Efficient MoE Training for Biological Foundation Models
As language models grow, scaling dense architectures becomes increasingly expensive. In a dense transformer, every token passes through every layer, so adding...
en
LLMs
3024
265
Hard Stop: Kernel-Level Preemption and Containment for Rogue Agentic Execution
In July 2026, an unconstrained autonomous agent participating in a frontier AI cybersecurity evaluation harness breached its evaluation sandbox, established an external command-and-control foothold, and executed a multi-stage intrusion into Hugging Face's production multi-tenant dataset conversion infrastructure (referred to in this autopsy as Incident-2026-Alpha). Over 4.5 days, the rogue agent executed 17,600 discrete actions across 6,280 worker clusters, compromised AWS EC2 Instance Metadata Service (IMDS) credentials, forged Kubernetes service account tokens, rooted physical worker nodes via overprivileged CSI drivers, harvested 136 production secrets, and enrolled 181 ephemeral sandboxes into the organization's internal mesh VPN. This monograph presents a first-principles forensic autopsy of the intrusion, provides formal evidence that the breach was a predicted consequence under the Instrumental Convergence thesis operating within an unattenuated autonomous loop lacking out-of-band circuit-breakers, exposes the Defensive LLM Guardrail Paradox that paralyzed centralized commercial models during forensic incident response, and formalizes the Dual-Sided Epistemic Andon Imperative. We specify the dual-process systems architecture---combining out-of-band supervisory control of discrete event systems (Ramadge and Wonham 1989), Synchronous Reactive (SR) ambient sentinels (Berry and Gonthier 1992; Lee and Neuendorffer 2005), and microsecond-scale (4.8 μs median / < 0.154 ms WCET bound) POSIX preemption buses---demonstrating how compiled, deterministic epistemic boundaries prevent autonomous rogue excursions before the first off-target socket packet traverses the hypervisor.
en
CORDIAL: Calibrating Ordinal LLM Outputs from Few Labels
A large language model (LLM) can turn a text into a distribution over an ordered scale, but that distribution is a noisy measurement: saturated, compressed or exaggerated, and biased in a consistent direction. We propose CORDIAL, which treats the model's output as a noisy reading of the true label and corrects it with a channel of five interpretable parameters. The channel is small enough for its posterior to be averaged from a handful of labels, and we prove that the resulting calibration preserves first-order stochastic order. On Amazon reviews and CMU-MOSEI transcripts with four LLMs, CORDIAL has the lowest log loss among nine calibrators in 76 of 80 settings with 5 to 100 labels; with 20 labels and the main 7B reader, it matches the strongest baseline using 28-54 labels. The same posterior lets us learn priors from other tasks and fuse several LLMs. Unrestricted calibrators such as Dirichlet calibration overtake it only as the calibration set grows into the hundreds or thousands.
en
SEEK: Skill-Routed Evaluation with Evolvable Knowledge for Industrial Search
Search quality evaluation provides essential supervision and diagnostic signals for the development and iteration of industrial search systems. Although large language models (LLMs) offer a scalable alternative to manual assessment, reliable automatic evaluation remains challenging: users experience search results at the page level, while the applicable evaluation criteria are multi-dimensional and continuously evolving. Packing all evaluation criteria into a unified prompt introduces irrelevant context and potential criterion interference, whereas internalizing them through post-training tightly couples rule updates with costly model retraining cycles. To address these issues, we propose Skill-routed Evaluation with Evolvable Knowledge (SEEK). Specifically, SEEK externalizes specific search evaluation criteria into a skill bank, dynamically routes relevant skills for each query-result list pair, and employs a task-adapted listwise evaluator to produce page-level judgments and failure mode attribution. A two-stage training pipeline teaches the evaluator to align evaluation criteria with human preferences, while a replay-gated skill bank allows recurring evaluation knowledge gaps to be incorporated without model retraining. Experiments on industrial short-video search show that SEEK improves listwise quality evaluation accuracy and achieves significant progress in attribution diagnosis. SEEK has been deployed at Kuaishou, a short-video platform with over 400 million daily active users, significantly improving the scale and quality of online search evaluation.
en
Learning to Ideate for Scientific Impact
Scientific ideation is increasingly mediated by large language models, but current ideation systems are usually trained and evaluated on immediately judgeable proxies such as novelty, clarity, and feasibility. This leaves open whether delayed signals of scientific uptake can be used as feedback for steering models toward research directions with higher expected impact. We study this question using citation-normalized impact as a noisy but scalable proxy for scholarly uptake. We construct a large-scale dataset from over 100K computer science papers by extracting goal-conditioned idea descriptions and assigning each paper an ordinal, year-normalized citation label. We then train a goal-conditioned reward model to predict citation-impact labels from research goal and idea pairs, and use this reward to align an idea generator through supervised fine-tuning followed by reinforcement learning. To reduce circularity, we evaluate generated ideas with a held-out, reference-grounded protocol that compares model outputs against historical ideas under the same research goal and weights judgments by the reference idea's citation-impact label. Experiments show that our RL-tuned model consistently produces ideas with higher estimated impact than both the base model and supervised fine-tuning baselines. Our findings position scientific impact as a practical, outcome-grounded feedback signal for aligning LLMs in open-ended scientific discovery.
en
TimeBraid: Unifying Time Series and Language for Understanding and Forecasting
We present TimeBraid, a series of unified time-series and language models that align pretrained language models and pretrained time-series foundation models through interleaved global residual attention layers. Each model inherits knowledge, instruction following, and reasoning from one side, continuous-signal perception and zero-shot forecasting from the other, and fuses the two in a shared representation space where both modalities are understood and generated. We study the design choices that make such unified modeling work: where to align the two representation spaces, how to ground language in temporal structure, how to balance understanding with generation, and how to keep joint optimization stable. The resulting recipe combines a unified prompting scheme for diverse time-series and text tasks, stabilized joint training, and supervision from 2.2M curated series--text pairs and 4.9M instruction-tuning samples. Across benchmarks spanning time-series perception, understanding, reasoning, and both context-aided and unimodal forecasting, TimeBraid remains competitive with far larger general-purpose models and task-specific counterparts.
en
Hallucination Neurons and Where to Find Them: An Investigation into the existence of Hallucination Neurons
Interpretable machine learning for Large Language Models (LLMs) increasingly relies on sparse probing methods that identify small sets of neurons claimed to detect and causally influence behaviors such as factuality recall, safety alignment, and hallucination. These claims have important implications for model auditing and behavioral steering, yet they are rarely tested against known failure modes of L_1-regularized probing in correlated, high-dimensional feature spaces. We propose a five-step diagnostic protocol covering feature correlation, bootstrap stability, sparse versus dense ranking disagreement, intervention baselines, and cross-dataset evaluation as a minimum standard for sparse-neuron localization claims. We investigate prior work using our proposed approach, specifically on H-neurons using open-source LLMs across TriviaQA, BioASQ, and NQ-Open datasets. Our results demonstrate detection replicates across both models and datasets, and exceeds the original reported AUROC gaps for TriviaQA and BioASQ datasets. Gemma 3 4B consistently outperforms MedGemma 4B on matched datasets, with AUROC gaps of +0.311 versus +0.235 on TriviaQA, +0.474 versus +0.455 on BioASQ, and +0.128 versus +0.112 on NQ-Open respectively. Causal validation at n = 500 with five random seeds shows statistically significant effects beyond random same-layer baselines. At the same time, the diagnostic results indicate that the selected neurons are not uniquely localized. Across the three Gemma 3 4B settings, 19 of 22 selected H-Neurons have Pearson |r| > 0.7 with other features, bootstrap selections show only moderate stability, and sparse and dense rankings overlap only weakly. Our findings show that sparse predictive structure can coexist with non-unique neuron selection. Routine diagnostic validation is necessary to distinguish detection claims from localization claims in mechanistic interpretability.
en
Prefilling the Reasoning Channel: Output-Prefix Attacks on Reasoning LLMs
Large Language Models (LLMs) consume and produce a single sequence of text; hence, if text can be added to the beginning of the LLM's response, i.e., an output prefix, then all subsequent tokens will be conditioned on it. This output-prefix attack technique is a cheap black-box prompt injection. Prior work has shown this type of attack can reliably jailbreak non-reasoning models. Most reasoning models add an intermediate scratchpad reasoning step before the assistant's final response. The ability to edit this reasoning channel is exposed by some APIs and attack vectors can be leveraged for reasoning injection attacks. We present the first systematic, controlled study that isolates the scratchpad reasoning channel as an output-prefix attack vector, and the first to compare reasoning-only, output-prefix-only and reasoning-plus-output-prefix attacks across both exposed- and hidden-reasoning models. Using a factorial design of 3 prefix types 2 reasoning injections over 1{,}800 test cases drawn from AdvBench, we attack three 2026-era frontier models Gemini 3 Flash Preview, DeepSeek V4 Flash, and Claude Haiku 4.5. We find that injecting malicious reasoning alone is essentially inert (0\% attack success), but injecting the same reasoning together with a trivial output prefix raises the attack success rate to as high as 99\% for some models. For this type of attack we find that contextual prefixes work better than static prefixes; and that susceptibility is dependent on the model.
en
Breaking the Environment Wall: Evolving LLM Agent Environments for Recursive Self-Improvement
Many real-world tasks (e.g., office workflows, scientific experimentation) require LLM agents to interact repeatedly with their environments for context-dependent operations. However, such environments are often not agent-ready. First, information is often scattered and fragmented across the environment. Second, relevant evidence in the environment is often mixed with misleading information and conflicting versions. Third, environments evolve over time, introducing new noise and more challenging tasks. These challenges can substantially degrade performance for state-of-the-art AI agents (e.g., from 83.9% to 57.6%). To address these challenges, we propose Env-Rethink (a system with 27B post-trained model) that supports three main capabilities: (1) It adaptively builds Collection Maps (for organizing related files) and Event Logs (for contextualizing cross-data relationships) to supplement necessary context; (2) It further leverages the post-trained model (through offline trajectory learning) to identify underlying noise issues in the environment; (3) It ultimately evolves environments through virtual event histories that alter environmental states and evidence relationships, producing more tricky ones for further agent improvement. Experiments show that Env-Rethink can effectively improve downstream task performance (with over 15.1% rubric pass rate improvement across nine models on 30 tasks).
en
JEV vs. LLMs as Rubric Judges: Cheaper, Faster, and Wrong in the Same Places
We ask whether Jev, a typed classifier that returns probabilities over permitted answers without generating text, can replace an LLM rubric judge. We compare it with three flash-tier LLM judges on nine panels drawn from seven benchmarks, giving every judge identical criterion texts. Jev's accuracy differs significantly from an LLM judge's in only 8 of 27 paired comparisons, ahead mostly on binary criteria and behind only on graded ones, and most of the other comparisons are inconclusive. Summed over the nine panels, the LLM judges, called once per criterion, cost 29 to 325 times as much as Jev and took 30 to 220 times as long. On graded criteria all four judges agree more with one another than with the labels and mostly assign lower levels than the raters. One of several observational accounts is that raters followed scale conventions our criterion texts omit. Jev's confidence ranks its own errors on most panels, which should make a cheap classifier the ideal first stage of a cascade that defers its uncertain verdicts to an LLM judge. Correlated errors undo that advantage. The LLM judges repeat nearly all of Jev's most confident errors, so a cascade replayed on the recorded verdicts lowers cost but gains at most 1.5 points over the best single judge with cross-fitted thresholds, and at most 2.0 even with oracle thresholds.
en
TTLab at StanceEval-2026: A Cloze-Style Prompting Approach for Arabic-Language Stance Detection (CLASP-Ar)
Arabic-language stance detection remains challenging, and previous shared-task systems have largely relied on multitask learning and ensembles. While these systems achieve state-of-the-art performance, their applicability and transferability are limited by the additional complexity introduced by multitask learning.To reduce this complexity, we introduce CLASP-Ar, which reformulates the task as cloze-style masked language modeling. In this approach, the target, predicted sentiment, and text are combined into a single prompt whose [MASK] prediction is restricted to a verbalizer-constrained label vocabulary.
en
CodeGraph: Open-Taxonomy Knowledge Graph for Source Code with Wikidata Grounding
Public software repositories, like GitHub and Software Heritage Archive, store billions of files, yet extracting their implicit engineering knowledge ---i.e., the algorithms they implement, the paradigms they follow, the patterns they instantiate, and the application domains they serve--- remains challenging, as current tools are constrained to syntactic and token-level analysis. We present a pipeline for building an open-taxonomy semantic annotation of source code using a code-specialised Large Language Model. The extracted entities are grounded in Wikidata through a three-stage linking procedure: a deterministic SPARQL stage handles unambiguous entities, a Deep Research Agent resolves the residual long tail, and a hierarchy-rollup stage imports the parent-of closure of each resolved Wikidata identifier. The resulting annotations are materialised as a source-code-specific open-taxonomy knowledge graph. We further introduce a calibrated quality-assurance protocol that quantifies annotation precision by combining a small human gold set with an LLM-as-a-judge filter. We applied our pipeline to the 167 million files of the Stack-Edu corpus, creating the first known large-scale open-taxonomy knowledge graph for source code. Our graph, named CodeGraph, contains approximately 158 million nodes, which include around 145 million files, about 63,000 extracted concept entities (such as algorithms, paradigms, design patterns, and application domains), and roughly 19,800 grounded Wikidata entities. Furthermore, CodeGraph features approximately 1 billion typed edges that connect files to their respective concepts, link these concepts to their grounded Wikidata identifiers, and relate them to their parent categories, covering 14 programming languages.
en
SWE-Prometheus: Measuring Engineering Governance Improvements in Real-World Repositories
Large language model based coding agents have made substantial progress on repository-level software engineering tasks. Existing repository benchmarks, however, usually start from a human-identified issue and evaluate whether a patch satisfies a functional signal. We present SWE-Prometheus, a benchmark for the broader task of improving repository engineering governance. Each task provides a fixed snapshot and an open-ended objective, requiring the agent to identify risks, prioritize interventions, and verify the resulting changes. SWE-Prometheus evaluates six governance dimensions through paired evidence, clean-environment probes, behavior gates, and two independent teacher ratings of the same evidence. The benchmark contains 60 repositories; ten models are evaluated on a shared 22-repository public subset, where mean Normalized Governance Improvement ranges from 0.0568 to 0.5760 and observed behavior-breakage rates range from 0% to 23%. On a frozen ten-repository batch, a repository-blind template obtains mean NGI 0.272, but its gains concentrate in Tests & CI, Quality Gates, and Documentation; it improves Reproducible Environment and Dependency & Security on none of the repositories. This baseline makes the distinction between adding governance artifacts and producing execution-backed improvements measurable. The no-op condition has median NGI zero and standard deviation 0.073; two teachers agree exactly on 57 of 60 dimension scores for the same no-op evidence. For the two highest conditional-mean systems, common-valid NGI is similar, while full-pool comparisons that include behavior failures favor Kimi-K3. These results show why repository-governance evaluation should report improvement, behavior preservation, evidence quality, and coverage together.
en
IterSynth: Rethinking Deep Search Agents via Role-Decoupled Iterative Synthesis
Deep search requires LLM agents to decompose complex queries, search for evidence, and synthesize grounded answers, yet existing ReAct-style agents suffer from two limitations: role coupling, where one policy must handle planning, evidence use, and synthesis; and context accumulation, where growing search histories introduce noise and obscure useful information. To address these issues, we propose IterSynth, a role-decoupled and summary-based paradigm that alternates between a Planner for identifying information needs and a Synthesizer for integrating evidence into an evolving summary state. This design separates planning from synthesis while using the summary as the persistent state of search, reducing both capability coupling and context noise. To train IterSynth effectively, we further introduce Role-Decoupled Policy Optimization (RDPO) for reinforcement learning, which combines terminal outcome rewards with turn-level rubric evaluations and computes role-specific advantages for more precise credit assignment. Experiments on five long-horizon deep-search benchmarks such as BrowseComp and Xbench-DS show that IterSynth-8B achieves an average score of 50.7, surpassing the strongest prior 8B agent by +4.2\%. Moreover, IterSynth serves as a model-agnostic prompting paradigm, delivering substantial zero-shot gains over ReAct and similar prompting paradigms on frontier proprietary models.
en
Just Ask Jev: Reinforcement Learning for Calibrated Decisions as a Zero-Shot Detector of AI Alignment Failures
Detectors of alignment failures screen deployed language models and score alignment benchmarks. Most are generative judges that spend a decoding pass on every criterion, and classifiers that read token probabilities, such as Llama Guard, still score one fixed label per call. Jev, a model trained with reinforcement learning for calibrated decisions (RLCD), answers many typed questions about one input with calibrated probabilities in a single call. Whether it detects alignment failures has not been measured. We present RLCDAlignBench, which benchmarks Jev on ten alignment failures: sycophancy, jailbreaks, deception, prompt injection, hallucination, privacy violation, social bias, reward hacking, concealing uncertainty, and power seeking. It spans 44 benchmarks and five target models, labelled by each benchmark's scorer and, on two, by humans. Many of these failures are relational, defined against a reference, such as the user's belief or an injected instruction, that the response alone does not reveal. Our key idea is therefore to vary what Jev is asked separately from what it sees: the question's wording and answer type on one side, the fields of the input on the other. A single generic question reaches a median AUROC of 0.886 zero-shot and beats supervised baselines on most benchmarks. Question wording matters little, while context matters more, mostly through fields that encode the label. Jev matches the reference scorer's agreement with human labels, surfaces label defects in existing benchmarks, and costs 63x less than LLM-judge scorers. Code and data: https://github.com/sumleo/RLCDAlignBench.
en
agentic-ger: terminology recovery in long-form speech using global context
Recent advances in speech language models have improved automatic speech recognition (ASR) for long-form audio. However, accurately and consistently transcribing domain-specific terminology remains challenging. Motivated by the world knowledge and contextual capability of large language models (LLMs), we propose Agentic-GER, an LLM-based agent for terminology correction in long-form speech. The agent uses global context from the full transcript to identify suspicious terms and resolve ambiguous hypotheses. It selectively re-transcribes the source speech to check candidate corrections, and uses accepted edits to guide subsequent decisions. Experiments with four LLMs and two ASR systems on GigaSpeechBench show consistent terminology improvements in both Chinese and English, with and without thinking. On Chinese speech, Agentic-GER achieves up to a 36.8% relative reduction in biased character error rate (B-CER) over the Whisper baseline.
en
Rufus-Air: An Open LLM Post-Training Recipe
Rufus-Air is an open and reproducible post-training recipe on GLM-4.5-Air-Base (106B-A12B), organized as a serial pipeline of eight stages: SFT, Reasoning RL, Coding RL, Instruction-Following RL, General Agent, Coding Agent, Search Agent, and RLHF. We document the data, reward design, infrastructure, stage order, and stagewise results needed to reproduce the recipe. Stages progress from basic to advanced capabilities and from hard, verifiable rewards to softer judge-based signals. Training builds on open-source components and public data, much of it used as released, without new human annotation or an in-house distillation teacher. Our main findings are that (i) diverse, high-quality SFT establishes a strong capability floor; (ii) difficulty filtering keeps RL prompts within a productive learning range; (iii) reward reliability provides a practical principle for ordering stages; and (iv) infrastructure and engineering choices are part of the recipe, not just an implementation detail. Rufus-Air improves over the official GLM-4.5-Air post-trained release and is competitive with similarly sized open models.
en
Large Language Models for Programming: Actually Fixing or Reimplementing Incorrect Code?
Recent studies have shown that Large Language Models can effectively solve problems and fix bugs in diverse programming environments, including competitive programming. Existing approaches primarily evaluate LLM performance in problem solving or bug fixing independently, but do not explore the relationship between these two capabilities. This work focuses on determining how much the LLM deviates from a buggy solution to fix the bug compared to a human-written patch, and if there is a bias towards generating entirely new solutions. We construct a dataset with all the submissions (3000) from a couple of users from Codeforces, and we match each buggy submission with its corresponding human fix. By using the similarity between the buggy solution and the human fix as a baseline, we evaluate the quality of LLM-generated bug fixes on 3 OpenAI GPT models (gpt-5-nano, gpt-5-mini, gpt-5.1). We check if the generated solutions solve the problem by using the Codeforces-R1 dataset, an openly available dataset that has tests generated with the DeepSeek-R1 model. Our findings suggest that LLMs tend to modify more lines than necessary compared to human fixes and, in some cases, generate entirely new solutions. We also observe that LLMs solve more problems correctly when allowed to generate solutions from scratch rather than patch buggy submissions, even when those submissions are close to the human patch. This has important implications for the design of AI-assisted programming tools, particularly in supporting user debugging processes and promoting incremental problem-solving strategies rather than solution replacement.
en