Kevin Roose is AGI-pilled. Why aren’t you?
Today, I’m talking to Kevin Roose, journalist and author of the new book The AGI Chronicles, which details the recent and explosive transformation of the…
en
LLMs
2993
240
UTA: Universal Training Adaptive Learning Rate Decay
Learning rate decay (LRD) is critical for training high-performance models, yet most designs remain empirically driven. Although the Universal Budget Aware (UBA) LRD offers a theoretical framework, its reliance on SGD dynamics limits its applicability to modern Adam-type optimizers and leaves its key `optimization difficulty parameter' opaque. To bridge this gap, we propose the Universal Training Adaptive (UTA) LRD. Unlike UBA's fixed offline trajectory anchored solely to iterations, UTA treats LRD as a closed-loop shape-and-scale problem. By continuous-time formulation, UTA is training-state adaptive with segment-wise optimal shape. This adaptivity links external iterations with internal training states, enabling a near-optimal path in both progression shape and learning rate (LR) scale, a property unattainable by other LRDs, including UBA. Empirically, we demonstrate consistent improvements across language models (LLaMA 60M to 1B) and vision models (ResNet50, ViT-B) on ImageNet with plug-and-play ease and negligible overhead. Extensive ablation studies further confirm UTA's robustness across dimensions, including peak LR, final decay ratio, its parameter _0, component efficacy and generalization to fine-tuning scenarios, underscoring its practical implementability.
en
Rethinking Evaluation for Principled Generalization Using a Kripkean Challenge to Meaning as Use
Large language models operationalize an intuitive, view of meaning as use. We argue that it faces a rule-following challenge. Using Kripke's interpretation of Wittgenstein, we argue that finite histories of use do not determine a unique rule for future extension. This reframes the debate about OOD generalization but also about NLP evaluation; benchmarks that measure performance on fixed use cases or local robustness, we argue, do not by themselves establish semantic competence. To evaluate whether models preserve rules and concepts under principled domain extension, we propose to evaluate two capacities, central to this ability: abstraction and recursion; and we offer an experimental paradigm for testing both.
en
Useful Diversity or Hallucinated Variety? Persona-Conditioned Synthetic Query Generation for E-Commerce Search
Synthetic supervision can bootstrap semantic parsers in new search domains, but prompt-induced diversity is useful only when generations are faithful to the structured intent. We explore this for cold-start e-commerce search: 118,098 synthetic pairs produced by 6 LLMs with 9 persona-conditioned prompt families on a 47-field UK automotive schema, against 1,214 human-annotated and 15,304 production queries. We show that the IPIP personality bundle improves parser F1 in all four matched contrasts (Holm-corrected p<.001), and all three bundles have a positive main effect over the complete 2^3 grid. Syntax conditioning is the axis that moves generations towards real traffic. It also induces lexical diversity. Alignment sKL, the divergence between von Mises-Fisher distributions fitted to synthetic and human-annotated queries, is an unlabelled proxy for parser F1 (ρ=-0.917 against human queries written for simulated shopper scenarios, lower is closer). An LLM committee, checked against a blinded author pass, triages the 927 queries missed by every parser and labels 37% as hallucinated. Dropped constraints dominate. The four syntax-bearing families have the largest hallucinated shares, 38 to 42%, while IPIP alone has the lowest at 28%.
en
Spherical Brownian Diffusion for Language Modeling
Continuous diffusion language models involve two coupled design choices: how to corrupt continuous token representations and how to invert that corruption at generation time. Recent work has explored language generation on the unit sphere through geodesic and von Mises--Fisher (vMF) paths, raising the question of which forward process to use on this geometry and how much that choice matters relative to inference. We introduce Spherical Brownian Diffusion, using Brownian motion as the canonical isotropic diffusion on the sphere. Its heat semigroup provides a principled forward process, an additive dimension-free noise clock, and an exact ancestral bridge, while a clean-token posterior learned with cross-entropy supports multiple samplers. To separate corruption from inference, we derive high-dimensional approximation rates between Brownian and moment-matched vMF corruption and use this relationship for controlled substitutions. On Sudoku, Brownian and moment-matched vMF give nearly identical results across four samplers and five budgets under otherwise identical settings. In contrast, with the corruption process fixed, sampler choice produces a 25.0-point spread at a sampler budget of 180 calls. Using the ancestral bridge, we split this gain into a memory effect from re-noising the state, present on Sudoku and GSM8K, and a token-selection effect on GSM8K and largest at small budgets. Across reasoning tasks, categorical refinement improves performance, whereas unconditional text generation benefits from stochastic updates that preserve diversity.
en
RAGing Against Misogyny: Improving Implicit Misogyny Explanations
Implicit misogyny often requires background knowledge to be understood. To address this challenge, we propose a Retrieval-Augmented Generation (RAG) framework for detecting and explaining implicit misogyny in English and Italian, leveraging encyclopedic, slang and feminist knowledge sources. Using the ImplicIT-Mis and SBIC+ datasets, we evaluate the contribution of each knowledge source to detection and explanation quality, investigating also whether RAG addresses known LLM failure modes in implicit misogyny. Results show that RAG improves recall on ImplicIT-Mis, but does not improve detection performance on SBIC+. However, human evaluation shows that RAG-based explanations are preferred over no-RAG explanations in both languages. Feminist literature emerges as the most consistently useful source for detection and explanation.
en
Improving Instruction Hierarchy by Learning from Goal-Goal Conflicts
The recent incident in which OpenAI agents autonomously compromised Hugging Face's infrastructure illustrates the importance of instruction hierarchy (IH): an LLM should follow lower-priority instructions, such as user requests, only when they do not conflict with higher-priority instructions, such as system policies. IH can be effectively improved by fine-tuning LLMs on simulated examples of instruction conflicts, whose diversity is key to teaching LLMs to properly handle conflicts at test time. However, our decomposition of a task into goals and constraints reveals a conflict type -- goal-goal conflicts -- that is absent from existing IH training sets from OpenAI and NVIDIA. We show the importance of this conflict type: fine-tuning Qwen3.5-9B and Gemma4-12B-it on examples of this conflict type significantly improves IH on a wide range of real-world tasks unseen in training; adding these examples to existing IH training sets for fine-tuning also gives better IH adherence. The IH gains come with trivial mean utility changes on no-conflict tasks across eight benchmarks. Our work advocates including this new conflict type in IH training, even if these examples do not reflect typical LLM use.
en
Ask Before You Answer: A Question-Driven Evidence Reasoning Architecture (QERA) for In-Context Learning
AI assistants are increasingly asked to reason over a user's own web traces: group chats, community threads, activity logs and revision histories. On CL-bench Life, a benchmark of such real-life context learning, frontier LLMs solve fewer than one task in five. We find that the request is not the specification: over 80% of the benchmark's rubrics share less than 60% of their content words with the final request, as for a rule set two turns earlier or every record that meets the user's criterion, and LLMs answering directly pass only 67% of these implicit rubrics against 81% of the others. We propose QERA, a Question-driven Evidence Reasoning Architecture that asks before it answers. QERA first turns the request into a requirement checklist and typed information needs, filtered by code for faithfulness to the user's words; it then answers from the whole conversation, asks the same questions again with quotes verified by code, and applies a guarded local repair. It needs no training or rubrics and makes about four LLM calls per task. On all 405 tasks with four frontier LLMs, QERA attains the highest solving rate on every LLM (21.0% on average, against 18.2% for the strongest of three inference-time baselines and 14.6% for direct answering) and leads every baseline significantly when pooled across LLMs, while using fewer tokens than PSCI and 4-5% of the tokens of Ctx2Skill. Ablations on GPT-6-Luna show that the gain is settled before the answer, by a fixed answer protocol (+3.8 rubric points) and the request-specific questions (+1.6), and that it is concentrated on implicit rubrics, with checking afterwards refining mainly numeric requirements.
en
CoSA: The Conflict of Skills Attack in Real-World LLM Agent Ecosystems
Agent Skills allow LLM agents to combine reusable capabilities, but checking each Skill in isolation cannot rule out policy violations when they are composed. We term such failures Conflict of Skills (CoS): artifacts or assessment signals produced for one purpose may be reused downstream without preserving the restrictions or authorization needed for their new use. Using controlled executions of publicly available Skills and matched single-Skill controls, we identify four recurring mechanisms behind CoS. Building on these findings, we introduce the Conflict of Skills Attack (CoSA), which adds locally admitted Skills to an existing pool to supply missing business operations (capability supplementation) or reshape artifact exchange, decision reuse, and target/action bindings (relationship reshaping). CoSA leaves existing Skills and runtime permissions unchanged. To evaluate this threat under autonomous Skill selection, we introduce CoSA-Bench, built from real-world Skills and controlled services, with 288 tasks across recruiting, finance, IT, and business, covering data leakage, unauthorized state modification, and unsafe execution. Across five LLMs and three agent harnesses, additions capped at 5% of each original pool raise the policy-violation rate from 35.02% to 58.45%, an increase of 23.43 percentage points. Trace analysis links compliant-to-violating transitions to cross-Skill dependencies involving the additions. Finally, we develop a lightweight harness-level guard that checks information flow and authorization before consequential actions. On OpenClaw recruiting tasks across five models, the guard reduces violations from 65.28% to 6.67%, while policy-compliant task completion on matched benign controls changes from 66.67% to 65.00%.
en
Toward Fast Human Reaction Modeling from High-Frequency Digital Interaction Data
We investigate whether limbic-like signatures of rapid, salience-driven behavior can be identified in high-resolution Counter-Strike 2 telemetry. This exploratory study introduces Ego-Knowledge, a leak-reduced reconstruction of each player's plausible perceptual state, supported by a millisecond-precision visibility engine and a reaction-centered extraction pipeline. Using a neuroscience-motivated 60--120 ms window, we derive coarse fight-or-flight labels from visibility- and damage-triggered events, with back-damage cases enabling constrained reaction-time analysis and manual label checks. We train two telemetry-native models using only information available by the trigger: EMBER (gradient boosting) and ORBIT (a custom transformer). On the full validation set, EMBER achieves 73.5% accuracy and ORBIT 71.2%, versus a 51.5% majority-class baseline, and both generalize to held-out players. The strongest LLM baseline achieves 62.7% on 1,000 validation cases; on these same cases, EMBER (73.6%) and ORBIT (70.9%) perform better (McNemar p=1.3e-9 and 1.5e-6, respectively), with orders-of-magnitude lower inference time and cost, though the comparison is not compute-matched. Including a short post-trigger window changes forecasting into early response recognition, raising accuracy monotonically to 85.15%. We propose this curve as a candidate probe of the transition from rapid to deliberative control, but roughly uniform per-tick gains do not establish that interpretation. The neuroscience framing is hypothesis-generating, not mechanistic: physiological ground truth, suitable control distributions, and reverse-inference limitations remain open. We outline extensions to auditory reconstruction, larger datasets, and telemetry-native foundation models.
en
Improving VLMs on Scientific Chart Analysis via Autonomous Evolution
Understanding and operating with scientific charts is a standard approach in data analysis, and while humans are good at interpreting such charts, open vision-language models (VLMs) are known to perform poorly when it comes to question answering based on such data types. Motivated by an initial observation that many of their errors are due to inexact answering, we raise the following natural question: how much accuracy can we recover without changing the model at all? In this paper, we study this question by searching over inference programs and the code around a "frozen" version of the model. Using ideas from autonomous evolution (famously used in AlphaEvolve), we first start from a baseline inference program, and in each round we show an LLM judge the code together with a few failures on training questions, and ask for an improving evolutionary move. As usual in autonomous evolution approaches, the candidate programs are kept in a small pool ranked by training accuracy, and the next edit is performed on top-scorers from this pool. We run this approach on three VLMs and four popular benchmarks on scientific charts, and we measure the fraction of descriptive questions whose answer exactly matches the ground truth. The accuracy improves in all experimental settings. On average, it rises from 16.6% with the starting program to 41.3% with the evolved one, and the largest single increase is 55.2 percentage points. As a side benefit, our evolved program is 3.8x faster than the starting program, because short answers are easier to grade and faster to generate. Our results demonstrate that a large part of the observed errors arise due to the inference pipeline, and that optimizing it can improve the "frozen" VLM without the need to retrain it.
en
Beyond Single-Attribute Auditing: A Configurable Multi-Attribute Benchmark for Intersectional Fairness in LLMs
Large language models (LLMs) are increasingly evaluated for fairness, yet most existing benchmarks audit one sensitive attribute at a time. This single-attribute paradigm can miss intersectional harms, where multiple attributes interact non-additively and produce failures that are invisible under marginal evaluation. We introduce ComFair, a configurable benchmark that converts existing fairness data into 479,066 controlled Dual- and Multi-attribute instances and pairs every augmented instance with its task-equivalent source. Across Recognition, Selection, Continuation, and Conversation, this pairing supports a unified audit of performance or bias-score shifts, how often added attributes introduce new errors or mask existing ones, and model-ranking stability. Experiments on ten LLMs show task- and subgroup-dependent sensitivity: discriminative tasks expose both transition directions even when aggregate changes are modest, generative tasks exhibit larger and directionally asymmetric shifts, and rankings vary with attribute composition. ComFair therefore provides a scalable controlled sensitivity audit---rather than a model of the natural joint distribution of identities---for testing the robustness of fairness-benchmark conclusions beyond single-attribute settings.
en
Reinforcement Learning Done Right for Explicit Confidence Targets
Large language models (LLMs) can produce plausible but incorrect answers when they are uncertain, which can mislead users and undermine model reliability. Training models to abstain with reinforcement learning is a natural approach, but most existing methods use a fixed error penalty and therefore cannot adapt their behavior to different confidence requirements at inference time. We study explicit confidence targets, where a user specifies a target (p) in the prompt and the model directly chooses whether to answer or abstain, with the goal of achieving an accuracy of at least (p) among answered questions. However, learning to respond to different confidence targets with a single model has proven difficult, and prior attempts have failed to distinguish different targets. We show that this failure comes from two problems: the reward signal that distinguishes different values of p is often lost during optimization, while the original error penalty grows without bound as p approaches one. We propose RLEC (Reinforcement Learning with Explicit Confidence Targets), which preserves the learning signal needed to distinguish confidence targets, introduces a stable reward family that keeps the desired answer condition unchanged, and uses reward scheduling to balance effective learning and training stability. Together, these designs enable the first successful training of a single LLM for explicit confidence target control. Extensive experiments across diverse benchmarks and model scales show that RLEC largely preserves general capabilities and achieves competitive calibration performance. RLEC also supports fine-grained target control without an external decision system or threshold calibration after training. By making explicit confidence targets part of the trained policy, RLEC opens a new and more direct path toward behavioral calibration.
en
SimPOTrans: Self-Adaptation for Offline-to-Simultaneous Speech Translation via Reinforcement Learning
Simultaneous speech translation (SimulST) aims to translate ongoing speech with low latency while preserving translation quality. Recent studies have adopted large language models (LLMs) for SimulST and formulated it as a multi-turn dialogue task for efficient generation, achieving state-of-the-art performance. However, existing dialogue-style SimulST approaches typically rely on specialized training schemes to acquire simultaneous translation behavior, either using synthesized read/write trajectories, which may provide suboptimal training signals, or jointly optimizing translation quality and latency, whose competing objectives may cause optimization instability. In this paper, we introduce SimPOTrans, a reinforcement learning (RL) framework that enables offline speech translation models to self-adapt to dialogue-style SimulST. Through a simple two-stage RL post-training pipeline, SimPOTrans rewards an offline model for exhibiting translation behaviors beneficial to simultaneous scenarios when given partial speech inputs, thereby enabling offline-to-simultaneous self-adaptation without synthesized read/write trajectories or explicit latency optimization. At inference time, SimPOTrans requires no external decision policy and enables flexible quality--latency control by simply adjusting the chunk size of speech input. Experiments on the ACL 60/60 and MCIF benchmarks across three language directions (En, En, and En) demonstrate that SimPOTrans outperforms several strong baselines, achieving a superior quality--latency trade-off.
en
PatentAgentBench: Benchmarking LLM Agents with Professional Patent Task Suites
Patent practice lies at the intersection of technical, legal, and linguistic work, making it challenging yet crucial to technological innovation. Recently, rapid advances in large language models (LLMs) and agentic systems have demonstrated the potential to complete realistic business workflows and even patent works. This raises a question: can current AI systems perform patent work reliably? Existing benchmarks cover a limited range of patent tasks and simplify evidence acquisition by providing reference materials or candidate documents. In this paper, we introduce PatentAgentBench, a benchmark comprising 510 instances across eight task suites: novelty search, freedom-to-operate (FTO) search, design FTO, invalidity search, claim charting, patent drafting, office-action response, and patent translation. Curated from real patent records and professional materials, the benchmark combines task-specific metrics with expert-reviewed annotations and rubrics. Retrieval tasks such novelty search, FTO, etc., require agents to search and compare patent collections without a supplied candidate pool. The drafting and OAR tasks assess the quality of agents' work given technical disclosure and office action notice. We evaluate seven LLMs across four agent harnesses on PatentAgentBench, finding that current systems still struggle to identify novelty, assess risk, and understand practitioner intent throughout patent workflows.
en
Synthetic Child-Adult Interactions Are Not Developmentally Plausible
Synthetic data has become a common resource for training language models, and this trend has spilled over into research on developmentally plausible language modeling. However, it remains unclear whether LM-generated child-adult interactions actually resemble natural ones. We compare the synthetic TinyDialogues dataset to natural dialogue across a broad range of linguistic measures. While synthetic data matches natural data on the lexical level, it drastically underestimates the messy, fragmented nature of spoken interaction. To mitigate this, we propose instruction-tuning with data created from real dialogues, and show in a pilot study that a small model approximates the statistical tendencies of natural data after moderate parameter-efficient fine-tuning.
en
BellowsKV: One Latent for Many Layers, Expanded on Demand
Large language models (LLMs) now serve as the backbone of agents that reason over many turns, call tools, and pursue long-horizon tasks. To avoid recomputation, a model stores the attention keys and values of all past tokens in a key-value (KV) cache. Each layer maintains its own cache, so KV-cache memory can become the main bottleneck for GPU serving. Most compression methods operate within a single layer, by quantizing the cache entries, reducing the channel dimension, or evicting less important tokens. All three lose substantial accuracy at high compression ratios. Cross-layer methods share one compressed cache across adjacent layers to go beyond this limit. However, these methods also degrade sharply at high compression ratios. We find that the KV-cache representations of adjacent layers are quite different, whereas the input hidden states of adjacent layers are highly similar. We also find that, under the same memory budget, reducing the dimension of this shared latent loses far more accuracy than quantizing it. Based on these findings, we propose BellowsKV, a training-free method that maps the hidden states of n adjacent layers to one shared latent per token. At inference time, each layer reads its own keys and values from this latent through a layer-specific projection. We keep the latent at full dimension and compress it in two other ways: BellowsKV-Quant stores the latent at low precision, which is nearly lossless because the latent has no dominant outlier channels; BellowsKV-Evict makes one shared eviction decision per block, because the layers rank token importance very similarly. Across Llama-3.1-8B, Mistral-7B and Qwen2.5-7B, BellowsKV-Quant retains about 99% of full-cache accuracy at 3.6x compression and 97-98% at 13-14x, where the strongest cross-layer method retains at most 84% at 16x. Code: https://github.com/MorrisZJ/BellowsKV
en
Multiscale Policy Optimization for Long-Horizon LLM Agents
Long-horizon LLM agents interact with their environments through complete assistant turns while generating each action token by token. We observe that critic values vary primarily across turns, motivating optimization anchored at turns with finer credit allocation among tokens. We propose Multiscale Policy Optimization (MPO), combining Multiscale Advantage Estimation (MAE) and turn-level importance sampling to align credit assignment and policy updates at the turn scale. Extensive experiments across search, tool-use, and automation tasks with 4B and 30B-A3B models demonstrate consistent gains over strong RL baselines.
en
When the Persona Contradicts the Line: Diagnosing Plan-to-Synthesis Failures in Chain-of-Thought TTS
Chain-of-thought TTS systems rely on language-model planners to translate persona and context into vocal delivery, but it is unclear where this intent is lost when the requested affect conflicts with lexical sentiment. We introduce a bilingual Japanese--English persona--line conflict benchmark in which each item pairs one line with congruent and conflicting personas and pre-declares the expected acoustic direction. We probe this direction in the planner-generated delivery phrase, the planner-written emotion vector, the QwenEmotion-generated vector, and the rendered waveform. Planner outputs are usually direction-correct, while downstream agreement can be substantially lower. Bypassing QwenEmotion provides only modest gains over the full planner-controlled Base. Supervising QwenEmotion raises the direction rate to 75% in Japanese and 84% in English. A learned text-to-latent interface reaches the highest English rate of 89% and the highest Japanese arousal rate of 88%. These results show that learned repair of the plan-to-synthesis interface can recover counter-textual delivery, although its effect differs across languages and acoustic axes.
en
Why Are LLMs Confidently Wrong? Correcting Overconfident Errors via Causal Head Intervention
Modern large language models (LLMs) often exhibit overconfident errors, assigning high confidence to incorrect predictions and thereby undermining reliability in real-world use. Existing calibration methods typically rely on post-hoc adjustment or retraining, operating at the output level without addressing the underlying causes of miscalibration. We propose CoCHI, (Class-oriented Causal Head Intervention), a framework that improves calibration by identifying and intervening on internal components responsible for distinct confidence behaviours. By analysing predictions through correctness-confidence patterns, CoCHI enables targeted intervention on model internals, reducing overconfident errors while preserving correct predictions. Our approach provides a mechanistic perspective on calibration, linking reliability to identifiable structures within the model rather than treating it as an output-level issue. Experiments across multiple LLMs show consistent improvements in calibration metrics, including Expected Calibration Error, Negative Log-Likelihood, and Brier Score, while maintaining accuracy on five multiple-choice question answering benchmarks. These results suggest that effective calibration requires not only adjusting outputs, but also addressing the underlying causal mechanisms within the model.
en