When drug discovery fails: scientists share their frustrations with the process
In clinical trials, there are many times (and ways) a therapeutic can come up short. In clinical trials, there are many times (and ways) a therapeutic can come up short.
Academic Intelligence · Curated Daily
AcademicHub 汇聚顶级期刊与预印本平台的实时文献。定制您的专属科研雷达,利用大语言模型自动生成交叉领域文献分析简报。
In clinical trials, there are many times (and ways) a therapeutic can come up short. In clinical trials, there are many times (and ways) a therapeutic can come up short.
Continual learning for pre-trained vision-language models requires balancing three competing objectives: retaining pre-trained knowledge, preserving knowledge from a sequence of learned tasks, and maintaining the plasticity to acquire new knowledge. This paper presents KeepLoRA++, balancing these objectives through a unified dual-dimensional knowledge retention mechanism. We analyze knowledge distribution of Transformer architecture from both inter-layer and intra-layer perspectives. The inter-layer perspective examines how retention is distributed across layers, while the intra-layer perspective focuses on the parameter space within each layer. Our analysis reveals a structural property: general transferable knowledge is mainly encoded in the shallow layers and the principal subspace of the parameters, while task-specific adaptations are localized in the deep layers and the residual subspace. Motivated by this insight, KeepLoRA++ introduces a layer-scaled residual gradient adaptation method. New tasks are learned by restricting LoRA parameter updates to the residual subspace, combined with a shallow-to-deep layer scaling, to prevent interference with previously acquired capabilities. Specifically, the gradient of a new task is projected onto a subspace orthogonal to both the principal subspace of the pre-trained model and the dominant directions of previous task features, while simultaneously assigning smaller update magnitudes to shallow layers and larger ones to deeper layers. Our theoretical analysis and empirical evaluations confirm that KeepLoRA++ successfully balances these three competing objectives, consistently outperforming representative baselines across image classification, visual question answering, and video understanding tasks.
When humans see a bird, they recognize far more than just "bird" – they see a head, wings, and talons, a structured assembly of reusable parts that can be identified across every bird they have ever seen. We ask whether a self-supervised visual model can discover the same compositional structure on its own. To this end, we propose RATS (Register Attention Transformers), which decomposes the classification token into N learnable register tokens that route patch information through an L->N->N->L bottleneck via a three-step compress-communicate-broadcast attention. The N registers are partitioned across the H attention heads, so that registers assigned to different heads do not interact with each other. Without auxiliary losses or part annotations, each register spontaneously specializes into a proto-semantic region whose emerging structure resembles object parts. RATS surpasses all baselines by +12 mIoU on average across five segmentation benchmarks, with consistent gains on ADE20K (+1.11 mIoU) and COCO (+0.2 AP^m). Its register dictionary further exhibits part-level consistency and semantic proximity across related categories. Our results suggest that RATS may provide a useful architectural prior for structured and interpretable visual representation learning.
3D Gaussian Splatting (3DGS) has shown strong potential for high-fidelity talking head synthesis. However, enabling fine-grained, interpretable, and editable facial expression control remains fundamentally challenging due to intrinsic conflicts between speech-driven facial dynamics and explicit expression signals. Existing methods rely on implicit multimodal fusion, leading to spatial entanglement and temporal instability. We present EmoZone-Talker, a novel framework that reformulates audio-driven facial animation as a structured spatial-temporal coordination problem under cross-modal conflicts. Our approach introduces an explicit spatial disentanglement and temporal dynamics modeling of facial motion. Specifically, we propose Synergy Zones with Prioritized Attention Bias (SZ-PAB) to explicitly decouple modality contributions via region-wise constraints guided by anatomical priors, and a Channel-Independent Temporal AU Encoder (CIT-AE) to model temporally coherent AU dynamics. By integrating these representations into 3D Gaussian deformation, EmoZone-Talker enables precise and interpretable control over facial expressions. Extensive experiments demonstrate that our method improves expression controllability and realism, with notable gains in upper-face accuracy and temporal coherence, while preserving high rendering quality and accurate lip synchronization. Code will be publicly released to facilitate reproducibility and further research.
arXiv:2606.12334v1 Announce Type: new Abstract: High-precision robotic manipulation requires fine-grained spatial reasoning that is often difficult to achieve with RGB-only policies due to depth ambiguity and perspective scale issues. Policies that leverage 3D information directly, such as those based on point clouds, offer a stronger geometric prior over purely image-based ones, yet their performance remains highly task-dependent. We hypothesize that this discrepancy may be due to the spectral bias of neural networks towards learning low frequency functions, which especially affects architectures conditioned on slow-moving Cartesian features. We thus propose to map point clouds from Cartesian space into high-dimensional Fourier space, effectively equipping the point cloud encoder with direct access to high-frequency features. We experimentally validate the use of Fourier features on challenging manipulation tasks from the RoboCasa and ManiSkill3 benchmarks and on a real robot setup. Despite their simplicity, we find that Fourier features provide significant benefits across diverse encoder architectures and benchmarks and are robust across hyperparameters. Our results indicate that Fourier features let policies leverage geometric details more effectively than Cartesian features, showing their potential as a general-purpose tool for point cloud-based imitation learning. We provide source code and videos on our project page: https://fourier-il.github.io/fourier-il
arXiv:2606.12629v1 Announce Type: cross Abstract: We show that the standard basis of transformer hidden states already provides a training-free, architecture-general feature basis. Individual dimensions encode semantic content via their signs and confidence via their magnitudes, functioning as independent binary registers. We validate this Bag of Dims framework across three model families (Qwen 3.5-4B, Gemma 3-4B, Mistral 7B) through four progressive experiments. Sign patterns alone carry predictive content: replacing all magnitudes with unity achieves 72-93% top-5 next-token accuracy through the LM head, and pure Hamming scoring without any decoder reaches 80-90% top-4096. These sign patterns organize into semantic features: using a single-token type cache (one forward pass per vocabulary token, no context), we discover 175 categories via per-dimension sign consistency (mean AUC 0.80) from 50 anchors with zero training. A trained probe adds only +0.018 AUC and converges to axis-aligned weights, confirming negligible cross-dimension structure. This structure extends to attention: all 175 categories remain discoverable in K and V projections. On the write side, static FFN weight inspection links 20% of features to individual writer neurons (>0.70 agreement; random controls: 0%), with top-200 neuron coalitions achieving >0.70 agreement on 99.9% of prototypes via majority vote. Fully unsupervised discovery (random seeds, no labels) scales to 1500 features at 100% yield and 99% sparsity across all three models, with pairwise MI of 0.0014 bits confirming low inter-dimension coupling. These results establish that the standard basis already suffices for feature reading throughout the transformer compute pathway, requiring no training, no optimization, and no GPU-days beyond a single forward pass per vocabulary token.
arXiv:2606.18722v1 Announce Type: new Abstract: In this short article we consider a preferential attachment random graph model with edge steps, studied by Alves, Ribeiro and Sanchis. Starting with an initial graph $\mathbb{G}_1$ formed by a vertex with a self-loop attached to it, the model evolves as follows. At every subsequent (discrete) time step, either with probability $p$ we add a vertex to the graph and connect it to exactly one of the older vertices selected with probability proportional to its degree, or with probability $1-p$ we add one edge between two existing vertices, both selected (independently) with probability proportional to their degrees. Let $\omega(\mathbb{G})$ be the clique number of a graph $\mathbb{G}$, i.e.\ the number of vertices in a largest complete subgraph of $\mathbb{G}_{}$. Alves, Ribeiro and Sanchis showed that, for any given $\varepsilon>0$, we have $\omega(\mathbb{G}_{2t})\geq t^{\frac{1-p}{2-p}(1-\varepsilon)}$ with high probability (i.e.\ with probability tending to $1$ as $t\rightarrow \infty$). Here we strengthen this bound by showing that, for any function $f:\mathbb{N}\mapsto \mathbb{N}$ that satisfies $f(t)\rightarrow \infty$ as $t\rightarrow \infty$, with high probability \[\omega(\mathbb{G}_{2t}) = \Omega\left(t^{\frac{1-p}{2-p}}\Big(\log^{\frac{1}{2-p}}(t)f(t)\Big)^{-1}\right).\]
An economist sets out how communities can direct economic decisions and not just endure their consequences. An economist sets out how communities can direct economic decisions and not just endure their consequences.
arXiv:2606.09022v2 Announce Type: replace Abstract: This paper studies heterogeneous mean-field models in which agent parameters are sampled from a population distribution. We establish an $O(1/M)$ bound on the steady-state mean-square error between the occupancy measure of the $M$-agent system and the corresponding annealed mean-field equilibrium. The analysis extends Stein's method for homogeneous mean-field models and reveals a fundamental difference between homogeneous and heterogeneous systems. While stability of the mean-field dynamics is sufficient in the homogeneous setting, heterogeneous systems further require uniform robustness of the occupancy dynamics with respect to perturbations of the initial condition. The results are illustrated through a heterogeneous SIS epidemic model.
arXiv:2509.08965v3 Announce Type: replace Abstract: We study the capacity of a quantum channel for retrocausal communication, where messages are transmitted backward in time, from a sender in the future to a receiver in the past, through a noisy postselected closed timelike curve mathematically represented by the channel. We completely characterize the one-shot retrocausal quantum and classical capacities, and we show that the corresponding asymptotic capacities are equal to the average and sum, respectively, of the channel's max-information and its regularized Doeblin information. This endows these information measures with a novel operational interpretation. Furthermore, our characterization can be generalized beyond quantum channels to all completely positive maps. This imposes information-theoretic limits on transmitting messages via postselected-teleportation-like mechanisms with arbitrary initial- and final-state boundary conditions, including those considered in various black-hole final-state models.
arXiv:2606.19133v1 Announce Type: cross Abstract: Scalable prediction of optical spectra is a critical component of high-throughput materials screening for optoelectronic applications such as solar cells. Existing surrogate models are trained on spectra computed from lower levels of theory or rely on rotation-invariant scalar features, limiting their geometric expressiveness. We explore the use of equivariant graph neural networks for optical spectra prediction, adapting GotenNet to this task and evaluating it on multiple datasets including a recently published collection of 10,533 structures with spectra computed at the level of the random phase approximation (RPA). The proposed model outperforms the current state of the art, with the largest gains in the 0-8 eV range and on predicting the static real permittivity, both of particular relevance for thin-film optics.
Editing natural images using textual descriptions in text-to-image diffusion models remains a significant challenge, particularly in achieving consistent generation and handling complex, non-rigid objects. Existing methods often struggle to preserve textures and identity, require extensive fine-tuning, and exhibit limitations in editing specific spatial regions or objects while retaining background details. This paper proposes Context-Preserving Adaptive Manipulation (CPAM), a novel zero-shot framework for complicated, non-rigid real image editing. Specifically, we propose a preservation adaptation module that adjusts self-attention mechanisms to preserve and independently control the object and background effectively. This ensures that the objects' shapes, textures, and identities are maintained while keeping the background undistorted during the editing process using the mask guidance technique. Additionally, we develop a localized extraction module to mitigate the interference with the non-desired modified regions during conditioning in cross-attention mechanisms. We also introduce various mask-guidance strategies to facilitate diverse image manipulation tasks in a simple manner. CPAM can be seamlessly integrated with multiple diffusion backbones, including SD1.5, SD2.1, and SDXL, demonstrating strong generalization across different model architectures. Extensive experiments on our newly constructed Image Manipulation BenchmArk (IMBA), a robust benchmark dataset specifically designed for real image editing, demonstrate that our proposed method is the preferred choice among human raters, outperforming existing state-of-the-art editing techniques. The source code and data will be publicly released at the project page: https://vdkhoi20.github.io/CPAM
Background. Ortho-positronium (o-Ps) lifetime, now measurable in vivo on long-axial-field-of-view (LAFOV) PET/CT, has been proposed as a biomarker of tissue oxygenation and hypoxia. Because o-Ps lifetime is dominated by tissue free-volume structure while the oxygen- specific contribution is small, whether an in-vivo lifetime contrast reflects oxygenation rather than anatomy is an open, identifiability-limited question. Aim. To test the oxygenation hypothesis directly using the heart's natural arterial/venous oxygenation contrast, with a built-in anatomical control. Methods. We re-analysed a public [82Rb]Cl human cardiac LAFOV PET/CT dataset (5.30 x 10^8 evaluated three-photon events). Per-compartment o-Ps lifetimes were extracted with a background-plus-two-component exponentially-modified-Gaussian (EMG) model. The list-mode to image mapping and right/left ventricle (RV/LV) identity were established lifetime-free (the mapping reproduces the provider's reconstructed image at block-correlation 0.998 and wins a joint multi-organ alignment panel). We applied a confound battery: registration stress test, blood-core vs wall, lung-air and wall-myocardium partial-volume, tissue density; and a structure/position-matched control (pulmonary artery, deoxygenated, vs aorta, oxygenated). An isotope-matched 82Rb uniform-quartz reference bounded the instrument's positional behaviour. All results were produced by two independent analysis pipelines. Results. RV o-Ps lifetime exceeded LV by delta tau = +0.304 ns (RV 1.700 +/- 0.172, LV 1.396 +/- 0.130 ns; about 1.4 sigma), in the oxygen-expected direction; the contrast was stable across +/-16 mm registration perturbation (sign preserved in 100% of 342 shifts) and resided in the blood core, not the wall. However, the matched-vessel control was null: pulmonary artery minus aorta = -0.011 +/- 0.344 ns. Lung-air and wall-myocardium partial-volume were disfavoured, and the effect fell within the isotope-matched 82Rb instrumental positional envelope (about 0.1-0.35 ns over 40 mm in uniform material). Conclusion. On this single subject, the cardiac o-Ps lifetime contrast does not provide a clean readout of blood oxygenation: an oxygenation effect of the observed (about 0.3 ns) magnitude is ruled out by the matched control, while a small physiological effect cannot be excluded. We provide a reusable confound-control battery for evaluating future in-vivo o-Ps oxygenation claims. Multi-subject replication with anatomy decoupled from oxygenation is required.
arXiv:2606.11266v1 Announce Type: new Abstract: The cost signal that constrained-RL algorithms optimize against is almost always reactive: the simulator emits a non-zero cost only after a collision has begun, and the Lagrange multiplier of PPO-Lagrangian grows only after the episode budget has been exceeded. At race speeds, where collisions are instantaneous and irreversible, any safety mechanism that waits for cost to accumulate is structurally too late. We present VLM-Safe-RL, a framework that integrates a frozen vision-language model into the CMDP Lagrangian update as an anticipatory cost term. The framework comprises four contributions: (i) Decoupled Dual-Path CLIP, independent reward/cost paths that respect the CMDP's factorization; (ii) VLM-Lagrange, an augmented multiplier update that incorporates a per-step VLM cost as an anticipatory term; (iii) Confidence Gating, a Bayes-optimal weight derived from a logistic noise model on the CLIP margin; and (iv) VLMPPOLag, the composed algorithm. On Safety-Gymnasium FormulaOne L2, our principal evaluation ($n{=}5$ seeds, $10^{6}$ steps, budget $d_{lim}{=}25$) VLMPPOLag$+$Conf is the only configuration in our default budget comparison that simultaneously retains substantive return ($J_r{\approx}40$) and holds cost within budget on a majority of seeds; the five constraint-aware baselines (PPOLag, CPO, CPPOPID, CPO-CLG, PPOLag-RND) each fail at least one requirement. The mechanism generalizes to held-out MetaDrive Medium (catastrophe rate $41\%{\to}26\%$, 95\% bootstrap CI $[-26,-5]$\,pp) and shows directionally consistent transfer to Bullet Safety-Gym; we report honestly where it does not (MetaDrive Easy/Hard, Qwen2-VL backbone) and trace the Hard failure to a Lagrangian-regulation pathology rather than the VLM signal itself. To our knowledge, this is the first work to use frozen VLM signals as an anticipatory cost term inside the CMDP Lagrangian update.
Synthetic data for autonomous driving is surging, powered by diffusion models that promise scalable scene generation. Yet key obstacles remain, as enforcing multi-view and temporal consistency often relies on backbone fine-tuning or added layers, which erodes pre-trained knowledge and weakens text alignment. Models also stay close to the training distribution, struggling under adverse weather and unseen configurations, and fidelity favors frequent over rare classes. We address these gaps with FrozenDrive, a controllable generative framework that preserves a pretrained diffusion models knowledge while achieving strong consistency. FrozenDrive conditions on rich driving-stack signals and text prompts, and introduces knowledge-preserving spatio-temporal attention to impose cross-view alignment and temporal coherence in a single pass within a parameter-free frozen diffusion backbone. An additional object-focused constraint improves per-object fidelity for rare categories. Without any weather- or scene-specific fine-tuning, our model synthesizes globally coherent multi-view driving scenes from text, particularly under adverse and rare conditions, and surpasses prior baselines. On nuScenes, FrozenDrive augmented data significantly improves AD models performance, especially at night and in rain, demonstrating stronger robustness when trained with our scenario-targeted data.
arXiv:2606.14356v1 Announce Type: cross Abstract: Applications in the 3D Computing Continuum, which unifies edge, cloud, and space, require combining multiple AI tasks such as object detection, time-series analytics, and natural language processing into Compound AI systems. These systems must satisfy stringent Service Level Objectives (SLOs) on accuracy, latency, and cost. A key mechanism for maintaining SLO compliance of Compound AI systems is runtime model selection, where AI models are dynamically switched for each workflow task. However, existing distributed and compound AI frameworks do not natively support runtime model selection. We present PLAIground, a framework that enables runtime model selection for Compound AI systems. PLAIground introduces Compoundable AI Model (CAIM) abstraction, which decouples task semantics from AI model implementations via Task and Data Contracts, enabling model switching without workflow changes. Additionally, PLAIground introduces Pixie, an SLO-driven runtime model selection algorithm, which dynamically selects the most suitable model for each task during execution. Our evaluation on two realistic Compound AI workflows demonstrates that Pixie achieves up to 91.3% accuracy while maintaining SLO compliance where fixed-model strategies either violate cost and latency budgets up to 21x or miss accuracy targets by 4%.
Hateful and propagandistic memes exploit the interplay between images and text to convey harmful intent that neither modality reveals alone. Although thinking-based multimodal large language models (MLLMs) have advanced vision-language understanding, their application to meme content moderation remains underexplored. We propose a reinforcement learning-based post-training method that improves classification performance and reference-based explanation quality in thinking-based MLLMs via task-specific rewards and Group Relative Policy Optimization (GRPO). Concretely, we (i) conduct a systematic empirical study of off-the-shelf MLLMs for hateful and propagandistic meme understanding across English and Arabic benchmarks, (ii) extend existing meme datasets with weakly supervised chain-of-thought (CoT) rationales via distillation and multi-LLM fine-grained propaganda annotations, (iii) introduce a GRPO-based objective with thinking-length regularization that jointly optimizes classification accuracy and explanation quality, and (iv) investigate self-supervised GRPO on unlabeled memes using consensus-based pseudo-labels. Experiments on the Hateful Memes and ArMeme benchmarks show that our approach improves over previously reported results on FHM accuracy (up to +2.1%, from 79.9% to 82.0%) and on ArMeme macro-F1 (up to +7.6 points, from 0.536 to 0.612 with explanations; +6.1 compared to the original ArMeme benchmark), while also generating natural-language explanations. On ArMeme, sequence-classification baselines remain stronger in terms of raw accuracy, whereas our approach provides more balanced per-class performance along with explanations. We publicly release our code, data extensions, and evaluation resources.
arXiv:2603.21396v5 Announce Type: replace Abstract: Recent work has shown that LLMs can sometimes detect when steering vectors are injected into their residual stream and identify the injected concept – a phenomenon termed "introspective awareness." We investigate the mechanisms underlying this capability in open-weights models. First, we find that it is behaviorally robust: models detect injected steering vectors at moderate rates with 0% false positives across diverse prompts and dialogue formats. Notably, this capability emerges specifically from post-training; we show that preference optimization algorithms like DPO can elicit it, but standard supervised finetuning does not. We provide evidence that detection cannot be explained by simple linear association between certain steering vectors and directions promoting affirmative responses. We trace the detection mechanism to a two-stage circuit in which "evidence carrier" features in early post-injection layers detect perturbations monotonically along diverse directions, suppressing downstream "gate" features that implement a default negative response. This circuit is absent in base models and robust to refusal ablation. Identification of injected concepts relies on largely distinct later-layer mechanisms that only weakly overlap with those involved in detection. Finally, we show that introspective capability is substantially underelicited: ablating refusal directions improves detection by +53%, and a trained bias vector improves it by +75% on held-out concepts, both without meaningfully increasing false positives. Our results suggest that this introspective awareness of injected concepts is robust and mechanistically nontrivial, and could be substantially amplified in future models. Code: https://github.com/safety-research/introspection-mechanisms.
Background: Peru lacks an integrated information system for patients with Inborn Errors of Immunity (IEI). Although disease registries are essential tools for data management and health planning, their success depends on implementation science approaches that account for local contextual factors. This study reports Phase I of a three-phase mixed-methods implementation project to design and develop a national IEI registry. Methods: Phase I consisted of a phenomenological qualitative study exploring stakeholder perspectives. Semi-structured focus groups and in-depth interviews were conducted with 29 key stakeholders across four groups: policy-makers, clinical experts, end-users (immunologists, residents, allied health personnel), and patient organization representatives. Interviews followed a guide structured around four a priori domains (structure, navigation, feasibility, and perception of existing systems). Discussions were conducted in Spanish, audio-recorded, transcribed verbatim, and coded using ATLAS.ti. A hybrid thematic analysis combining deductive and inductive coding was performed. Data elements proposed for the registry were triangulated with qualitative findings. Results: Thirty-six initial codes were consolidated into 15 categories, which were further integrated into four overarching themes conceptualized as pathways toward intention to use: (1) Environment, where governance, regulatory backing, and sustainable financing were identified as key enablers, while limited interoperability emerged as a structural barrier; (2) Technical Dimension, emphasizing usability, alignment with clinical workflow, and a hierarchical data architecture (demographic, clinical, therapeutic); (3) Users, highlighting clinical leadership, protected time, digital readiness, and perceived usefulness as stronger motivators than financial incentives; and (4) Patients, underscoring data protection, transparency, trust, and advocacy as essential for legitimacy and sustainability. Conclusions: A national IEI registry in Peru is perceived as necessary and feasible if implemented with strong regulatory foundations, interoperable design, robust data security, and user-centered architecture. These findings informed the development of an initial functional prototype and the operational plan for Phase II, focused on usability evaluation.
arXiv:2606.20518v1 Announce Type: new Abstract: Flow-matching text-to-speech systems achieve remarkable zero-shot quality but remain static after deployment: pronunciation errors on out-of-vocabulary proper nouns persist unless the model is retrained. We introduce FlowEdit, a life-long adaptation framework for frozen flow-matching TTS that learns pronunciation corrections as latent conditioning edits rather than weight updates. When corrective feedback is provided, FlowEdit optimizes a token-level perturbation in the text embedding space, then stores the correction in a Modern Hopfield Network serving as content-addressable episodic memory. At inference, corrections are retrieved via soft attention with a similarity gate, enabling fuzzy morphological matching. On our curated benchmark of 312 multilingual proper nouns across 18 language families, FlowEdit reduces target-word Phoneme Error Rate by 92.7% relative to the zero-shot baseline while maintaining identical general-speech quality. Corrections complete in approximately 15 seconds on a single GPU.
CAR T cell therapy can induce deep responses but also severe toxicities in patients with smoldering myeloma, an asymptomatic cancer precursor; given the potential risks and benefits, its success depends on careful patient selection.
arXiv:2606.12096v1 Announce Type: new Abstract: Any quantum computation consists of a sequence of unitary evolutions described by a finite set of Hamiltonians. For the case where this set consists of only products of Pauli operators, known as Pauli strings, we provide a necessary and sufficient condition for it to generate $\mathfrak{su}(2^n)$, i.e., to be universal for quantum computation on $n$ qubits. When combining Pauli strings with a general Hamiltonian, we show a sufficient (and in certain circumstances even necessary) condition for universality based on the Pauli-basis expansion of the Hamiltonian. As an application of these results, we prove two corollaries: (i) a necessary and sufficient condition for the universality of a general Hamiltonian given arbitrary single-qubit control on all qubits, and (ii) the universality of an XYZ Heisenberg Hamiltonian with local control of just two adjacent qubits.
arXiv:2606.18847v1 Announce Type: new Abstract: To assist humans over extended periods in real homes, embodied agents must remember user routines, world states, and past interactions. Existing long-term memory benchmarks mainly evaluate language-centric retrieval and question answering, while embodied benchmarks often focus on short-horizon task execution without testing long-term memory use in dynamic environments. We introduce WorldLines, a project-driven benchmark for long-horizon embodied household assistance. It constructs temporally extended household traces with dialogues, actions, execution feedback, object and device state changes, and converts them into evidence-linked samples for Memory QA and Embodied Task Planning. We further propose ObsMem, an observer-grounded memory framework that maintains visibility-aware memories and action-native state trails for state-aware decisions. Experiments reveal persistent challenges in partial observability, overwritten world states, and translating long-term memory into embodied plans, while ObsMem offers a stronger reference architecture for this setting.
arXiv:2606.14022v1 Announce Type: new Abstract: LayerNorm-based GNNs routinely erase the topology signals (degree, centrality, $k$-core) that node-selection policies should depend on, but the literature has not located where in the residual block the erasure happens. We answer that question: a positive per-node scalar inserted before LayerNorm is divided out up to a stabilizer term, while the same scalar inserted after LayerNorm reaches the score head as representation magnitude. The surviving slot is the post-LayerNorm position. We instantiate it with PostDeg, a parameter-free post-LayerNorm inverse-degree scale, and pre-register four falsifiers (graphwise scalars, extra LayerNorm, expressive same-slot capacity, backbone-agnostic source) that would reject the rule. PostDeg gains $+3.5\%/+2.5\%/+5.6\%$ over the LN backbone on influence maximization, network dismantling, and maximum independent set, with $10/10$ paired-seed wins per task; none of the four falsifiers fires. The takeaway is that placement, not parameterization, carries the gain – a small invariance check that generalizes to any positive topology scalar in any normalized residual stack.
Reinforcement learning (RL) has become a key component in modern large language models, yet the rollout stage remains the key bottleneck in RL training pipelines. Although Multi-Token Prediction (MTP) offers a natural solution to accelerate rollouts through speculative decoding, many studies have observed that MTP acceptance rates degrade significantly during RL training, leading to limited speedup performance. To address this bottleneck, we present Bebop, a systematic study of MTP in LLM post-training, and offer practical recipes to integrate MTP into large-scale RL pipelines. First, we reveal that the MTP acceptance rate is fundamentally bounded by the fluctuation of model entropy, which demonstrates a clear negative linear relationship with the rise of entropy in the RL stage. Second, we show that probabilistic rejection sampling largely alleviates the disturbance introduced by entropy in RL compared to greedy draft sampling. We further identify that the conventional MTP training objectives (cross-entropy or KL) are suboptimal in such settings, and therefore we propose a novel end-to-end TV loss that directly optimizes multi-step rejection sampling acceptance rate, yielding ~10% acceptance rate improvements, achieving up to 95% acceptance rates and up to 25% extra inference throughput gains across mathematical reasoning, code generation, and agentic tasks. Third, we test various online MTP training strategies during RL and show that pre-RL MTP training with e2e TV loss and rejection sampling achieves a consistent acceptance rate and speedup throughout the entire RL, eliminating the need for costly online MTP updating. We provide extensive experiments and analysis that validate our findings. Experimental results show our method achieves up to 1.8x end-to-end acceleration in async RL training of Qwen3.5, Qwen3.6, and Qwen3.7 models.