. However, this relationship weakens under substantial distribution shifts. Inductive prompting performs most consistently across domain shift, adversarial perturbations, compositionality, and length increase, while gains from scale, architecture, and deductive and abductive prompting strategies are uneven and dimension-specific. We conclude that LLM generalization in temporal extraction tasks cannot be predicted from any single dimension alone and cannot be reliably inferred from in-domain or single-dimension evaluations, highlighting the need for reasoning strategies that generalize across dimensions.

Vision-language models (VLMs) have emerged as powerful candidates for universal vision backbones, with representative architectures including autoregressive (AR) models and diffusion transformers (DiTs). Yet, adapting them efficiently for all-in-one low-level image restoration remains a challenge. Crucially, the field lacks an understanding of how VLMs organize hidden-layer representations and whether these structurally distinct paradigms share a common geometric organization for pixel-level perception. Such shared organization is a prerequisite for building highly transferable, unified restoration VLMs and adapters. In this paper, we systematically investigate representational similarity across 24 low-level tasks spanning 5 categories. We propose GeoSim, a unified four-level framework that analyzes task-conditioned representations from global similarity, local geometry, sparse feature decomposition, and topological verification perspectives. Our formulation applies to the analysis of hidden states in AR models and feature maps in DiTs across same- and cross-task/model settings. Our results reveal the organizing principles of low-level visual representations while exposing their limits in cross-task and cross-model agreement. Ultimately, GeoSim provides an interpretability lens for probing latent transferability in low-level vision and diagnosing model limitations in task- or model-specific scenarios.

Multi-token prediction (MTP) improves the throughput of autoregressive generation by enabling the language model to draft multiple next tokens per forward pass, while a verification step over draft tokens ensures that token distribution of the backbone is preserved. Every open MTP-family release (MiMo-7B, DeepSeek-V3, Qwen3) trains its heads jointly with the backbone over the full pretraining run of tens of trillions of tokens, thus setting the drafter quality at pretraining time. We ask whether a lightweight post-training pass on target-generated chain-of-thought is enough to reach the same expected throughput speedup on a frozen reasoning model, and study how a serving-time system built on such a checkpoint can be optimized. We present three findings. 1) On a frozen Qwen3-8B with $K{=}3$ chained MTP heads, we show that a post-training recipe with plain cross-entropy on $\approx\!2.5$B tokens reaches or exceeds the expected speedup of jointly trained MiMo-7B on math, coding and knowledge benchmarks. Our post-training recipe utilizes $10^3$-$10^4\times$ less MTP-training tokens as compared with joint pre-training of MiMO-7B MTP baseline. 2) We propose a chain-aware relaxation of draft token verification rule that allows a bounded drift from backbone language model token distribution. We show that this relaxation lifts expected speedups by $+12$ to $+16%$ per benchmark while preserving task accuracy. 3) We propose an adaptive controller that dynamically chooses the number of MTP heads to be engaged at inference time and demonstrate recovery of upto $11$--$14%$ loss in speedup using fixed maximum MTP draft length.

Biological reasoning models use post-training to connect LLMs to biological foundation model representations and biological text. Their benchmark accuracy is taken as evidence that LLMs reason over these inputs. We test this assumption in six biological reasoning models across DNA, protein, and single-cell tasks. We perturb one biological input while holding the query and other inputs fixed, construct evidence conflicts that pair the foundation model representation of one genome, protein, or cell with the text of another, fit linear probes to the representations the language model receives, and analyze reasoning traces against the biological inputs. Evo2 and ESM3 contribute little to BioReason and BioReason-Pro performance on the evaluated tasks. Shuffling the DNA sequence barely changes BioReason disease prediction accuracy, and in evidence conflicts the two models follow the text in 97.9% and 99.7% of cases. Linear probes trained on the Evo2 and ESM3 representations predict the task targets, so these foundation models encode information relevant to the task, but provide limited overall performance improvement to BioReason and BioReason-Pro. In contrast, foundation model inputs contribute to ChatNT, Prot2Text-V2, and CellWhisperer performance, and differentially expressed genes in the gene sentence contribute to Cell2Sentence-Scale performance. Across SFT and RL checkpoints of BioReason-Pro and 42 BioReason checkpoints, increases in accuracy do not imply greater performance contributions from biological inputs. BioReason traces misstate nucleotide changes, while BioReason-Pro traces describe functions omitted from final predictions under evidence conflicts. We find that current post-training strategies do not ensure that foundation model representations contribute to task performance.

Looped Transformers introduce recurrent depth as a new scaling axis for LLMs: by repeatedly applying shared Transformer blocks, they increase effective depth without increasing parameter count. However, the benefits of looping remain unclear for large MoE LLMs under FLOPs-matched comparisons. The main reason is that the gains from additional iterations diminish quickly and can even turn into degradation, so the extra FLOPs spent on looping yield little substantial improvement. Consequently, prior work typically settles on two loops. We identify two main obstacles to scaling looped MoE. First, looping inherits and amplifies the curse of depth: hidden-state variance grows with each iteration as residual updates accumulate, which destabilizes deep recurrence and causes representations to drift. Second, looped MoE suffers from expert selection collapse: routers repeatedly select the same experts across loops, so extra iterations add computation without adding computational diversity. Guided by this diagnosis, we propose LOOM, built on a single principle: each loop should contribute new computation while keeping the recurrent state stable. LOOM stabilizes recurrence by scaling residual updates to bound variance growth and re-injecting the input embedding at every loop, and diversifies it through per-loop routers that engage different experts and a Looping Residual that carries earlier outputs forward. Experiments across 100M-1.7B models show stable scaling to 9-12 loops. Under near-iso-FLOP, the 700M model performs best at 5 loops, reducing perplexity from 18.36 to 16.54 and improving average zero-shot accuracy from 38.84% to 39.53% over the non-looped baseline. Without FLOP matching, the 1.7B model trained on 60B tokens peaks at 9 loops, reducing perplexity from 9.62 to 7.77 and improving average zero-shot accuracy from 42.4% to 47.7%. Code is available https://github.com/hed-ucas/LOOM.

Generalization in large language models (LLMs) is the ability to produce consistent and semantically stable outputs when the same input is expressed in different ways. Existing work typically evaluates generalization through aggregate accuracy on a single prompt format, task, or set of variations, which conflates robustness with overall benchmark performance. In this work, we show generalization evaluation at the level of individual examples, across multiple input variants, and across different aspects of model behavior, focusing on variability rather than reducing performance to a score that can be improved through narrow training or other ways that obfuscate generalization evaluation. Following this view, we introduce the Stability-Aware Generalization Objective (SAGO), a framework that measures how much model behavior changes for the same input under different variations and benchmarks, capturing variability across several dimensions including generation consistency, internal activations, confidence, and response mirroring. We show that many commonly used models exhibit statistically significant and consistent generalization instability: no model generalizes uniformly, behavioral axes capture independent failure modes, and cross-dataset variation can reverse model rankings.

An emerging hypothesis about reinforcement learning (RL) post-training of large language models (LLMs) is that it merely sharpens existing behaviors of a base model, improving single-shot accuracy at the cost of solution coverage. Although this trade-off has been observed in math and coding tasks, it need not extend to agentic tasks, where multi-turn tool use and interaction may require capabilities newly acquired during post-training. Our surprising finding is that pre-trained LLMs, equipped with a light inference harness, can serve as capable agents. Despite far lower accuracy (pass@1), they often surpass their post-trained counterparts in solution coverage (pass@K) given a sufficient test-time budget. We further analyze the underlying mechanism and show that post-training pushes tasks toward two extremes, always solved or never solved, and thereby improves sampling efficiency and consistency at the cost of solution coverage. To measure this cost, we propose Sharpening Tax, a diagnostic metric that quantifies the loss in test-time scalability after post-training. Across 14 base/post-trained model pairs from four families and three agentic benchmarks (42 cases in total), the tax is prevalent in most settings, can be estimated from a few rollouts, and correlates well with other metrics. Finally, we present posterior-tempered group sampling (PTGS), a simple plug-and-play Bayesian sampler that adapts the sampling temperature per prompt to its estimated difficulty. Applied during RL training in two agentic environments, PTGS pays a smaller tax than the fixed-temperature baseline, solving more tasks under repeated sampling while also improving single-shot accuracy.

Large language models are often asked which input factors influenced their outputs. For structured inputs, such reports can be checked by counterfactual perturbation, but each factor must be queried multiple times to estimate its effect, so verification is usually budget-limited. We study how this limited-budget setting changes the incentive to report factor-level influence truthfully. We formalize the interaction as a verification game and show that proper scoring alone is not enough when auditing depends on the report: report-dependent auditing creates a suppression incentive, because factors reported as important are more likely to be checked and penalized for estimation noise. In contrast, report-independent auditing, or a mixed rule with a small report-independent floor, removes this channel and makes truthful reporting preferable to full suppression. We instantiate the framework with the Counterfactual Brier Score (CBS) and evaluate its predictions on four NLP benchmarks. A synthetic rational agent matches the theoretical prediction exactly, and real LLMs follow the same incentives when they are made explicit. The main design implication is simple: under partial verification, factor-level explanation systems should include a report-independent audit component so that under-reporting cannot be used to avoid scrutiny.

Model intelligence and fast response jointly shape the quality of interaction with speech language models, yet remain difficult to achieve together. Explicit chain-of-thought (CoT) improves reasoning and audio understanding, but generating intermediate reasoning tokens delays responses. Describing fine-grained acoustic cues further lengthens CoT and increases latency. Latent reasoning can reduce this overhead, yet existing methods often trail CoT and remain limited by single-path supervision and reasoning budgets that do not adapt to problem difficulty. We introduce AURAL, which models a distribution over multiple plausible reasoning continuations in latent space and jointly predicts chunks of future states to reduce sequential forward passes and reasoning latency. To provide initial supervision for latent reasoning, we construct AuralReason-683K: 683K bilingual speech utterances (about 1,000 hours) with concise CoT for emotion recognition, empathetic dialogue, and general reasoning. AURAL-RL then explores beyond these traces, rewarding concise reasoning that yields high-quality answers and adapting reasoning effort to each problem. Across two backbones, AURAL-RL achieves performance comparable to CoT-RL, with larger gains over the respective supervised checkpoints on most metrics. Analysis further shows that harder questions elicit more latent reasoning steps. On Qwen2.5-Omni, it reduces time to the first answer token by 11.8x, from 1.22 to 0.10 s, versus 0.05 s for direct answering.

In-context learning (ICL) enables a pretrained model to infer a task from demonstrations without updating its parameters. While much of the existing theory focuses on linear target functions, in this paper we study nonlinear cases by comparing two one-layer attention architectures on the same family of single-index tasks. A kernel learner first maps inputs through a fixed nonlinear feature map and then applies linear attention, whereas a feature learner applies attention to the original input, followed by a learned nonlinear readout. We derive predictions for their memorization and generalization errors using the replica method, retaining the effects of pretraining size, task-pool diversity, and training and inference context lengths. The resulting predictions closely match numerical experiments across a broad range of regimes. Our analysis yields phase diagrams that characterize when each architecture is advantageous as the amount of pretraining data, task diversity, and context lengths vary. We further identify qualitatively different context-length scalings for the two learners. Together, these results clarify how architectural choices interact with the dataset and govern nonlinear in-context learning.

Vision-language-action models benefit from the understanding and reasoning capabilities of pretrained vision-language models, but action-only supervision provides limited grounding in world dynamics. Conversely, world-action models inherit spatiotemporal priors from video generation models, yet remain limited in semantic understanding and reasoning under distribution shifts. We introduce UniWAM, a unified architecture that integrates a physical reasoner, a world generator, and an action predictor to jointly learn semantic understanding of the physical world, visual generation, and action prediction. To ensure the quality of the training data, we developed a rigorous data cleaning and annotation pipeline for both human egocentric data and robot data. To adapt the vision-language component to embodied tasks while preserving its inherited language capabilities, we represent low-level actions in natural language and introduce a pre-training recipe that assigns complementary supervision from visual question answering (VQA) data, human egocentric data, and robot demonstrations to the appropriate model components. During post-training, future visual noise augmentation reduces reliance on precise future predictions, while history-conditioned flow matching uses encoded action history to initialize action generation. Together, these designs significantly reduce denoising steps while maintaining performance. UniWAM achieves state-of-the-art (SOTA) performance across multiple evaluations, including in-distribution performance, robustness, generalization, instruction following, and long-horizon task execution. Furthermore, we uncover a log-linear scaling law of unified human-robot co-training, demonstrating the effectiveness of large-scale pre-training on a mixture of human and robot data.

Despite progress in vision-language models, 3D spatial reasoning from 2D images remains challenging. Text-based methods describe intermediate geometry with discrete tokens, limiting fidelity for continuous spatial relations. Continuous latents offer richer representations, but a single latent type does not explicitly separate the cues needed across spatial tasks. Decomposed spatial latents address this by representing position, direction, and global geometry separately under geometric supervision. Yet the geometry representation can still collapse toward one dominant direction, and unrestricted attention can leave the latents underused during answer learning. We introduce GeoLatent, combining Common--Residual Geometry Alignment (CR-GEO) with routed optimization to structure the geometry states while promoting latent-mediated answer learning. CR-GEO separates shared from residual teacher geometry; routed optimization jointly trains geometry and language, temporarily directs visual answer learning through the latents, and restores full attention with geometry supervision. In controlled comparisons, CR-GEO raises geometry effective rank from 1.00 to 3.87, while blocking latent readout at the bottleneck lowers direction accuracy from 89.1% to 25.8% on 128 fixed questions. After recovery, the differentiated geometry representation and latent-mediated visual route remain available alongside direct image access. GeoLatent achieves 73.0% on SPAR-Bench and 72.1% on SPBench, outperforming previously reported methods on both.

Asynchronous reinforcement learning (RL) improves the efficiency of large language model post-training but introduces stale rollouts generated by earlier policies. Theoretical understanding of how this staleness affects convergence and how to mitigate its impact remains limited. We derive a convergence bound for GRPO-style algorithms that explicitly characterizes the tradeoff between the gradient estimator's second moment and bias. For trajectory-level importance-weighted estimators, our analysis shows that once the second moment is uniformly controlled, delay enters the bound through the bias introduced by clipping or rescaling. Guided by this insight, we propose a novel group mass capping GRPO (GMC-GRPO) method, which minimizes a ratio-based bias bound within a class of weighted estimators sharing a common second-moment guarantee. We establish convergence guarantees for asynchronous GMC-GRPO and show that, compared with TIC-GRPO, it improves the threshold dependence of the fourth-order delay term from $O(ε^{-4})$ to $O(ε^{-2})$ as $ε\to0$, where $1+ε$ is the ratio threshold. Under local policy overlap, the delay-dependent term decreases as $G^{-2/5}$ after tuning the step size, where $G$ is the group size. For fixed behavior and current policies, the bias introduced by group rescaling also vanishes as $G\to\infty$, whereas the bias from trajectory-wise clipping can persist. Experiments across Qwen3 models and reasoning benchmarks demonstrate improved robustness to stale rollouts, with GMC-GRPO achieving the best performance among stable baselines under large rollout delays.