每天知道
你的领域
出了什么

Autonomous agents challenge conventional LLM serving by coupling repeated inference with persistent context and sandboxed tool execution. We present Aries, a full-stack experimentation framework that separates task semantics from execution configurations, reconstructs cross-component agent trajectories with correlated system telemetry, and exposes stateful tool execution through a consistent interface across heterogeneous sandbox substrates. We use Aries to conduct reproducible experiments on open agent harnesses and benchmarks. We complement these experiments with production traces from a commercial platform, grounding low-level systems research in observed production behavior. Our results show that (1) token-centric metrics miss non-inference bottlenecks, (2) retaining additional context yields diminishing accuracy benefits while reducing serving capacity, and (3) tool sandboxes alternate between long idle periods and short resource bursts, while current snapshot-based state management makes aggressive suspension costly. A complementary security analysis further highlights the need to reduce the sandbox attack surface. We then discuss the vision for agent-native serving systems designed around trajectory-level metrics, adaptive context management, elastic sandbox resource management, and sandboxes with minimized attack surface.

Adaptive designs are increasingly used in clinical trials and digital experiments to improve estimation efficiency by updating treatment randomization probabilities as data accumulate. While most existing work focuses on settings with a single-stage treatment, adaptive designs for longitudinal studies with multi-stage, time-varying treatments remain relatively underexplored. In this work, we develop a general semiparametric efficiency framework for designing longitudinal adaptive experiments to optimize the estimation efficiency of a broad class of target estimands of interest. An efficiency-oriented design criterion is proposed to accommodate both single-estimand targets and joint optimization across multiple estimands. We demonstrate that optimal randomization at earlier stages depends on later-stage allocations, yielding a backward-recursive strategy for deriving the oracle design, and propose a longitudinal adaptive design to sequentially learn and target the oracle design using accumulating data. We further develop an adaptive-design-likelihood-based longitudinal targeted maximum likelihood estimator (ADL-LTMLE) for asymptotically normal and semiparametric efficient estimation of statistical estimands from dependent data collected from adaptive experiments, without relying on parametric model assumptions. Applying the framework to time-to-treatment-initiation effects that compare initiating treatment at a given stage with delaying initiation until a subsequent stage, we show that designs optimized for a particular stage-specific effect can substantially compromise estimation efficiency for effects defined at other stages, highlighting the design trade-offs addressed by our framework. Simulation studies show the proposed design and estimation approaches achieve substantial variance reductions relative to non-adaptive designs, with performance close to that of the oracle design.

Although world-action models (WAMs) enhance long-horizon robot control by predicting visual evolution before acting, long-horizon reliability demands repeated re-grounding in real observations--not recursive rollout. Existing WAMs address this by refreshing history or KV cache with ground-truth data between chunks. However, such chunk-wise feedback operates at a coarse temporal granularity and thus fails to correct prediction errors at the individual time-step level. To address this, we propose Feedback Flow Matching (FBFM), a training-free inference mechanism that pushes re-grounding inside the actively generated chunk. During flow matching, FBFM applies a masked pseudoinverse correction to the conditional velocity field: it leverages the preceding action chunk to guide generation of the next action chunk, and uses the image observed after executing that preceding chunk to guide the next frame prediction. This cross-chunk pairing--where feedback from one chunk arrives in time to shape the next--creates an asynchronous loop that corrects errors without waiting for chunk boundaries. Being training-free, the mechanism improves responsiveness to unexpected events and suppresses drift in long-horizon tasks. We evaluate FBFM on both a joint-generation WAM (DreamZero) and a stage-wise WAM (LingBot-VA). On selected LIBERO and RoboTwin2.0 tasks, it improves success rates by over 5% in favorable settings, and real-world robot observation-prediction diagnostics show notably better tracking. We argue that FBFM offers a new paradigm for fine-grained online correction, bridging open-loop flow generation with closed-loop real-world dynamics.

We address the challenge of securely and efficiently outsourcing AI computations from a trusted but computationally weak client to an untrusted but powerful server, in the setting where the client holds both the input and the model, and the server must learn neither. We present MOSAIC, whose core is a novel matrix-multiplication masking protocol that scales to far larger matrices than prior work, enabling the safe outsourcing of modern workloads such as large transformer inference. By introducing small amounts of noise to the multiplication result and thereby relaxing correctness, MOSAIC achieves optimal asymptotic client overhead and concrete runtimes orders of magnitude faster than prior work. Its security reduces to the decisional LWE and LPN assumptions. Because this noise accumulates across the many layers of a transformer, a key technical challenge is bounding error growth; MOSAIC addresses this with an error-scaling mechanism based on random Hadamard rotations. On large 70B transformer models, MOSAIC's perplexity is comparable to popular quantization approaches and even matches full-precision BF16 inference on HumanEval. Finally, we present an end-to-end implementation showing how ideas like MOSAIC can promise a path towards large-scale confidential AI in modern data centers. Non-confidential inference is already distributed across phase (prefill/decode), layer, and time to maximize utilization of heterogeneous hardware, using RDMA-like networking to move activations, cached KV values, and weights across nodes. MOSAIC enables scaling of confidential compute by keeping the trusted computing base (TCB) small and outsourcing the bulk of the AI computation to untrusted accelerators.

Diffusion models have revolutionized generative tasks but incur high latency due to iterative denoising. While cache-based strategies accelerate inference by reusing intermediate features, they largely rely on static, sample-agnostic schedules. We argue that this rigidity overlooks two facts empirically validated in this paper: (i) generation difficulty varies across prompts, requiring adaptive resource allocation--complex inputs demand more computation while simpler ones require less; (ii) error sensitivity fluctuates across timesteps, where static policies may cache high-error steps or waste computation on low-error ones. We therefore propose OnlineCache, a dynamic caching framework that jointly learns when to cache and how to correct approximation errors. We leverage policy gradient to train a lightweight network for adaptive speed-quality trade-offs, and incorporate a learnable corrector to mitigate caching-induced errors. Both modules are jointly optimized under a bilevel optimization framework, with the policy targeting global generation quality and the corrector minimizing local errors. Our method automatically allocates computational resources across both samples and timesteps, improving overall generation quality. Extensive experiments demonstrate clear superiority. On FLUX.1-dev model, OnlineCache achieves nearly 3 speedup while preserving generation fidelity. On DiT and CogVideoX, it similarly delivers competitive acceleration without compromising quality; across all scenarios, it consistently outperforms existing cache-based acceleration baselines.

Training-free acceleration makes diffusion-based multimodal large language models (dMLLMs) more deployable, but it may silently change generated content. We study this serving-time consistency problem on 300 real images, comparing Fast-dLLM outputs with the same model's unaccelerated outputs. Across the mild parallelism induced in our long-form setting (1.05--1.25 committed tokens per step), confidence-threshold tuning changes decoding behavior but not baseline agreement. State-refresh ablations and an image-swap intervention instead identify stale visual and generated-text states as contributors to drift. For the tested Fast-dLLM implementation, shortening the KV-cache refresh interval yields a monotonic speed--agreement frontier and near-exact agreement at a measured 1.3x speedup. The initial diagnosis also appears with dLLM-Cache and LaViDa, although dLLM-Cache recovers agreement only after both caches are tightened, which removes its speed advantage. Independent prompts and images reproduce the threshold-insensitivity and refresh recovery. A targeted audit finds genuine content substitution in half of 50 low-agreement pairs. In a separate blinded two-annotator evaluation, the pooled accelerated-minus-baseline factual-error difference is 0.00 (95% CI [-0.17,+0.17]); this sample detects no difference but does not establish factual equivalence. Finally, none of the tested adaptive or smoothed-refresh variants beats the fixed interval at matched compute. Our contribution is a paired diagnostic and an implementation-scoped consistency control, not an accuracy or safety guarantee.

LLM serving stacks cache prompt KV state, yet the front end still re-tokenizes the full request text on every call. Coding agents pay the most: each call resubmits a long transcript after a small append, and reuse is hard because a short append can move token boundaries near the end of the prior sequence. Across 153,951 agent calls, the median append is 1.4K characters; only 1.0-3.6% of calls start or rebuild a session, but those carry multi-million-character contexts. At the fleet's 94.1% prompt-cache hit rate approaching 0.99, tokenization grows from 10% to 64% of time to first token. TokTier is a stateful CPU+GPU tokenization service for this two-mode workload with one contract: emitted token IDs are always identical to full reference tokenization of the request text. For session continuations it re-tokenizes a small window around the append and splices only when a per-request check finds a stable pre-tokenization boundary, else it widens or falls back. For calls without a reusable prefix it decomposes GPT-family regex pre-tokenization into run-local rules and runs exact pre-tokenization and BPE on a GPU. A sampled shadow verifier re-checks live traffic. Differential campaigns over 17 production tokenizer families ($1.5\times10^{10}$ split checks, a 12.4TB real-text corpus, 93,000+ replayed agent steps) show zero divergence. Incremental repair takes 0.5-1.1ms from 100K to 3M characters, up to $437\times$ faster than HF tokenization and $2.1\times$ faster at 1M characters than the strongest cache-based baseline (Gigatoken) fully prewarmed. GPU full tokenization encodes 1M characters in 0.87ms, $491\times$ below HF and $23.4\times$ below the fastest published CPU method. With vLLM, median time to first token drops 16-34% and P99 23%; under a 50ms P99 objective, four repair cores plus one GPU sustain 1,821 requests/s where a 16-core stateless front end saturates at 40.

Large language model (LLM) serving commonly increases batch size to improve throughput, but performance eventually reaches a deployment-dependent plateau beyond which larger batches provide marginal gains while increasing latency and GPU memory consumption. Previous studies have attributed this behavior to HBM/DRAM bandwidth limitations, but the underlying causes have primarily been supported by conceptual arguments or high-level performance observations. As our first contribution, we present a detailed GPU characterization using hardware profiling techniques, demonstrating that throughput saturation originates in the attention kernels during the decode phase. Specifically, we show that their nearly constant arithmetic intensity as active-context lengths increases -not merely larger batch sizes- drives DRAM-bandwidth saturation, while the achieved compute throughput remains far below the hardware limit. Building on this analysis, we present the Batching Configuration Advisor (BCA), which selects the highest-throughput batching configuration satisfying a target latency constraint and identifies up to 55 GB of GPU memory allocation that can be avoided for the evaluated OPT models with minimal throughput loss. To enable these recommendations, we introduce SLIM (Saturation-Aware Lightweight Performance Model), a semi-analytical model that predicts LLM inference throughput and latency from analytical formulations of Transformer computation and memory traffic. Across the evaluated scenarios, SLIM outperforms representative performance-modeling baselines while successfully generalizing to previously unseen operating conditions.

Analog compute-in-memory (CIM) arrays have emerged as a promising substrate for energy-efficient LLM inference, particularly for weight-stationary computations in linear layers. However, extending analog CIM to attention mechanisms introduces a fundamental challenge: KV cache operations demand repeated in-situ weight updates, and the resulting mismatch with the weight-stationary paradigm exposes dynamic computations to significant hardware noise, a critical problem that remains largely unexplored. In this paper, we present the first systematic study of dynamic attention computation on analog CIM arrays, revealing that initial and recent tokens exhibit disproportionate vulnerability to hardware noise. Motivated by this token-level insight, we propose a hierarchical token protection strategy that keeps sink tokens and a sliding recent-token window on a higher-precision digital path while processing the bulk KV cache on analog CIM. A co-designed scheduler combines analog programming, ownership transition, and bulk-MVM tile formation to bound digital overhead. Evaluations on nine LLMs show that our approach lowers average perplexity under analog noise from 33.91 to 11.95, close to the clean baseline of 11.06, while improving dynamic-KV programming-row utilization from 23.1% to 91.2%.

Deploying large language models in realistic server environments poses challenges, as the system needs to provide high-quality responses with low latency. Quantization is a common approach to reduce the memory footprint and improve inference efficiency, yet its impact on latency and throughput is rarely evaluated under controlled, orchestration-level workloads. In this work we study the quantization trade-offs of two translation model families, EuroLLM \citep{martins2025eurollm} and Hy-MT2 \citep{zheng2026hy} across five models ranging from 1.7B to 22B for efficient deployment on a single A100 or H100 GPU. We demonstrate that combining a document-chunking strategy with W4A8 or W8A8 quantization improves the latency-throughput Pareto-curve under a wide range of workloads. Furthermore, since standard machine translation (MT) benchmarks rely on isolated sentences and fail to capture long-context dynamics, we introduce a document-level evaluation from WMT24++ to assess how text chunking strategies affect translation quality under quantization. Our results reveal that standard segment-level evaluation can fail to predict the interaction between quantization and long-context document translation. While Hy-MT2 remains robust under quantization, EuroLLM shows strong sensitivity and translation quality collapses rapidly for all considered quantization formats. Overall, our experiments show that the trade-off between inference efficiency and translation quality depends not only on the quantization format, but also on the choice of text chunking strategy.

Heterogeneous Large Language Model (LLM) systems increasingly rely on shared contexts, retrieved evidence, and multi-agent dialogue histories, yet their internal key-value (KV) caches remain model-specific and cannot be reused across architectures. Consequently, each model must repeatedly prefill or store caches for the same context, limiting the scalability of multi-model reasoning and long-context generation. We propose Mixture-of-Translators(MoT), a cache translation framework that maps context KV caches from a source LLM into the cache space of a target LLM. Unlike prior approaches that depend on a single projection path or global shared latent space, MoT uses multiple translator modules to capture diverse source--target mappings. To further reduce residual translation error, we introduce a Context Correction Loss that aligns the replayed target trajectory with the native target trajectory. We reveal two competing failure modes in cache translation: propagated translation shift from early injection and last-state shift from late injection. MoT addresses them through translator mixtures and target-side correction. Across homogeneous and heterogeneous translations among Qwen2.5, GPT-2, and OPT models, MoT preserves downstream QA performance, including Qwen2.5-7B-scale translation with 51.0% average closed-set QA accuracy and 0.43 average extractive QA F1. In practical case studies, MoT enables quality-preserving memory reuse for multi-agent reasoning and retains 96.3% of direct-context quality in long-context cache-augmented generation, demonstrating scalable KV cache reuse across heterogeneous LLMs.

KV cache compression is essential for efficient long-context inference. Existing eviction methods permanently discard unselected tokens and consequently remove their aggregate contribution to attention. Merging-based alternatives preserve more information but can perturb retained keys and values that should remain exact. We observe that the information omitted by cache eviction can be formulated as residual statistics in both the numerator and denominator of softmax attention. Based on this observation, we propose ResKV, which divides a fixed KV budget into an exact main cache and a compact residual cache that reconstructs the contribution of omitted tokens. ResKV lets main-cache tokens and residual entries participate in the same softmax normalization, so residual entries restore both attention numerator and denominator mass rather than acting as a post-hoc correction. A construction-time validation proxy determines residual allocation for each layer and KV head, while a decode-time dynamic gate adjusts residual contributions for individual queries. Comprehensive evaluations on LongBench and RULER, covering query-aware and query-agnostic settings, multiple backbones, cache budgets, and representative compression baselines, demonstrate broad improvements under the same retained KV budget while preserving the practical efficiency of compressed decoding, including peak memory usage and long-context decode throughput.

Sycophancy refers to the tendency for large language models (LLMs) to match user beliefs at the cost of factual correctness, thereby undermining model reliability. Prior work on evaluating sycophancy in LLMs aims to assess whether a model's output matches an authority's claim, but cannot reveal which part of the prompt drives this sycophantic behavior. To bridge this gap, we investigate the relationship of sycophantic responses with an authority's credentials, their assertive claim, and the problem statement. We introduce the Authority Share Index (ASI), an Integrated Gradients-based token attribution method, which measures the degree to which a model's decision is driven by authority-related text. Through extensive experiments across five models and 30 test configurations, we find that sycophantic responses consistently direct more attention toward authority tokens than resistant ones. Moreover, our token attribution method reveals that for the sycophantic cases, the claim asserted by the authority receives more attention than the authority's credentials. Building on these findings, we propose attribution-guided contrastive activation steering to mitigate LLM sycophancy. Our method constructs a steering vector from high-attribution tokens of sycophantic and resistant responses, selectively pushing models toward resistance. This enables inference-time steering without retraining, lowering sycophancy from 96% to 25% in the strongest case. Together, our results show that token-level attribution can both explain what drives sycophancy and directly inform a practical intervention.

Long-term memory lets large language model(LLM) agents reuse prior preferences and work flows, but it also turns untrusted observations into persistent action context. We identify memory provenance laundering: during LLM-based memory consolidation, an external observation may be rewritten as apparent user history or workflow support, preserving an action trigger while erasing the low-trust source that should limit its authority. Existing prompt filters, content sanitizers, and tool guards do not enforce source-authority non-amplification after lossy memory consolidation. We formalize this boundary and instantiate it as Provenance-Preserving Memory Fire wall (PPMF), a lightweight memory middleware that preserves platform-maintained provenance and authorizes tool calls by matching action risk to the authority of action-relevant memories. In our schema-grounded evaluation with fixed risk policies, vulnerable consolidated memories reach up to 1.000 attack success rate(ASR); with intact platform-maintained provenance, confirmation, and risk labels, no evaluated unauthorized high-risk action passes the PPMF gate while confirmed benign actions and targeted low-risk memory use remain executable.

With the ever-increasing pervasiveness of smart edge devices, the demand is growing for applications that can be tailored to users (e.g., custom keyword spotting) or patients (e.g., adaptive health monitoring). Yet, most edge devices rely on fixed inference algorithms and thus cannot learn on-device to personalize predictions. When they can, devices typically support only a specific learning scenario, such as few-shot learning (FSL): going beyond this requires resorting either to another specialized device or to cloud-based retraining, which implies significant energy and latency overheads, a lack of real-time capabilities, and privacy concerns. In this work, we introduce embedder-centric learning (ECL), a framework that unifies four different online learning scenarios: FSL for on-the-fly customization, continual learning (CL) for knowledge accumulation, zero-shot learning (ZSL) for leveraging semantic data, and in-context learning (ICL) for adapting beyond classification. We demonstrate in silicon that ECL can be deployed on resource-constrained devices across four real-world use cases representative of the aforementioned learning scenarios. Our approach establishes a new state-of-the-art performance for FSL character recognition (Omniglot: 96.8% for 5-way 1-shot, 83.3% for 32-way 1-shot), and the first hardware baseline for CL in keyword spotting (NeuroBench keyword FSCIL: 71.8% for 200-way 5-shot). Moreover, we present the first hardware demonstrations of ZSL with semantic data (60.6% for 5-way spoken sentence classification) and ICL (46.2% at the 500th token of RegBench) operating at micro-to-milliwatt power budgets. Therefore, by unifying multiple learning scenarios, we pave the way for smart and versatile devices that can adapt right at the edge, without reliance on the cloud.

The Abstraction and Reasoning Corpus (ARC) tests whether a model can infer an unseen transformation from a few input-output examples and apply it to a new grid. Looped visual reasoners refine predictions over multiple iterations, but conventional training constrains only the final output, leaving intermediate refinements unconstrained. We propose that these refinements should instead follow the transformation step by step. We introduce TraceViT, a looped visual reasoner trained with semantically monotonic transformation chains. We obtain these chains by rewriting and verifying programmatic task implementations, decomposing each solution into intermediate grid states. Each iteration is grounded by a task reference derived from the few-shot demonstrations and an object workspace representing the current grid state. Because these chains may differ in length from the loop, soft trace alignment enforces only their ordering, letting the model allocate iterations freely. TraceViT achieves 67.8% pass@2 on ARC-AGI-1 and 24.3% on ARC-AGI-2. Controlled ablations on ARC-AGI-1 show that trace supervision becomes beneficial only when paired with grounding. Code and data will be available at https://github.com/LiuBinnan/TraceViT.

Training large language models (LLMs) to write register-transfer level (RTL) requires large corpora of paired specifications and code, and such data is scarce enough that most public corpora are now synthesized. Synthesis provides scale but not correctness, and in two widely used RTL datasets only 24.4% and 53.5% of pairs pass generated functional tests. This raises the question of how much of such a corpus to keep and which part of it. Correctness alone is a poor answer. A pair that misbehaves in one corner case still shows valid syntax and interface conventions, and complex sequential designs are both harder to generate and harder to validate, so filtering by correctness leaves a corpus of short and simple modules. Correctness is also hard to obtain, since behavior leaves little trace on the surface in RTL, and validating an entire corpus only sorts pairs into passed and failed. We present RTLCurator, which learns a behavior-aware compatibility prior by contrasting each specification with implementations that fail simulation, and calibrates it to a new corpus using a small number of validated pairs. It then constructs the retained subset by balancing alignment, representation coverage, and RTL structural richness. On CodeV and RTLCoder, keeping 80% of the corpus this way improves on training with the full corpus across all reported metrics while validating only 10% of the pool, whereas ranking by the score alone falls below random selection and filtering the whole pool by simulation does no better.

Large language models (LLMs) are increasingly used to solve complex problems by searching over program space, offering a general paradigm for scientific problems that can be naturally represented and solved as programs. Despite recent progress, identifying effective optimization directions for a candidate program remains challenging. By analogy with automatic differentiation, existing methods typically guide the search using a textual ``gradient'': a first-order update direction expressed as textual edits. Such gradients are inferred either from previously evaluated programs or from LLM-generated feedback on the implicit program-score mapping. However, these estimates become increasingly unreliable as the program--score mapping grows more complex, limiting their practical utility. We argue that explicit gradients are not essential for effective program optimization. Leveraging their prior knowledge, LLMs can propose plausible atomic edits directly from the current program, thereby enabling a zeroth-order optimization strategy. However, zeroth-order search suffers from a weakest-link effect: when a bundle of edits is accepted or rejected as a whole, a single harmful edit can negate the benefits of all remaining edits. To address this issue, we introduce HERO, a program optimizer that prompts an LLM to generate diverse, non-overlapping atomic edits and then systematically selects and composes them into coherent program improvements using evaluator scores. We evaluate HERO across algorithmic problems, strategy games, the design of LLM-based agentic systems, and robotic path planning. Across these domains, HERO consistently discovers higher-scoring programs and converges substantially faster than prior LLM-based optimizers, while consuming fewer tokens.

Emotional video captioning (EVC) aims to describe a video with both factual correctness and affective expressiveness. It requires a model to perceive subtle, ambiguous, and temporally varying emotional cues and translate them into natural language without weakening objective visual content. Existing methods have progressively introduced contextual attention, emotion interpretation, emotion priors, dynamic emotion perception and emotion-cause reasoning. Nevertheless, most of them still depend on either global emotion vectors or rigid hierarchical priors. In recent methods, the tree-structured emotion prior establishes a coarse-to-fine connection between psychological emotion categories and daily emotion words, but its hard subordinate masking may irreversibly suppress correct lexical emotions once the coarse category prediction is inaccurate. It is also limited in representing mixed or overlapping emotions that frequently occur in real videos. To address the issues, we propose SAGML, an adaptive EVC framework via affective heterogeneous graph and multi-task language modeling. Instead of treating the emotion prior as a discrete tree, SAGML constructs a soft affective heterogeneous graph containing catalog-level emotion nodes and lexical-level emotion word nodes. The soft gate is injected into video-to-emotion graph attention as a continuous bias, allowing visually supported lexical emotions to remain recoverable rather than being removed by a hard mask. The resulting affective representation is fed together with visual tokens into a causal language decoder, while dual catalog and lexical heads impose explicit emotion distribution learning on the prompt hidden states. The overall model is trained with a joint objective that combines autoregressive caption generation and emotion distribution supervision. SAGML provides an error-resilient and multi-emotion-aware baseline for EVC.

Large language models (LLMs) generate text by auto-regressively sampling the next token. This inherently leads to a many-to-many mapping between prompts and responses, complicating the task of inferring prompts from observed outputs. Prior work on LLM inversion frames prompt recovery as a semantic reconstruction task. They rely on fine-tuning pretrained sequence-to-sequence models on large external datasets--and requiring access to model weights or logits--to generate semantically plausible prompts. In contrast, we present a functional approach to inverting a given LLM in a black-box setting, without auxiliary aids. We train an explicit inverse language model entirely from scratch on data synthetically generated from the target LLM itself. Analogous to forward next-token prediction, our inverse model is trained using previous-token prediction, establishing a generative link between the forward and inverse processes that enables faithful prompt reconstruction. Moreover, it naturally supports diverse prompt reconstructions through sampling, whereby all such prompts induce similar responses under the forward, target LLM. Our approach generalises across datasets and exhibits transferability in reconstructing prompts from responses generated by different LLMs. Further, across the set of token based evaluation metrics for prompt and response reconstructions, our approach outperforms prior work.

Deepseek's Engram, a conditional memory module, was introduced to trade-off storage versus reasoning in large language models. However, the module relies on token-level $N$-gram hashing for Engram embedding lookup, introducing a tight coupling to the tokenizer used: a model with a different tokenizer would have to train its own Engram embeddings from scratch. To improve the reusability of Engram embeddings, we propose a change to the hashing routine, enabling compatibility between Engram models using different tokenizers. Instead of modelling disjoint $N$-gram spaces, we treat $N$-gram as a method to sample potentially useful byte sequences, from all possible byte sequences across tokens. We replace the XOR-based hashing with the general polynomial hashing with a joint embedding space across $N$. This work investigates the possible trade-offs and shows that this simple substitution produces comparable performance and achieves tokenizer-agnosticism: hash equivalence for byte-equivalent token sequences.

Conversational AI developed by geopolitical rivals reaches citizens worldwide, raising concerns that it could sway public opinion or be rejected as foreign propaganda, with consequences for democratic discourse and information sovereignty. Yet, whether an AI's perceived national origin shapes its persuasive power is unknown. In a preregistered randomized experiment, 403 adults from a nationally representative United States sample held a three-round debate with a chatbot introduced as either American ("DiscoveryAI") or Chinese ("ZhengheAI"), discussing a political or non-political topic. In all conditions, participants actually conversed with the same model (GPT-4o), instructed to argue against their initial position. We combined pre- and post-conversation self-reports of attitudes, trust, and collective narcissism with computational analyses of 1,209 participant turns, including LLM-coded stance and argumentative conduct, stance-sensitive embeddings, and keyword-masked emotion and toxicity classifiers. The conversations produced substantial attitude changes in every condition. Critically, the nationality label affected neither self-reported attitude change nor expressed stance, concessions, counterarguing, or affect, and equivalence tests and Bayes factors largely supported these null effects. The label's only reliable footprint was lower pre-conversation human-like trust in the Chinese model, whereas functionality trust was unaffected. Political topics slowed stance movement toward the AI's position, and collective narcissism predicted less attitude change regardless of origin, acting as a general barrier rather than an out-group filter. Users thus initially withhold social trust from a rival's AI yet still assimilate its arguments; origin labeling and transparency requirements alone may offer weak protection against foreign influence operations conducted through conversational AI.

Interventional data is widely regarded as the gold standard for teaching models causal reasoning. We test this assumption in a fully controlled synthetic environment pitting observational correlation against causal effect, and find it fails instructively. In Simpson's-paradox worlds, where the two have systematically opposite signs, increasing the fraction of interventional samples in pretraining does not improve causal direction: the magnitude of the model's do()-response grows monotonically, yet its sign is copied from the observational context. What governs whether interventional evidence is used is not the training mixture but the evidence type present in the context at inference time. Under an identical training recipe, a purely observational context induces systematic sign reversal in 29/50 worlds, a mixed context in 19/50, while aligned interventional probes alone yield 41/50 correct. Erasing observational evidence from the context immediately releases the suppressed causal interpolation ability (ratio_true = +0.56); a four-state content manipulation shows the switch is content-mediated and graded. The suppression is stable across training seeds (11/11 strong reversals persist on a matched-protocol second seed) and robust as a rate at 0.93B parameters (31.8% vs. 6% reversals in the matched probe-only arm), even as absolute gains shrink four-fold. An external audit on CLadder exposes a learned positive-effect prior with a two-layer structure: sign-randomized retraining removes it in-distribution but not out-of-distribution. We summarize: the capability lives in the weights; the switch lives in the context, and activation patching localizes the switch to the middle layers' observational rows. We further quantify the sampling noise floor of probe-based causal evaluation and an evidence-averaging protocol that cuts sign errors from 26% to 9%.

Large language models generate computationally expensive yet semantically void reasoning on beyond-capability tasks, creating risks where plausible-sounding but incorrect derivations mislead users. We characterize this futile reasoning phenomenon through systematic analysis, revealing universal capability overreach and systematic miscalibration between capability and behavior. The dominant failure mode is specious reasoning, which outputs look superficially valid but contain subtle errors, escalating with task difficulty. To address this, we introduce CaRL (Capability-aligned Reinforcement Learning), which aligns model behavior with capability boundaries through reward shaping that incentivizes refusal over futile reasoning and hindsight refusal augmentation that converts failures into refusal supervision. Experiments demonstrate a substantial reduction in futile reasoning while preserving performance across task difficulties, effectively achieving capability-aligned behavior without sacrificing utility. \footnote{https://github.com/icip-cas/Knowing-When-to-Quit}

Adversarial training is one of the most effective defenses against adversarial attacks, yet the computational cost remains prohibitive at modern scales, especially for large language models (LLMs). While existing mitigation strategies, e.g., latent adversarial training (LAT), have been developed, they still incur a high computational cost. In this work, we comprehensively investigate computation-efficient strategies to speed up LAT from two complementary perspectives: (1) Defense-side optimization: We explore the representation fine-tuning (ReFT) within LAT, and reveal a potential issue if there is a mismatch on which tokens to apply ReFT and the attack. (2) Attack-side optimization: When computing adversarial attacks in each LAT iteration, we extract only the relevant circuits from the LLM to construct a lightweight surrogate model, avoiding the computation in the forward-backward passes through the full model during the attack generation. For both perspectives, we provide theoretical justifications and numerical evidence to illustrate the effectiveness of the proposed strategies. Ultimately, compared to standard LAT with full fine-tuning, our method on average reduces per-step adversarial-training FLOPs by 48.1% while requiring only 0.0118% trainable parameters.

Large language models increasingly generate optimization models from natural language, but existing evaluation often reduces a generated model and its ground truth to a single equivalent/not-equivalent verdict or an execution-success rate--labels that are neither independently checkable nor faithful to the multiple distinct senses in which two formulations can agree. We present ModelEquivBench, a certifying, multi-relational evaluation system that reports a per-pair semantic profile E0--E6: model construction and exact ingestion (E0), verified representation alignment (E1), same-space and projected feasible-set relations (E2, E3), objective-order equivalence (E4), optimal-value equality (E5), and optimizer-set equivalence (E6). Each decided entry carries relation-appropriate, independently re-checkable evidence: replayable traces or explicit maps for E0--E1, exact-rational certificates for positive E2--E6 conclusions, and explicit witnesses for supported negatives. Incomplete mapping search, unsupported structure, and resource limits produce typed UNKNOWN or N/A outcomes rather than guesses, while unmet prerequisites are reported as ABSENT. Using ModelEquivBench to evaluate three model snapshots--GPT-5.4, Claude Sonnet 4.6, and Qwen3.5-397B-A17B--on the same frozen cohort of 173 base problems (346 cells per model) under a no-repair protocol, the resulting profiles expose distinctions that coarse baselines do not represent: 49, 35, and 25 cells contain executable candidates that are nevertheless certified negative on at least one supported relation, and 25, 8, and 18 structural rejections occur on pairs for which E2 certifies mapped feasible-set equality under a verified map. The three model snapshots fail at different stages of the profile and therefore cannot be meaningfully reduced to a single accuracy score.

Symbolic Regression (SR) aims to discover analytical equations from observational data and plays a central role in scientific modeling. While recent Large Language Model (LLM) based approaches show promise, they face two limitations. First, they lack data analysis mechanisms for uncovering variable dependencies, which reduces the efficiency of equation discovery. Second, most methods rely on single-objective evaluation focused solely on fitting error. This neglect of structural complexity and generalization often causes models to converge prematurely to local optima, limiting their ability to explore the broader equation space. We propose Multi-Objective Tool-augmented Symbolic Regression (MOT-SR), a unified framework that integrates external analytical tools to extract structural priors and guide equation generation, while jointly optimizing for accuracy, complexity, and generalization via a multi-objective evaluation module that maintains a dynamic Pareto front. MOT-SR employs two collaborative LLM modules: a Meta Strategy Generator, which selects tools and synthesizes structural optimization strategies based on Pareto-optimal equations, and an Equation Generator, which produces new candidate equations accordingly. The system operates in a closed-loop manner, continuously refining both strategies and equation structures. Across 40 standard tasks, MOT-SR outperforms existing SR methods in accuracy, generalization, and efficiency. We further validate MOT-SR on extreme mass-ratio inspiral (EMRI) orbital modeling, an important problem in space-based gravitational-wave astronomy where small local errors can accumulate substantially over long-term evolution. The discovered interpretable correction achieves the lowest trajectory-level integration error on held-out configurations. These results demonstrate the potential of MOT-SR to enable reliable modeling of long-horizon scientific dynamics.

Large language models often improve task performance by generating long reasoning traces, but the resulting computation is frequently wasted on redundant verification and revision. Existing probe-based early-exit approaches mainly inspect explicit self-doubt expressions, leaving many earlier termination opportunities undetected. Expanding inspection to ordinary reasoning boundaries improves coverage, but also exposes highly diverse intermediate states whose predictive information may reside in different hidden layers. We present Boundary-Expanded and Layer-Adaptive Dynamic Exit for Efficient LLM Reasoning (BLADE), a lightweight framework that dynamically terminates reasoning by estimating whether the generated prefix is sufficient for correct answering. BLADE constructs multi-granular checkpoints from sentence, self-doubt, and paragraph boundaries, and derives robust training labels through repeated answer completions. It further learns a compact subset of informative probe layers instead of relying on fixed choices or expensive representations from all layers. At inference time, calibrated predictions are combined with checkpoint-specific confirmation rules to balance responsiveness and premature-exit risk. Experiments on five benchmarks and two Qwen3 reasoning models show that BLADE preserves near-baseline accuracy while reducing generated tokens by 24.8% on Qwen3-8B and 15.8% on Qwen3-4B. Ablation studies further confirm the benefits of diverse checkpoints and automatic layer selection, demonstrating an effective approach to more efficient LLM reasoning.

AI coding agents like GitHub Copilot, Claude Code, and Codex interleave multi-step LLM inference with tool execution, creating a workload different from chatbots. We present the first production-scale characterization of this workload using sampled GitHub Copilot traces from June 2026, comprising 3.2M users, 13M sessions, 761M LLM calls, and 95T tokens. Our analysis reveals distinctive workload properties with important systems implications. For example, agentic coding sessions consist of sparse user-initiated turns, each unfolding into an autonomous agent loop of LLM calls almost always coupled with tool execution. This structure yields KV cache hit rates averaging 90% within a turn, but falling to 55% across turn boundaries and drastically invalidated after events like model switches or context compaction. Diverse workflows and user behaviors are observed with variable and long-tailed token consumption, time span, and tool calls. We highlight the difference between quick agentic turnaround times and the minutes-long user idle periods at turn boundaries, and design a lightweight idle-time predictor that captures 86-90% of total idle time, enabling proactive decisions for efficient resource orchestration. These findings challenge assumptions underlying current LLM-serving systems and provide an empirical foundation for agent-native infrastructure.

Gene Ontology (GO) enrichment analysis is a foundational tool for translating large-scale genomic data into biological insights, but typically yields hundreds of redundant terms that obscure overarching themes. Existing summarization tools rely on fixed similarity metrics (REVIGO, GOSemSim, clusterProfiler::simplify()), gene-overlap measures (Metascape), or static hierarchy mappings (GO-slim), and therefore cannot incorporate biological context. Manual curation provides context-aware grouping but is subjective and labor-intensive. A scalable, context-aware framework is needed to cluster GO terms into interpretable higher-order biological domains. Here we present LLMBDC (Large Language Model for Biological Domains Oriented Clustering of Gene Ontology), a training-free framework that leverages zero-shot semantic reasoning of LLMs with confidence scoring to cluster GO terms into BioDomains using only ontology information at inference time. Benchmarked across Alzheimer's disease (AD) and Fragile X syndrome (FXS) against six baseline methods including SapBERT, LLMBDC achieved substantially higher precision, recall, and clustering performance. Against ground-truth annotations, LLMBDC improved ARI from 9.7% to 73.3% (AD) and from 15.7% to 66.6% (FXS) over REVIGO, with corresponding NMI gains from 59.9% to 73.4% (AD) and 66.0% to 79.5% (FXS). A Cauchy combination test further confirmed that aggregated BioDomains retained statistically significant functional signals. LLMBDC provides a scalable, reproducible, and interpretable route to context-aware, system-level interpretation of GO enrichment results while preserving biological specificity.