Skip to content
AI NEWS JOURNAL
AI, IN PERSPECTIVE

Search AI news

Search headlines, summaries and articles to connect the dots.

Newest first

An Operationalization of Opaque Serial Depth

arXiv:2611.17495v1 Announce Type: new Abstract: Chain-of-thought (CoT) is currently a valuable tool for overseeing AI models, but some architectural shifts could significantly reduce CoT monitorability. We operationalize a measure that serves as a proxy for the amount of unverbalized serial cognition a model can perform, as a specific instantiation of the notion of opaque serial depth: roughly, the length of the longest path through a model's computation that does not pass through an interpretable bottleneck. Our main contribution is a particular standard for what counts as such a bottleneck. We define natural-language-rooted (NL-rooted) nodes as nodes that were initialized from a pre-training prior to produce human-understandable outputs, whose output space has not been expanded beyond that prior, and whose outputs the model has not been modified to interpret as a different type of data. By construction, the resulting measure, NLS depth, can be adjudicated before training from an architecture, training recipe, and deployment configuration alone, and is insensitive to ordinary post-training such as prompting, distillation, and RL. We find that, across open-weight models, NLS depth correlates strongly with layer count, with some exceptions, and has grown slowly over recent years. It would likely grow slowly under continued scaling of standard transformers, while latent-reasoning architectures such as full-bandwidth transformers could increase it by a large amount. We work through how NLS depth applies to a range of existing and proposed architectures, discuss alternative operationalizations, and give a worked example bounding the depth of Gemma3-1B. We argue that AI developers should track and report the NLS depth of their models as a complement to direct measurements of CoT monitorability.

Does the Ranking Hold? A Practitioner-Facing Audit of Hallucination Detectors on Small, Quantized, On-Device LLMs

arXiv:2611.17339v1 Announce Type: new Abstract: Hallucination detection for large language models is dominated by methods evaluated on 7B-70B research models with abundant sampling budgets and exact-match or ROUGE correctness labels. The small models practitioners actually ship on-device, Phi-3-mini, Llama-3.2-3B-Instruct, Qwen2.5-3B-Instruct, and Gemma-2-2B-Instruct, typically at 4-bit precision, sit outside this evidence base. We present a reproducible one-GPU audit of four detectors spanning three paradigms: length-normalized token-margin calibration, a lexical n-gram baseline, and two sampling-based consistency methods (semantic entropy and SelfCheckGPT-NLI) on two small open instruction-tuned models (Llama-3.2-3B-Instruct and Qwen2.5-3B-Instruct) loaded at 4-bit NF4 quantization on a single Colab T4 GPU, evaluated on closed-book TriviaQA. Cross-model ranking transfer is poor: Spearman rho is -0.80 and Kendall tau-b is -0.667, meaning the detector ranking actively inverts between the two models, and the cheap lexical baseline is statistically indistinguishable from the sampling-based consistency detectors on both models under bootstrap 95 percent confidence intervals. As a second, methodological contribution, we show that the standard token-overlap-to-multiple-choice correctness labeler for TruthfulQA-Generation produces near-chance AUROC (0.46-0.59) for all four detectors on both models, a labeler failure, not a detector failure, since the same detectors reach 0.69-0.89 AUROC on TriviaQA under the identical pipeline. We release the harness, prompts, and per-item scores and frame the detector choice explicitly as a fairness-of-error-rates allocation problem for practitioners deploying on resource-constrained, on-device hardware.

HarmonyText: Harmonious Scene Text Editing via Reverse-to-Refined Pair Synthesis and Region-Decoupled Frequency Learning

arXiv:2611.17733v1 Announce Type: new Abstract: Natural language-guided, mask-free scene text editing aims to modify text in real-world images using only textual instructions, yet remains challenged by unreliable pseudo-GT supervision, over-smoothed text generation, and the absence of GT-free evaluation. Existing methods often rely on imperfect pseudo-targets with background discrepancies, while standard MSE-based velocity regression tends to suppress high-frequency details, leading to blurred text boundaries and weakened local textures. To address these issues, we propose HarmonyText, a mask-free scene text editing framework with region- and frequency-aware supervision. HarmonyText introduces Region-Decoupling Optimization (RDO), which aligns edited text regions with pseudo-targets while anchoring preserved regions to the source image, and Cascaded High-Frequency Texture Guidance (CHFTG), which applies cascaded wavelet-domain constraints to preserve sharp boundaries and fine-grained textures of textual content. We further propose Reverse-to-Refined Pair Synthesis (RePair) to construct STE25K, a high-quality 25K paired dataset, and introduce a GT-free evaluation protocol with Color Distribution Divergence and a Harmony Assessment Model. Extensive experiments show that HarmonyText outperforms existing methods in text editing accuracy, background fidelity, color consistency, and perceptual harmony.

QCAP-VQA: Question-Conditioned Capped Prior-Mass Supervision for Medical Visual Question Answering

arXiv:2611.17708v1 Announce Type: new Abstract: Medical visual question answering models can answer correctly without attending to the anatomy that the question names. Automatic region priors allow spatial supervision without expert masks, but they are noisy, and it is unclear how strongly they should shape attention and how their effect should be verified. We present QCAP-VQA, which supervises a subset of cross-attention heads with a question-conditioned prior built from ranked segmentation proposals. Its capped prior-mass objective penalizes insufficient attention inside the prior only below a threshold. We evaluate attention against SLAKE structure boxes that play no role in training, separating attention on the named anatomy from how attention shifts between questions about the same image. QCAP-VQA matches the accuracy of a baseline without spatial supervision (84.26% vs. 84.23%), and every evaluated prior-supervised model places more attention on the annotated structures. Over a 64-fold range of regularization strength, the capped objective keeps prior mass between 0.67 and 0.77, whereas a pooled region-mass objective drives it to 0.97 and loses accuracy. The cost lies in question conditioning: as the constraint widens or strengthens, attention shifts agree less with the expert-defined contrast. Across 13 configurations, prior-referenced and expert-referenced evaluation rank the models in largely opposite order (Spearman correlation of -0.87), and the capped models it favors track the prior's own errors. Because clinical use of medical visual question answering requires answers grounded in the queried anatomy, supervision from automatic priors should be judged against independent anatomical references.

Autonomous Mobile Robots on a Computing Field

arXiv:2611.16652v1 Announce Type: new Abstract: We introduce CABOT, a computational model for autonomous mobile robots in which the discrete plane the robots operate on is not a passive stage but an active computing medium. In the classical Look-Compute-Move (LCM) model every increase in computational capability is obtained by strengthening the robots themselves, with lights, memory or wider visibility, at a cost paid once per robot. CABOT strengthens the environment instead: each cell of the grid holds a state from a finite set Q, the states evolve synchronously under a uniform local rule that also senses whether a robot stands on the cell, that is, the floor is a cellular automaton, and the Look operation returns the states of the visible cells in addition to the positions of the visible robots. The robots are as weak as we can make them: oblivious, silent, lightless, myopic with visibility range 1, without multiplicity detection and unable to write to the floor. The number of field states |Q|, rather than any robot capability, becomes the measure of the resources required. On an inert floor (|Q| = 1) such robots can neither form a line, nor gather from scattered positions, nor explore a grid. We show that a handful of states changes this. Line Formation from connected configurations is solvable with four states, and field-silently, so that the floor also reaches a fixed point, with five. Gathering from arbitrary, possibly disconnected, configurations is solvable with three states and Line Formation from arbitrary configurations with six; we prove that neither can be solved field-silently with any number of states. A single robot explores every finite grid forever with four states and every torus with two, which is optimal. Further rule sets, validated by simulation only, equalize the spacing of robots on a row, fill a pattern painted on the floor and guide robots through a maze, and indicate the range of the model.

Emergent Vision-Shaped Thinking in Text-Only Reasoning through Multimodal Training

arXiv:2611.14942v1 Announce Type: new Abstract: Does multimodal training shape how a language model thinks, even when the input is text only? We test this with targeted causal ablations across 10 multimodal configurations from 6 model families, with matched text-only backbones and cross-family transfer controls. We contrast spatial-imagery tasks that invite mental scene construction with symbol-manipulation tasks solvable by symbolic rules alone. We identify a vision-language axis in hidden-state space and ablate its projection at inference. In Chameleon-7B, this ablation drops accuracy on BIG-Bench Hard (BBH) navigate by 27.6 pp while leaving boolean expressions intact (+5.6 pp), a 33.2 pp double dissociation. The effect is direction-specific and layer-localized, extends to mental-simulation tasks such as mental rotation, paper folding, and perspective taking, and is not explained by task difficulty, reasoning depth, or sequential update load. Together, these results reveal emergent vision-shaped thinking (EVT): text-only reasoning that causally recruits representations introduced or strengthened by vision-language training, without assuming an image-like internal format. The measured effect is spatially concentrated at 7-8B and broadens to non-spatial visual attribute binding in the tested 32B configuration. In dual-path unified models (BAGEL, Janus-Pro), the causal handle localizes to the generation pathway as a compact one-dimensional carrier. The understanding pathway also encodes spatial information in a higher-rank, probe-decodable form, but is not the causal handle. To separate decodability from causal use, we analyze concentration, alignment, and semantic location. These diagnostics explain why behavior, probing, and intervention can diverge. Multimodal training thus reshapes the geometry of text-only reasoning, not just how visual input is processed.

ForeSpark: Joint Draft and Verification Budget Scheduling for Parallel Speculative Decoding

arXiv:2611.13442v1 Announce Type: new Abstract: 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).

SELD: Detecting Silent Evidence Loss in Context Compression beyond the Answer

arXiv:2611.12903v1 Announce Type: new Abstract: Context compression is now routine for language models with long inputs, but a compressed context is useful only if it keeps the evidence for the answer. Existing checks compare the reader model's answer or its most likely words, and thus miss a frequent failure, silent evidence loss: key evidence is removed while the answer stays the same, leaving the answer unsupported and risky for other models that reuse the context. To address this challenge, we propose SELD (Silent Evidence Loss Detector), which needs no training and no labels. Its principle is that when evidence is lost, a reader model's preferences among the answers it leaves unsaid shift even if its answer does not, so comparing these preferences before and after compression reveals losses that answer checks cannot see. We also build SEL-Pairs, a benchmark of length-matched evidence and control deletions from three question-answering datasets. Quantitative and qualitative experiments with five reader models on it show that SELD consistently outperforms strong and recent baselines across compressors, compression rates and models, identifying the evidence-removing edit in 83% of matched pairs with unchanged answers, against 70% for the strongest baseline. These results show that evidence loss in compressed contexts can be detected even when the answer is unchanged.

Tracing Psychometric Inference in Large Language Models

arXiv:2611.12937v1 Announce Type: new Abstract: Do large language models represent psychological constructs internally, or do they only generate outputs that resemble psychometric structure? We address this question using a cross-persona paradigm across 14 model variants from two families (Llama 3 and Qwen 2.5, 0.5B-14B, base and instruct). Given responses on one psychological scale from real individuals (N = 272 Taiwanese university students), models predict responses on six other scales. At the behavioral level, larger variants reproduce and amplify human cross-scale correlations; amplification persists after correcting the seven-scale human correlations for measurement attenuation. We then examine whether this structure is reflected internally. Contrastive direction analysis finds geometry aligned with the human matrix in larger instruct variants, but little alignment in base variants under this assay. Across models, geometry strength is associated with behavioral amplification (r = 0.68, p = 0.008; partial r = 0.69 controlling for log size). To relate internal structure to output behavior, a matched activation-probing paradigm shows that probe slopes vary less than behavioral slopes (1.38-1.77 vs 0.52-1.53). A synthetic control shows that ridge probing can inflate these slopes; after subtracting a baseline based on a reliability proxy from another analysis, adjusted slopes remain associated with behavior (r = 0.88). We term the measured probe-behavior gap readout attenuation. These results establish decodability and geometric alignment under the tested assays, without showing that the measured directions mediate the cross-persona answers.

CulturalCrosswords: A Multilingual Benchmark comprised of Answer Constraints and Grid Completion

arXiv:2611.10044v1 Announce Type: new Abstract: When several answers fit a riddle, exact-match scoring can penalise a valid response. CulturalCrosswords tests how answer constraints affect large language models (LLMs) solving cultural riddles, whose solutions draw on knowledge specific to a community. We constructed 53 crosswords from crowdsourced riddles in 42 languages and evaluated 11 LLM configurations on five versions of each clue set: plain batches, batches with lengths, batches with lengths and partial-answer hints, and crosswords requesting an answer list or a completed grid. On 962 clues with alternative answers, lengths improve agreement with the reference by 22.8 percentage points, compared with 8.4 points when accepted alternatives also receive credit. Part of the gain therefore appears to come from selecting a required answer form. Crossword input increases the share of all crossings with compatible answers by 9.7 points; requiring a completed grid reveals large differences between models. The benchmark distinguishes agreement with a reference from the ability to produce answers that fit together and occupy the correct cells.

OneReel: Efficient Multi-Scene Video Generation with Scene-Structured Attention

arXiv:2611.10594v1 Announce Type: new Abstract: Multi-scene video generation must follow scene-specific prompts while preserving consistency across scenes. Dense joint generation enables cross-scene information exchange but incurs quadratic attention cost, whereas separate generation avoids this cost at the expense of cross-scene consistency. We present OneReel, a training-free system that exploits scene structure to enable efficient multi-scene generation. Our key insight is that multi-scene generation can preserve complete intra-scene attention while restricting cross-scene information exchange to a bounded shared memory. OneReel uses scene prompts and durations to establish scene-specific attention scopes and dynamically selects a bounded set of representative latent frames for cross-scene sharing. This reduces video attention complexity from quadratic to linear in the number of scenes. To translate these savings into execution speedups, OneReel packs scene-local and shared K/V into dense variable-length attention and employs work-balanced multi-GPU partitioning with communication-aware selection between K/V gathering and Ulysses-style parallelism. Our evaluations across four video generation models show that OneReel achieves 1.7-4.9x end-to-end speedup over dense attention and 1.2-2.2x over existing sparse-attention advances, while improving text-following accuracy by 34-37 percentage points and maintaining cross-scene consistency.

Adaptive Scheduling Pipeline for Multi-Instance Asynchronous Reinforcement Learning

arXiv:2611.08864v1 Announce Type: new Abstract: We present the Adaptive Scheduling Pipeline (ASP), an offline rollout scheduler for multi-instance asynchronous reinforcement learning (RL) of large language models (LLMs). ASP jointly reduces three sources of rollout underutilization - padding bubbles from intra-minibatch length heterogeneity, skewness bubbles from long-tail outputs, and sorting bubbles from fluctuations in training-queue occupancy induced by length-sorted dispatch - through a single offline preprocessing pass. This enables efficient multi-instance rollout that overcomes the early scaling plateau of single-instance inference while smoothing the arrival of generated samples at the trainer. We integrate the scheduler into an asynchronous architecture that decouples rollout from training while keeping policy lag observable and bounded. Complete ASP improves end-to-end throughput by up to 3.22 times and reduces iteration time by up to 76.8% relative to MindSpeed-RL, while reducing iteration time by up to 62.5% relative to verl. Within ASP, scheduling reduces mean iteration time by 10.9%. ASP also scales more robustly with additional generation resources than these baselines and shows no visible convergence degradation in the measured bounded-lag regime. ASP has since been used in an internal telecom-domain RL post-training workflow, reducing the reported training period from more than three months to about two months and producing a domain-enhanced model for knowledge answering and tool calling.

AttentionSwitch: Jailbreak Defense for LLMs via Joint Learning of Sparse Attention-Head Selection and Scaling

arXiv:2611.08752v1 Announce Type: new Abstract: Large language models have achieved remarkable performance across a wide range of tasks, yet they remain susceptible to jailbreak attacks that circumvent safety mechanisms and elicit harmful or unintended outputs. Although safety can be improved through model fine-tuning, such approaches can be computationally expensive and may degrade previously acquired capabilities through catastrophic forgetting. Recent jailbreak defenses have therefore explored targeted interventions on specific model components associated with unsafe behavior. However, existing approaches typically decouple two inherently related decisions: where to intervene and how to modify the selected components. Such sequential optimization may be suboptimal because the optimal intervention location can depend on the intervention applied there, and vice versa. In this work, we introduce AttentionSwitch, the first jailbreak defense to jointly optimize intervention sites and strengths within a frozen LLM. Specifically, AttentionSwitch learns a sparse set of attention heads together with query-conditioned multiplicative scaling factors, allowing the outputs of the selected heads to be either attenuated or amplified without altering the directions of their representations. Across popular language models and jailbreak attacks, AttentionSwitch outperforms existing defenses, with particularly strong results when training data is scarce.

FishSense Lite: A Camera-Based, Single Laser Hardware Framework with Software System for In Situ Fish Length Measurement

arXiv:2611.08732v1 Announce Type: new Abstract: Fish stock assessments need fish lengths paired with species from far more places than trained surveyors can reach. Recreational divers already visit many of those places, but measuring a fish underwater has required trained observers or specialist equipment. We present FishSense Lite, which recovers fish length from a single photograph taken with a recreational dive camera, a dive laser pointer, and a 3D-printed mount. The diver aims the laser at a fish and takes a picture; afterward, the dot's position in the image gives the range to the fish, and the range turns the fish's size in the image into a length. A citizen-science program, trained by the system's builders, taught its own divers and collected lengths paired with species for 73 wild fish across seven open-water dives. On seven wild fish also measured by an underwater stereo-imaging system, the two instruments agreed without detectable bias (mean difference +0.8%). On rigid targets of known length, FishSense Lite recovered every target to within 1.3%, and 99% of 1,055 frames fell within a 15% error budget. From the error sources we measure, we derive a field protocol: calibrate against a dive slate on every dive, photograph fish within 15 degrees of perpendicular and beyond 1 m, and take at least ten frames of each.

IWIBench: Evaluating Gradient-Free Model Driven Intra-weight Injection

arXiv:2611.08647v1 Announce Type: new Abstract: Intra-weight injection lets one model change another model's behaviour by editing its weights directly, without gradients. Such edits can range from altering lexical associations to enforcing output formats or updating factual knowledge. Despite growing interest in these approaches, there has been no systematic way to evaluate how well models perform them. We introduce IWIBench, which measures a large editor model's ability to modify the weights of a smaller target model (typically Gemma 3 4B IT) to achieve one of ten classical text-editing objectives while preserving its other capabilities. Editors are limited to four tools for reading and writing weights, with bounded tensor access and a fixed budget for generating and inspecting rollouts. A validator applies each proposed patch in a clean environment and compares the edited model against the unmodified target. Evaluating frontier closed- and open-weight models reveals sharply diverging capabilities, both across models and across tasks. A small number of models, most notably Astra and GPT-6.1 Sol, perform far ahead of the rest. The recency of this ability (little of it is evident in GPT-5.6 Sol, for instance), together with qualitative analysis, suggests it has been acquired through recent training. Tasks also vary widely: structured response resisted every attempt, and answer-length control yielded only one successful solution, whereas alias acquisition and recoding proved most amenable. Success in one task family correlated poorly with success in others, indicating an early and uneven capability frontier. We discuss strategies that editors repeatedly adopted and conclude with an outlook on intra-weight injection as an emerging but already potent approach to gradient-free behaviour modification.

ReKV: Overlapping Quantized KV Transfer and Suffix Recomputation

arXiv:2611.08770v1 Announce Type: new Abstract: Prefill--decode disaggregation moves a prompt's key--value (KV) cache between workers before the first token can be generated. Reducing this handoff to either cache compression or recomputation misses a useful choice: which layers to transmit, at what precision, and which layers to execute again on the decode worker. ReKV transfers a quantized prefix KV cache and one BF16 boundary state, then reconstructs a contiguous suffix while the prefix is in flight. In a model-shaped synthetic benchmark on cross-node H100s, the Q8 configuration meeting separately evaluated quality criteria reduces 8K handoff latency by $2.36\times$, $2.84\times$, and $3.65\times$ over BF16 transfer for Qwen2.5-7B, Qwen2.5-3B, and Qwen3-8B. It is also faster than the best measured transfer-only or recompute-only action at those points. A separate real-weight BF16 recomputation experiment reduces prefill-to-token-return time by $1.16$--$1.63\times$. The preferred action changes with prompt length and link bandwidth.

Structuring Survey Histories for LLM Digital Twins: A Controlled Study of Background, Decision Procedure, and Evaluation Profiles

arXiv:2611.08730v1 Announce Type: new Abstract: LLM-based digital twins aim to predict how individuals would respond in new settings from representations of their prior responses. Existing evidence that compression of survey histories need not reduce predictive accuracy suggests that representation structure, rather than information volume alone, may be an important design choice. We study BDE, a fixed representation that organizes each respondent's survey history into Background, Decision procedure, and Evaluation. We evaluate it on Twin-2K-500 (Toubia et al., 2025) by predicting each respondent's held-out answers on 17 tasks. In a paired evaluation with 100 respondents, a fully inferred BDE profile improves overall accuracy over the raw survey transcript by 1.27 percentage points (pp) on gpt-5.4-nano (p < 0.01) and 1.19 pp on gpt-5.4-mini (p < 0.001). Three analyses then examine the schema, extraction depth, and component composition. Against a length-matched free-form profile written from the same transcript by the same extractor, BDE improves pricing accuracy on both simulators (2.15 pp on nano, p < 0.01; 1.38 pp on mini, p < 0.05), while its overall advantage is significant only on nano (1.05 pp, p < 0.05). Allowing deeper inference during extraction yields no statistically significant gain on either simulator. By contrast, adding Evaluation to a Background--Decision procedure profile improves overall accuracy by 1.44 pp on nano (p < 0.001) and 1.22 pp on mini (p < 0.001). These results indicate that the advantage over raw transcripts is not explained by compression or length alone, and attribute it primarily to profile composition rather than deeper extraction.

Finite-Time Convergence and Regret Guarantees for Neural Reinforcement Learning in POMDPs

arXiv:2611.07069v1 Announce Type: new Abstract: Reinforcement learning (RL) in Partially Observable Markov Decision Processes (POMDPs) is computationally challenging due to the curse of long-range dependence on history. Fortunately, recent studies have shown that a range of POMDPs that satisfy the filter stability assumption can be well approximated by a superstate Markov Decision Process (MDP) based on a truncated history in a finite-memory window, with a value gap that decays with the window length. However, existing finite-time convergence analyses on such superstate MDP surrogates are either limited to over-simplified linear function approximation or rely on an i.i.d. sampling oracle that is violated by POMDP trajectories, both of which severely limit the applicability of these theoretical results in practice. This motivates us to bridge the gap by investigating RL in POMDPs with nonlinear neural value function approximation. Our main contributions are threefold: i) We first establish a new finite-state MDP point-wise approximation error bound, which lays the foundation for our RL algorithm design and analysis in POMDPs; ii) By applying Temporal-Difference (TD) learning to POMDP-generated superstate transitions and targeting the Bellman fixed point of the superstate MDP, we prove a finite-time critic error bound with nonlinear neural value function approximation; and iii) By combining the neural critic function approximation error bound with the policy mirror descent framework, we obtain a finite-time sublinear regret against the optimal POMDP value. To the best of our knowledge, such a finite-time sublinear regret based on neural function approximation is the first in the literature for finite-memory POMDPs. Our numerical experiments also verify our theoretical results.

LowRAM: Low-Rank Keys and Amortised Values for Composable KV Cache Compression at Test-Time

arXiv:2611.07025v1 Announce Type: new Abstract: Autoregressive decoding relies on the KV cache, whose memory footprint grows linearly with context length and caps batch size and throughput. Existing post-training methods exploit redundancy across tokens, layers, and feature dimensions, or lower numerical precision. We identify a complementary axis: within each attention head, values are largely predictable from their corresponding keys. We exploit this with value cache amortisation, which fits a lightweight per-head predictor at test time and stores residuals only for the tokens it reconstructs worst, giving $4\times$ near-lossless value compression. The predictor can be initialised by meta-learning for substantially reduced prefill overhead, or fitted from scratch in settings that allow a one-off cost, such as multi-query inference. We pair value amortisation with cross-layer low-rank key compression to produce LowRAM, an asymmetric compression pipeline that stays close to lossless across three model families. By reconstructing only a selected subset of KV pairs at each decoding step, LowRAM can keep decoding overhead minimal and achieve up to $5\times$ higher peak decode throughput at $4\times$ compression. LowRAM further composes with quantisation and eviction, reaching $26.9\times$ compression at $94.5\%$ of baseline performance on long contexts. These results show that treating keys and values as structurally distinct enables efficient, aggressive, and composable KV cache compression.

Trust Responsive Runtime Authorization for Agentic AI with Empirical Characterization of Policy Behavior and Parameter Sensitivity

arXiv:2611.06996v1 Announce Type: new Abstract: The policy that converts runtime estimates of an autonomous agent's condition into operational authority is rarely specified or measured. We empirically characterize the Trust Responsive Authorization Policy (TRAP), which maps an Authorization Readiness Score (ARS), a weighted combination of runtime trust, operational stability, and contextual risk, onto discrete authorization levels through an asymmetric rule: escalation requires consecutive observations above a band edge plus a margin, while a sharp drop in readiness restricts authority at once. On one matched set of 100 Monte Carlo traces, we compare TRAP with static, trust-only, risk-only, and symmetric policies. Under abrupt degradation, TRAP restricts at onset and re-escalates sooner after recovery than the symmetric policy, with the highest and, descriptively, least variable stability among the adaptive policies, although more of its post-recovery escalations occur below the selected trust threshold. Paired bootstrap intervals exclude zero for the remaining differences. In one-factor sweeps, recovery time rises with the escalation margin and run length, shifts with the score weights, and is unchanged by the restriction threshold. In an exploratory post-freeze experiment, when the same degradation is spread over 5-20 intervals, TRAP restricts only on the 4 traces that retain a single-interval drop above that threshold; averaged over all 100 traces, it holds authority above what readiness supports for 15.9-18.9 of the 20 degradation intervals, against 0.8 under abrupt degradation and about 10 for the symmetric policy. Its restriction advantage therefore depends on degradation being abrupt. All reported values are traceable to the publicly available, versioned research release.