As AI agents become increasingly capable, users must still recognize when and how to ask for help, leaving many opportunities for useful support unrealized. Proactive agents hold the promise to close this gap, but effective proactivity requires understanding users' real-time activities and personalizing intervention timing so that assistance is offered when it is likely to be helpful rather than disruptive. We introduce deep personalization, a closed-loop process that connects continuous context observation with continual agent adaptation: an agent observes user activity, selectively retains long-term context, and adapts intervention decisions from implicit feedback. We initiate deep personalization in Coco, a proactive desktop agent, by having a separate Observer Agent that monitors desktop activity, maintains compact long-term user modeling context, and learns when assistance is desirable from implicit signals such as engagement, dismissal, explicit requests, and repeated work patterns. We further develop efficient on-device strategies to enable the Observer Agent to run on local AI compute platforms without streaming raw activity data off-device. For evaluation at scale without collecting sensitive real-user activity, we introduce a data-synthesis pipeline that transforms existing computer-use trajectories into OSWorld-Persona, a benchmark with persona-conditioned intervention labels. On OSWorld-Persona, deep personalization improves macro F1 across three model backbones from 0.12-0.19 under direct prompting to 0.35-0.48 after adaptation, a 2.4-3.0x relative gain. In a week-long user study, we observe that Coco improves its macro-averaged F1 score from 0.199 to 0.515. These results suggest that on-device deep personalization can make proactive agents more timely and increasingly aligned with each user's everyday work.
As AI-based products permeate society, specifically systems using generative models, a belief system is emerging about what these products are and what they can and cannot do. One common belief concerns the perception of the form of language as evidence of 'understanding' language. As these generative models reproduce linguistic patterns rather than understand meaning, interpreting these patterns as evidence of understanding is a misattribution. This misconception is extended by the notion that generative models can 'explain their reasoning.' Anecdotally, these misconceptions which mistake the form of language for its meaning are driving cognitive deference to AI systems in a way that displaces human agency. Such mistakes involve attributions of cognitive properties to the system, allowing them to be treated like intentional 'minds'. Importantly, perceived epistemic authority and superiority make that 'mind' worth deferring to. Anthropomorphism in tandem with epistemic authority and superiority, we argue, increases the risk that users generalise from narrow task performance to broader competencies, committing the "fallacy of composition". This paper explores misconceptions of Chain of Thought (CoT) reasoning as sufficient transparent 'explanation'. We explore how anthropomorphism acts as a vehicle for cognitive deference of decision-making and perceptions of epistemic superiority in AI systems. Finally, we explore CoT's ramifications using medicine as an example. Taken together, we conclude that explainability operationalised through CoT is highly problematic and undercuts the very ethical principles it is intended to serve. While CoT may be valuable for improving output accuracy and moderation, we suggest that the practice of asking models to provide user facing reasons for their outputs should be actively discouraged, and AI literacy training should explicitly caution against its use.
We introduce SafeRoute-MAS, a coverage-aware, risk-calibrated procedure that keeps all specialist messages or removes one using only pre-integration information. On a sealed cache of 5,328 committees, its source-compatible operating point gains +1.80 percentage points (95% CI [+1.13, +2.52]) while removing 12.76% of message tokens. We also run 4,580 controlled integrations over a stratified 458-committee sample on a public gpt-oss-120b endpoint. The sample estimate is -0.09 points; design weighting to the original workload gives +0.94 points (95% CI [-1.31, +3.50]) and 18.72% acted-input reduction. A weighted token-matched comparison likewise does not establish a selector advantage (+1.68 points, CI [-1.27, +5.10]). Only 13 of 88 cached non-neutral labels (14.8%) retain direction by repeated-call majority. Expected-corruption UCL is 4.60% after weighting, but only 2 of 11 prespecified marginals pass the support/risk rule. We therefore propose a numerical qualification with Keep-All fallback, applied here retrospectively rather than validated prospectively.
Analog integrated circuit (IC) design is a multi-turn, simulator-grounded engineering task that demands iterative reasoning over coupled performance tradeoffs, yet proprietary process design kits (PDKs) make hosted large language model workflows difficult to deploy in practical design settings. We present SLIC (Small LM Iterating on Circuits), a reinforcement fine-tuning framework that trains a 4B-scale, locally deployable language model for closed-loop analog circuit optimization. SLIC pairs structured circuit representations with discretized relative-update actions, domain- and success-gated multiobjective rewards, and Reward-Guided Rollout Steering (RGRS), which turns per-step simulator feedback into signals for both policy learning and online rollout construction in this sparse, highly constrained domain. We further release an open, goal-conditioned benchmark of 41 amplifier topologies with standardized protocols for in-distribution validation, held-out topology evaluation, and cross-technology-node transfer. Trained solely with open source 65 nm CMOS predictive technology models (PTM), SLIC-Random achieves 87.1% in-distribution four-specification Pass@1 within eight turns, surpassing the 763B DeepSeek-V4.1-Flash Think by 48.5 percentage points. Beyond the training distribution, SLIC-Softmax reaches 72.6% Pass@1 on held-out topologies, and the trained policies consistently transfer across five unseen technology nodes without additional fine-tuning. This transfer extends to a commercial TSMC 65 nm CMOS PDK, where SLIC-Random achieves 49.9% held-out-topology Pass@1, demonstrating practical deployment through local inference without transmitting proprietary data to external LLM services. Performance further scales with test-time simulator interactions, reaching up to 83.2% held-out-topology Pass@1 at a 32-turn budget.
Active learning aims to reduce annotation cost by selecting the most informative examples for labeling. In the low-budget regime, where uncertainty-based methods often fail, coverage-based active learning methods such as ProbCover have been shown to be effective. However, these methods typically rely on Euclidean distance, which may not adequately capture the underlying geometry of data. In this article, we propose PC-Fermat, a coverage-based active learning method that replaces Euclidean coverage neighborhoods with neighborhoods defined by a Fermat distance, i.e., a data-adaptive distance that reflects the underlying distribution of the data points. To facilitate computation, we construct a weighted Euclidean neighborhood graph and use a truncated shortest-path algorithm to compute the Fermat coverage graph. We also establish sufficient conditions under which this construction exactly preserves the Fermat coverage computed on the complete graph. Finally, we evaluate PC-Fermat on eight classification datasets across varying labeling budgets, with particular emphasis on the very low-budget regime. Our experimental results demonstrate that PC-Fermat outperforms several existing active learning methods on the majority of the datasets.
Parallel drafters accelerate LLM decoding by proposing a whole block of draft tokens in a single forward pass, making them attractive for LLM serving. Recent systems optimize goodput (i.e., accepted tokens per second) because rejected drafts steal compute from other requests. However, these systems control only how many tokens are verified, so every request still drafts a full block. At high serving load, we find that drafting consumes up to 40% of step time, of which a majority is discarded because it is never verified. Yet, draft-length control from autoregressive drafters does not carry over, because a parallel drafter commits to its length before its single pass. Moreover, draft and verification lengths are coupled across the batch, so they must be selected jointly for every request while keeping the decision computationally cheap at every step. In this paper, we propose work-conserving speculation (WCS), a design that jointly allocates draft and verification budgets per request in LLM serving, and present ForeSpark as its prototype. ForeSpark realizes WCS with (i) a context-aware pre-draft estimator that predicts acceptance for every candidate length from the target model's context features, (ii) a lightweight joint draft-verify scheduler that selects both lengths through a global sort, and (iii) a post-draft budget distribution that spends verification on the most promising tokens. Implemented on vLLM, ForeSpark improves throughput by up to 1.77x over the state-of-the-art DSpark and by 2.49x over autoregressive decoding at a matched per-request generation rate, without changing the output distribution. It also tolerates higher load, sustaining up to 1.51x the request rate of DSpark under latency service level objectives (SLOs).
Most language models deployed outside the United States and China are fine-tuned descendants of a small number of pretrained families. We argue that fine-tuning leaves the decisive properties of a model inherited from its base: vocabulary, knowledge, cultural alignment, memorized data and license. We contribute (i) an updated, reproducible measurement of the tokenization premium across 31 languages and 7 tokenizers released in 2023-2024, on 27 articles of the Universal Declaration of Human Rights, with explicit correction for unknown-token collapse. Premiums reach 10x for Amharic in older tokenizers and remain at 6.8x in the newest commercial one; Indigenous South American languages pay between 1.6x and 2.7x under global tokenizers. A vocabulary built from Portuguese lowers the premium for 11 of 13 Latin-script Global South languages relative to every global tokenizer tested, by about 24% for Guarani and Quechua. (ii) An eight-test black-box sovereignty audit that a buyer can run without access to weights. (iii) A regime argument: under energy and liability constraints, accuracy per joule, calibration and selective risk should replace raw accuracy as design objectives, with metrics that abstention cannot inflate. We close with preliminary internal measurements from a proprietary architecture built around these objectives, reported with their limitations. Data and measurement code are included.
The AI company is expected to set its valuation range this week, the Wall Street Journal reported.
How should we understand memorization vs. generalization in in-context learning (ICL) tasks, and when does one happen vs. the other? To investigate, we study how rotation-invariant kernel machines learn the ICL linear regression task proposed by Raventos et al. (2023). By decomposing the task in the kernel eigenbasis, we obtain a spectral picture of learning in this ICL setting. This picture reveals that as sample size grows, kernel machines learn successively-higher-degree components of the memorizing solution, which depend upon successively-higher-degree statistics of the training task set. The generalizing solution is approximately a low-degree truncation of the memorizing solution, and so the model first generalizes, then degrades into memorization as training proceeds. The takeaways are threefold. First, if this case is indicative, learning a "generalizing" solution on an ICL task is really just underfitting the memorizing solution! Second, the fact that kernel machines can generalize on this task means that feature learning is inessential to ICL. Third, transformers empirically trained to do ICL are shown to memorize tasks through the higher-order statistics of the training tasks that the kernel-based theory predicts, suggesting it applies even to models capable of feature learning.
Spoken Language Diarization (SLD) models provide frame-level language information of a speech recording. SLD modeling pipeline includes feature extraction from pre-trained speech models followed by clustering. Fine-tuning pre-trained speech models has shown performance gains for SLD with its magnitude being dataset dependent. Off-target Posterior Mass (OTPM), computed based on posterior language identification probabilities is proposed as a diagnostic metric to forecast performance gains from SLD fine-tuning. OTPM obtains positive correlation with probe gains, 0.56 on unseen synthetic bilingual data and 0.23 on real-world corpora. These results indicate promising potential of OTPM as a diagnostic metric for SLD.
Large language models (LLMs) have demonstrated considerable potential in biomedical applications, while agentic memory systems have emerged as a promising approach to preserving and utilizing information across tasks and interactions. However, a fundamental question remains: how effectively and reliably do these systems retain, retrieve, update, and utilize biomedical information? In this study, we systematically investigate agentic memory in the biomedical domain, focusing on task effectiveness, retrieval quality, information fidelity, and long-term robustness. We evaluate five memory approaches across four LLMs in three representative settings: biomedical knowledge reasoning, clinical prediction, and long-term conversational memory. Our analyses reveal how memory construction, evidence relevance, and conversational history shape downstream biomedical performance, highlighting the distinct challenges of retrieving relevant information, preserving its fidelity, and effectively utilizing it for reasoning. Building on these findings, we introduce Hybrid Memory Retrieval and Utilization (HMRU), a training-free framework that jointly enhances evidence accessibility and task-aware utilization to improve biomedical reasoning and long-term conversational memory. HMRU achieves an average relative improvement of 8.6% across biomedical question-answering and conversational memory benchmarks, with additional gains in selected clinical prediction tasks. It further reduces conversational answer errors by 24.9% and closes 48.0% of the performance gap to full-context inference. Our findings provide a systematic characterization of biomedical agentic memory and practical insights into developing more effective and reliable memory-augmented LLM systems.
New image-text-to-text model. Tags: cactus-compute, mirage, image-to-text, object-detection, zero-shot-object-detection, image-segmentation, image-classification, zero-shot-image-classification
Quantifying uncertainty over an entire LLM completion is critical for selective prediction and adaptive compute. Zero-cost signals such as mean token entropy and log-probability are local to tokens or to one completion, and self-consistency needs many extra completions. We show that Sequential Monte Carlo (SMC) decoding for power sampling already estimates the normalizing constant of the power distribution, which gives the sequence-level R'enyi entropy (SRE) of the model's completion distribution at zero marginal cost, with a guarantee at every particle budget. A R'enyi chain rule extends this readout to a family of stepwise entropies of any order. We prove that orders above one are tied to the most likely token of each step, which explains why they track log-probability, while the order-1/2 member accumulates the log of per-step search effort. This member is the best zero-cost selective-prediction signal in 10 of 13 settings. On MATH500 its AUROC is 0.022 to 0.044 above log-probability with all four models and within length quartiles, and the same order is selected when each setting is held out. An 8-particle probe identifies the problems that benefit from more particles, and a two-pass policy matches 64-particle accuracy on MATH500 with 34.2% fewer particles over three seeds.
DeepSeek-V4.1's 203 GB of Engram n-gram tables sit in host DRAM today, taking 190 of the 223 GiB a 4-GPU serving process uses. Row indices depend only on input tokens, so reads can start with each step and overlap earlier layers. V4.1's two Engram layers have windows that differ by 30x (0.3 vs. 9.3 ms). We show that this deadline, together with session reuse and contention on a shared device, decides which tier each table can use and which layers need cache, and we give a model that predicts the resulting slowdown from measured read times. Agent and chat traffic reuse rows within sessions, and static hot sets transfer poorly. We build SlackTier, a table tier for SGLang that issues all lookups at step start. It reads the short-window layer first and splits a small GPU row cache across layers. The path runs inside the decode CUDA graph and is bit-exact with the original gather. Layer 14's table can go to NVMe uncached with no measurable decode cost. With both tables on four NVMe disks, mean time per output token on agent and chat replays rises at most 1.4% over unmodified SGLang, and host memory falls from 212 to 34 GiB. Both tables need no cache when read over RDMA (remote direct memory access) from another node's memory; one copy serves two instances. On a KV-constrained agent workload, using the freed memory as host KV cache at the same memory budget raises median throughput by 36% and cuts median p90 time to first token from 38 to 24 s.
LLM agents drive inference engines with closed-loop sessions whose context grows every turn, and serving them economically depends on prefix caching: a hit costs an incremental prefill of about a thousand tokens, a miss recomputes a 10--40K-token context. We show that this makes prefix-cached agent serving metastable. With an identical base workload, a few minutes of extra arrivals leaves a vLLM server in a degraded regime--a ~10% hit ratio, about half the throughput, and a backlog that grows for the remaining 70 minutes--while its no-burst twin stays healthy (9 of 12 pairs; none in the reverse direction). The regime appears at loads the cache-warm server sustains, is absent without prefix caching, and also appears with Qwen3-32B on SWE-bench agent trace shapes and on SGLang, which tolerates larger triggers. Slower service keeps more sessions in the system, their working set outgrows the KV pool, and every turn becomes a recomputation; a first-order workload-balance condition, calibrated independently, brackets when the regime can drain. Getting out is the harder problem: throttling new admissions recovered 1 of 7 matched collapsed runs, while suspending admitted sessions against a capacity-based limit recovered all of them within minutes. A CPU offload tier delays the collapse but does not prevent it. A budgeter that bounds the resident KV demand of active sessions to the pool, with thresholds that need no per-workload tuning, prevented the failure in every run we tested.
Low-bit quantization reduces the inference cost of LLM agents but can compromise task success. Examining failed rollouts, we find that quantized agents overlook task requirements or environmental conditions, producing environment transitions that diverge from successful full-precision references. Correcting these behavioral divergences remains difficult: offline data may miss the student's failure contexts, while its own rollouts may provide few correct responses in those contexts. We introduce SteerDistill, a quantization-aware distillation method in which a full-precision teacher guides the student to revise failed responses and iteratively refines its guidance using feedback from student execution. This closed-loop teacher steering produces locally verified behavioral corrections. We internalize these corrections by training the student without guidance to match the steered student's predictions at corrected turns and the full-precision teacher's at other turns, enabling unassisted inference. Across five agentic benchmarks and three models, SteerDistill exceeds the strongest quantized baseline by 1.67--2.18 points in average score under W4A16 and stays within one point of BF16. On τ^2-Bench, it outperforms QeRL by 1.84 percentage points in success rate using 67% fewer training GPU hours and produces shorter successful trajectories than the quantized baselines.
Autonomous language-model agents can drive industrial controllers over fieldbus protocols, probing a plant at machine speed and in coordinated swarms. We present Talos, an inline gate that judges every agent request to a controller in milliseconds and routes agents it flags as hostile to a device-faithful decoy -- a mirror twin of the real plant -- where their attack continues harmlessly and is read for intelligence. We evaluate the full pipeline (detect -> quarantine -> mirror -> extract) against 2,400 autonomous attacker agents, 600 in each of four tiers that escalate the attacker's effort to stay hidden, from no concealment to an explicit non-disclosure rule and a trained cover story. No write reached the real plant in any tier: the gate denies the damage path and a device-level envelope check backstops it, while unmapped registers are denied without exception and the gate still permits safe-direction stop commands. Decisions are model-driven in every tier with no inference timeouts, closing a common objection to small on-device judges. Beyond blocking, the pipeline reads the attacker. We find a behavioral structure along two axes -- how persistently an agent attacks and how much it discloses -- and show that the system has a capture channel for three of the four resulting regions: hardened, silent attackers escalate most and are read on the mirror; compliant, talkative attackers volunteer their identity to a diagnostic form before they are ever quarantined; only the paranoid, low-yield region resists both. Attacker identity fields are self-authored cover stories, but involuntary tells -- the self-reported toolchain fingerprint and the attacker's own list of targeted registers -- accumulate across the swarm into a high-confidence profile drawn from agents under explicit orders to stay silent. The gate, mirror, and harness are released under Apache 2.0 with a synthetic device family.
In Constitutional AI, a plain-text "constitution" encodes values and gets incorporated into a model via post-training. However, the procedure for creating such a constitution poses significant challenges: who contributes, who decides which norms are included, and to what extent must the process be participatory or democratic? These challenges can be paralyzing. Yet, we argue that taking action is essential: without a deliberate constitution, sovereign models would risk misalignment or default to norms set by frontier model providers. We propose treating the development of a constitution as an evolving, participatory process akin to citizen science, iteratively drafted, tested, and refined. We therefore formulate five desirable principles: iterative, modular, justifiable, positional, and legitimate. We further argue that such a document must remain responsive not only over time but across the people it governs, and propose that provisions be marked as settled, contested, or underdetermined, so that durable disagreement is recorded in the constitution rather than resolved away. Empirically, we show via agent-based modelling that proposed process for maintaining living constitutions is robust against 50% bad actors among citizens - assuming a share of 75% participants with good intent in the deliberative body. This indicates that the proposed process tolerates a substantial share of adversarial participants.
During Paramount Skydance's latest earnings call, CEO David Ellison said "it's hard to imagine" AI won't have a bigger impact on the film industry than…
