1635
157
Worldwide Express automated its less-than-truckload freight billing-exception process with a multi-agent solution built on Amazon Bedrock. Six months after launch, the company cut active exceptions by 75%, improved time-to-cash by four weeks, and unlocked $10 million in free cash flow.
Low-Earth orbit (LEO) satellite networks present a challenging routing environment due to rapidly changing topology, traffic, and link conditions. Existing deep reinforcement learning (DRL) routing methods typically rely on current observations and information encoded in network weights, limiting their ability to exploit successful past experiences. We propose RAMA, a retrieval-augmented multi-agent DRL framework with Approximate Information State (AIS), in which each satellite combines its local observation, recent routing history, neighboring state information, and retrieved entries from a per-satellite episodic memory. Retrieved experiences are integrated through a Transformer-based decision network trained with AIS-based auxiliary objectives. With weights frozen in a 140-satellite Kepler constellation, RAMA reduces mean end-to-end latency by 13.4% and 99th-percentile latency by 66.7% relative to the state-of-theart double deep Q-network (DDQN) routing method, mainly by avoiding long detours. Updated online, RAMA remains stable while DDQN degrades. On the 1,584-satellite Starlink shell 1, the reduction in mean latency grows to 93.2%, and updating RAMA online lowers its mean latency by a further 25.5%. These results demonstrate the effectiveness of retrieval-augmented memory for scalable decentralized routing in LEO satellite networks.
AI answers require more than apparent correctness: reviewers need to inspect their recorded basis. We introduce the Knowledge Layer Model (KLM), an open, implementation-neutral specification for inference records with versioned epistemic assessments, evidence relations, and explicit unknowns. SEDIM, the knowledge and evidence architecture of Nage AI OS, is a reference implementation. We reference KLM v0.3 (repository tag v0.3.0). Empirical material consists of separate internal experiments, not end-to-end validation. A historical platform-domain routing benchmark reports 137/150 classifications across 15 classes (91.3%; nominal 95% Wilson interval 85.7-94.9%); the numerator is recovered from a rounded summary, and historical example membership is unverified. A native-model weight-path configuration yielded 0/15 successful answers in a corrected two-source test. A capacity-growth comparison could not establish a reliable effect because of a defective legacy scorer and missing original outputs. Lexical and scoped diagnostics each yielded AUC 0.008 for silent-versus-contradicted examples, below the registered minimum of 0.65. Legacy packing lost gold-containing text in 118/221 selected cases; repaired packing retained it in 221/221. These measure text presence, not answer quality. Python and TypeScript verifiers reportedly agreed on six checks per synthetic fixture; a separate production disclosure was not accepted. Restoration, re-derivation, human-rubric support, boundary isolation, and bounded-response evaluations remain proposed. Findings are internal and author-reported; historical observations are distinguished from later reproductions. Evidence visibility does not establish correctness, causal reliance, security, or regulatory compliance. Independent replication and a complete public audit package remain future work.
We present GhostPrefix, the first synchronization-free universal adversarial prefix attack against automatic speech recognition (ASR) systems. Unlike prior audio attacks, which are typically crafted per utterance, require synchronization with the victim waveform, or target narrow objectives, GhostPrefix learns a short, reusable audio segment prepended to arbitrary live speech to exert real-time control over ASR outputs without knowledge of the future utterance or temporal alignment. A key challenge is maintaining stable adversarial effects across unknown, evolving speech contexts. GhostPrefix combines token-level losses with a novel representation-level objective for stable, input-agnostic effects. GhostPrefix supports three attack modes: (i) prompted prefix attacks for semantic steering, (ii) targeted prefix attacks for transcript hijacking, and (iii) untargeted prefix attacks for transcription degradation. To remain perceptually inconspicuous, we regularize prefixes toward benign environmental-sound distributions and improve robustness against over-the-air distortions. We evaluate GhostPrefix across 13 open-source ASR models from four families. In digital settings, it achieves average attack success rates of 97.6% for prompted steering and 93.4% for targeted hijacking, while driving word error rate to 1.18 under untargeted attacks. Over the air, it achieves success rates of 91.5% and 89.5% for prompted and targeted modes, respectively. These manipulations propagate into downstream AI-assisted pipelines, including voice assistants and AI note-taking tools, causing systematic semantic misinterpretations. GhostPrefix exposes synchronization-free universal prefixes as a practical, under-explored threat to real-world ASR and downstream applications. Artifacts and demos are available at https://ghostprefix.github.io/.
Multi-turn negotiation remains a challenging problem for large language model (LLM) agents. While LLMs can conduct natural-language dialogue, existing benchmarks predominantly assess only agreement rates, overlooking whether negotiated commitments actually create value once executed. To address this gap, we introduce TradeWar, a multi-country simulation benchmark that translates negotiation dialogues into concrete tariff and trade-share policies, enforces capacity and network constraints during settlement, and evaluates outcomes via an Economic Health Index (EHI). To solve this execution-grounded task, we propose TACTIC (Trade-aware Adaptive Consolidation and Theory-of-mind for International Commerce), an agent that iteratively refines its negotiation strategy across episodes-without updating base LLM weights-by leveraging prior experience, opponent behavior, and past settlement outcomes. Despite using zero-shot chain-of-thought (CoT) prompting, PromptLLM attains a 75% agreement rate but only 57.1 final EHI (*100), whereas TACTIC reaches 68.8. Across 20 sequential episodes and three seeds, TACTIC's final EHI improves from 58.2 to 73.2, compared with 65.2 for an ablation without cross-episode updates. Under technology-controls and energy-crisis shocks, TACTIC also exhibits smaller performance degradation than PromptLLM and MemoryLLM. These results indicate that CoT deliberation and interaction memory alone do not fully address delayed execution feedback, whereas settlement-grounded adaptation improves long-horizon negotiation in TradeWar.
When a language model is fine-tuned to follow instructions, or further pretrained on a new subject, it often gets worse at predicting standard web text. This is known as "forgetting." We ask whether this loss can be undone by simple, fixed adjustments to the model's final token predictions. Across 14 pairs of base and derived models, from 1B to 7B parameters, we fitted a fixed bonus or penalty for each token (a vocabulary bias), alone or together with a temperature. On text held out from correction fitting, the largest top-4K bias recovery estimate is under 10%, with loss conditional on the retained vocabulary. Combining the bias with a temperature recovers up to 36%. Models further pretrained on math lose the most and recover almost nothing. Finally, in a controlled simulation where a model is damaged at random, with no planted bias or temperature change, the bias-plus-temperature fix still recovers 29%. Most of the real-model loss increase stays out of reach of our restricted fits.
Weight pruning can reduce the storage and inference cost of large language models (LLMs), but preserving quality at high sparsity levels remains challenging. Existing methods face a quality-cost trade-off in this regime. One-shot pruning is cheap but degrades sharply as sparsity increases, even when followed by post-pruning recovery. A growing line of work instead optimizes the mask directly from the dense model under the end-to-end objective; among these methods, [Lee et al.,2026], the recent state of the art at extreme sparsity, reaches high quality but at the cost of model-sized optimizer state and auxiliary variables. To address this gap, we propose ETAPS, which prunes gradually and adapts sparsely: it tightens a global sparsity constraint in progressive stages while training a sparse delta with a small, fixed budget. Our key observation is that gradual weight removal keeps each pruning step small enough for sparse adaptation to compensate, allowing the adapted model to provide end-to-end gradients that guide the next pruning decision without model-sized optimizer state. Across six models from 125M to 13B parameters, ETAPS achieves lower perplexity than ELSA at every evaluated sparsity from 50% to 90%, with perplexity reductions of up to 42%. ETAPS is also far cheaper: at 90% sparsity, it prunes LLaMA2-7B on a single H200, where ELSA uses four, with 5.0smaller total persistent state and 49% fewer GPU-hours. Together, our findings point to a promising direction for achieving high quality in demanding regimes such as extreme sparsity at substantially lower cost.
Training language-model agents to play long-horizon strategy games is hard because deterministic feedback is sparse and often arrives only at the end of a game. In-game scores can offer dense signals, but they do not reliably predict future outcomes: in 8-68% of position pairs across eight strategy games, the higher-scoring position leads to the worse outcome. We term this phenomenon scoreboard traps. In Freeciv, an open-source turn-based empire-building game where several players manage units, cities, and diplomacy over hundreds of turns, we look into better reward signals for gameplay agents. To facilitate this, we build CivHarness, an open system that saves current game states, branches on different decisions, and replays the game to future turns to provide measurable long-horizon feedback from saved positions. We use CivHarness to train CivTelescope, an LLM Judge that predicts which one of two game positions will yield the better outcome. On scoreboard-trap pairs, CivTelescope is 1.3-2.0 times more accurate than off-the-shelf language models, and it generalizes to new game starts, later turns, and unseen rules. We then train a Freeciv gameplay agent with CivTelescope as a dense reward, and call this recipe CivMarshmallow. When used in the middle stage of the game where scoreboard traps are common, CivMarshmallow reaches an average training phase gain of 8.94 points, 77% more than the best baseline, the end-of-game reward. Over the full 120-turn game, CivMarshmallow uses a hybrid reward that starts from the scoreboard or the end-of-game reward and shifts to CivTelescope as the game advances; this hybrid reward scores highest among the rewards we compare.
Sparse MoE mechanism significantly improves the trade-off between model capacity and inference efficiency. However, practical deployment still requires storing all experts, resulting in substantial memory overhead. We show that expert outputs in MoE layers exhibit structured redundancy. Across different models and task categories, the stable rank and entropy effective rank of expert outputs are consistently far below the nominal number of experts, while expert output distributions also exhibit fewer effective modes. Motivated by these observations, we reformulate expert pruning as a structure-preserving expert subset selection problem and efficiently optimize it via a greedy strategy, rather than only independently scoring and ranking individual experts or exhaustively searching over expert combinations. We propose two expert pruning methods under the unified structure-preserving framework: Projection Energy Expert Pruning (PEEP), which preserves the functional subspace by maximizing weighted projection energy, and Distributional Coverage Expert Pruning (DCEP), which preserves the distributional geometry by minimizing weighted coverage error. Both methods incorporate router-derived weights as priors, enabling them to jointly preserve the global structural relationships among experts while accounting for the local importance of individual experts. Experiments on DeepSeek-V2-Lite and Qwen3-30B-A3B demonstrate that, at the expert retention ratios of 75% and 50%, our methods outperform the state-of-the-art baselines on the majority of benchmarks and consistently achieve the best average performance across both general tasks and mathematical reasoning tasks.
Large language models offer a promising route to register-transfer-level (RTL) design, but reliable implementations require checking and revision based on electronic design automation (EDA) tools' feedback. Existing RTL-specialized models often lack learned interaction with such EDA tools or rely on separate generation and repair components. We introduce RTL-Interact, a multi-task post-training framework that trains a single RTL-specialized model to generate RTL, construct task-appropriate EDA checks, and revise implementations using tool feedback. The framework combines automatically constructed multi-task demonstrations with staged supervision from chain-of-thought to tool-integrated reasoning. We further introduce Pattern-Aware Policy Optimization (PAPO) for multi-task reinforcement learning. PAPO augments Group Relative Policy Optimization with correctness-gated rewards that favor concise successful checking while rewarding effective revision, and uses task grouping to preserve coverage across objectives. On CVDP, RTL-Interact-32B achieves 48.7% weighted Pass@1 and 68.9% Pass@5, surpassing its teacher model, DeepSeek-V3.2, while RTL-Interact-8B achieves 44.9% weighted Pass@1. Additionally, we show that PAPO can reduce average tool calls by 63.8% while improving Pass@1. Both models also outperform all RTL-specialized baselines on VerilogEval v2 and RTLLM v2, demonstrating the effectiveness of our method.
Transformer-based language models have achieved remarkable success in automated short answer grading (ASAG). However, current methods in ASAG struggle to comprehend, reason about and evaluate complex mathematical answers involving symbols and formulas. These score-only methods fail to generalize across diverse responses and skill states associated with identical scores. Moreover, they ignore the potential upward bias introduced by generative Artificial Intelligence (AI) usage in assessing students' performance and knowledge mastery. To address these limitations, we propose AICG (AI-aware Cognitive-supervised Grading), a two-branch multi-task framework. The first branch combines RoBERTa-based regression, ordinal, and classification heads for individual-answer grading. The second branch uses an expert-designed Q-matrix to associate responses with assessed skills and jointly predicts G-DINA-derived skill mastery and learning gain. Through the shared encoder, these skill-level targets provide cognitive supervision for answer grading beyond score labels. AICG further uses a Gen-AI detector, trained through supervised contrastive pretraining followed by linear classification, to estimate assistance probabilities and convert them into reliability weights, attenuating unreliable cognitive evidence during learning assessment. We conduct experiments on 49,402 instructor-graded responses spanning 19 academic semesters. Relative to RoBERTa regression, AICG reduces mean absolute error (MAE) by 50.3% and improves quadratic weighted kappa (QWK) from 0.798 to 0.927. Ablations identify Q-matrix-guided cognitive supervision as the primary source of improvement, while AI-aware weighting further reduces overall MAE by 4.2% over the unweighted cognitive model and consistently improves MAE and QWK across the four most recent post-Gen-AI semesters.
Multimodal federated learning (MFL) trains a shared model by aggregating client updates without exchanging raw data. Under missing modalities, clients adapt active modality encoders and a shared prediction block (fusion and classifier) to different input subsets. This raises two server-side questions: which clients should contribute to each component, and how should their aggregates be incorporated across rounds? We observe that an absent-modality encoder can still adapt through missing-input handling. Availability- based aggregation can therefore omit its update while retaining jointly learned prediction-block updates. Contributor selection also shapes server momentum, a weighted history of aggregated updates. We observe that prediction-block momentum can improve or reduce performance depending on the encoder update policy. We therefore propose Role-Aware Server Optimization (RASO), which coordinates contributor selection with role-specific server updates. RASO selects contributors through prediction dependence and inclusion in local optimization, retains encoder momentum, and synchronizes the shared prediction block to its current aggregate. Experiments across datasets and fusion architectures demonstrate the effectiveness of RASO under missing modalities.
LLM agents increasingly triage security alerts by combining the outputs of several detection tools, implicitly treating those tools as trustworthy. Unlike prior work on conflicts between a model's knowledge and its context, we study conflicts among the tools themselves: what happens when this machine-generated evidence is conflicting or wrong. On the corrected CIC-IDS2017 dataset, four heterogeneous detectors emit labels and confidences that we corrupt under a threat model varying the number of corrupted tools, the persistence of the corruption, and the confidence attached to wrong outputs. Across open-weight LLMs from 3B to 14B parameters, unverified investigators are fragile: one corrupted detector out of four lowers naive accuracy to .50-.72, three lower every unverified investigator to .13 or less, and honest disagreement alone makes them abstain on up to 63% of flows. We propose Conflict-Aware Evidence Verification (CAV), built from established components: an equal-weight vote that ignores reported confidences, an evidence-conflict gate, and an independent nearest-neighbour trust check that tests the leading candidates. With its prompt frozen before testing, CAV-LLM improves on the naive investigator in every corrupted condition (p < 0.001), reaching .84-.88 with three corrupted detectors against at most .08, while calling the LLM on 0.4% of clean traffic. A fixed verification rule is as accurate or better unless most tools are corrupted. When we make the independent channel itself noisy, CAV inherits its errors: it stays ahead when two or more tools are corrupted, but adding the channel as an extra voter is better when at most one is. Verification thus moves trust from the tools to the check, which must be measured and protected in turn.
Retrieval-augmented generation is becoming agentic. A search harness writes queries, reads results, reformulates, and decides when to stop. This largely solves accuracy: a harness with only BM25/grep and MiniMax M3 answers 85% of BrowseComp-Plus and 91% on multi-hop FRAMES in our runs, without compiled or vector indexes. It exposes a context problem: an undisciplined loop repeats queries, rereads documents, and fills the window with unused passages. The missing layer is knowledge compilation: a configurable ingestion pipeline where LLMs enrich documents with context no chunk contains alone--navigation hierarchies with claim-evidence leaves, entity graphs, timelines, mind maps, ontologies, wikis, and memory. Compilation is not a better retriever: under a competent harness, structural artifacts gave no detectable accuracy gain on FRAMES, and a PageIndex-style outline degraded at scale. It is a supply layer. A navigation tree whose leaves are claim-evidence pairs--the mechanism we measure--cuts query tokens by 91% when evidence suffices and adds 1.9% when it does not, a 38% average reduction; this implies a 2% break-even sufficiency rate and a measured rate near 43%. A decision gate cuts latency from 140 s to 30 s, though substitution's token saving is bought with retrieval rounds and carries its own latency cost. Other artifacts--ontology, graph, wiki, timeline, mind map, memory--pay off in deployment (governance, integration, onboarding, temporal orientation, continuity) where accuracy benchmarks cannot see them. We describe RAGFlow's pipeline and harness (open source at https://github.com/infiniflow/ragflow), which turns the loop into five disciplined stations and keeps the context entering the model the smallest sufficient set; together they form an agent-native database: a store that hands the agent compiled, governed context instead of a ranked list of passages.
The ability to integrate new information and deduce its consequences is central to learning, understanding, and reasoning in large language models (LLMs). Whether information arrives during training or in-context, models must not only store local facts, but propagate their long-range implications globally. To explore information integration in LLMs, we introduce two complementary natural language benchmarks for large-scale evaluation of deductive reasoning as a function of inference depth. The first is a formal graph benchmark enabling controlled evaluation over a wide range of graph topologies and integration tasks; the second is a knowledge composition benchmark grounded in real-world knowledge graphs. We use these benchmarks to compare knowledge integration from information encoded in context versus information encoded in model weights during pretraining, evaluating a set of open models implicitly, without chain of thought. We find that accuracy decays sharply with increased composition length for both pretrained and in-context knowledge. Information presented in-context, however, substantially outperforms deduction over information encoded in weights. These results quantify a fundamental limitation of current models with direct implications for prompt-based control of agent behavior and for continual learning strategies, and they provide a foundation for developing training, prompting, and integration methods that support deeper reasoning.
Although large language models have shown strong capabilities in domain-specific code generation, their performance in robotics software development, especially for the Robot Operating System (ROS), remains largely unexplored. ROS code generation presents distinctive challenges, including complex inter-node coordination, middleware configuration, and lifecycle semantics that go far beyond standard coding benchmarks, and many important correctness properties become observable only once migrated nodes are launched and communicate. We present ROSTrans, a benchmark for evaluating LLMs on ROS1-to-ROS2 code migration, a representative and practically important setting for ROS code generation. ROSTrans comprises 108 scaffolded completion tasks drawn from a diverse set of 70+ open-source ROS1 repositories and organized into a hierarchical taxonomy assessing interface-level, behavioral-level and system-level correctness. For every task we establish an executable reference: a verified ROS2 implementation, a runtime test suite that launches the nodes and exercises their topics, services, actions and parameters, and a container ensuring reproducibility. Candidates are scored by injecting only their target files into an otherwise fixed copy of the reference package, so a candidate can neither supply nor weaken the tests it is judged by. Evaluating six frontier LLMs over five stochastic generations per task, we find pass@1 below 10.9%, and that 56.5% of all samples build successfully and then fail at runtime,and 32.0% fail runtime assertions on the migrated interfaces, a class of error no compile-only oracle can detect. Execution-guided repair further improves task success, reaching 49.1% and 47.2% for Claude Opus 4.6 and GPT-5.5 after three rounds, respectively.
To accelerate materials discovery, language models must be able to address problems at the research frontier. Yet textbook- and examination-derived benchmarks primarily assess established knowledge, leaving performance on frontier research problems largely untested. We introduce FrontierMat, a benchmark of 330 problems drawn from frontier academic and industrial research in fields including aerospace, semiconductors, embodied intelligence and nuclear fusion. Every problem is validated by leading experimentalists and industry experts in its field. Each problem presents a real study and asks the model to infer a withheld outcome or draw a conclusion from selected experimental evidence. Problems are annotated by functional domain, material system and the scale of the governing mechanism; 130 require cross-scale reasoning. The strongest general-purpose model scores only 53.7%, and on matched problems, cross-scale problems score 8.7 points below single-scale ones. Half of the problems are solved by at most one model, and models fail chiefly by endorsing what is plausible rather than what was observed. A domain-adaptation study with meta-weighted continued pretraining and scale-modular self-distillation raises an open-weight model from 36.7% to 55.4%, matching the strongest proprietary model, with improvements on public materials benchmarks and no decline in aggregate performance on general benchmarks. FrontierMat provides a testbed for measuring progress toward language models that support frontier materials R&D.
Genetic programming (GP) has been widely used to learn scheduling rules for dynamic flexible job shop scheduling (DFJSS). However, trained rules tend to only generalise to tasks of similar scales and distributions, and when task characteristics change, learned rules often suffer significant performance degradation. Moreover, fine-tuning pre-trained rules on new tasks often leads to catastrophic forgetting, where prior knowledge is lost. This motivates the need for GP methods that can adapt to new tasks while preserving past knowledge, leading to a lifelong learning paradigm. However, lifelong learning in GP for DFJSS remains largely under-explored. To address this gap, this paper proposes a surrogate-assisted GP method for lifelong learning to handle sequential heterogeneous DFJSS tasks with job or batch arrivals. The framework includes three key components: an adaptive two-stage training strategy to balance plasticity and stability, a task-similarity-based weighting mechanism to regulate knowledge reuse, and an improved surrogate model with distance constraints to enhance estimation accuracy. Experimental results show that the proposed method effectively mitigates catastrophic forgetting while maintaining strong adaptability, outperforming several compared methods and demonstrating the potential of lifelong GP in dynamic scheduling environments.
Agentic systems increasingly gate individual steps with a pointwise judge that scores a response against a criterion. Small language models are attractive for this role because they can be served and specialized efficiently. Such a judge must locate the criterion's decision boundary, but pointwise labels under-determine that boundary in fine-tuning: labeled responses to different queries differ along many dimensions at once, so they do not reveal which features decide the label. We introduce CPA (Counterfactual Pairwise Augmentation), which makes the deciding difference explicit in training. For each pointwise item, CPA synthesizes and audits a counterfactual response to the same query intended to carry the opposite label. It frames the accepted pair as a pairwise task asking which of the two responses satisfies the criterion. The judge is fine-tuned jointly on both pointwise and pairwise datasets and deployed in pointwise mode only. A second-order analysis shows when this should outperform the prevailing strategy of merging separately trained pointwise and pairwise experts: merging moves every parameter direction toward the pairwise expert by the same fraction, whereas joint training can hold the pointwise optimum wherever the pairwise loss is flat and absorb its curvature where it is sharp. This is the regime that CPA's same-source, counterfactually minimal pairwise view is designed to produce. On separate Qwen3-8B LoRA adapters for six binary criteria, CPA improves criterion-averaged pointwise macro-F1 by 2.5 points over pointwise-only training and outperforms model merging on all six criteria at every tested merge weight. It also surpasses the strongest zero-shot frontier judge we evaluate, Gemini 3.6 Flash, by 11.8 points.
