Skip to content
AI NEWS JOURNAL
AI, IN PERSPECTIVE

Search AI news

Search headlines, summaries and articles to connect the dots.

Newest first

Smart Model Routing for Order-of-Magnitude LLM Cost Reduction

arXiv:2611.15911v1 Announce Type: new Abstract: Frontier large language models (LLMs) deliver the strongest response quality but carry per-token prices often more than an order of magnitude above those of smaller, task-adequate models. Smart model routing addresses this mismatch by assigning each query to the cheapest model expected to answer it adequately, rather than sending every request to the most capable, most expensive endpoint. This paper develops a formal cost-quality optimization framework for routing -- including a derived optimal threshold-policy structure, a calibration-versus-discrimination distinction, a cascade breakeven condition, and a sensitivity analysis over price ratio and traffic skew -- situates that framework against the principal routing architectures and the signals used to estimate query difficulty, and works through a numerical example showing how a calibrated router can reduce blended inference cost by roughly an order of magnitude relative to an always-frontier deployment under specific, stated pricing and traffic assumptions, reporting both the gross reduction from routing policy alone and the net figure with the router's own overhead priced in. We situate this figure against the range reported by published routing systems (roughly 2x to frontier-level accuracy at one-fifth the cost), and close with the practical risks -- under- and over-routing, aggregate quality constraints masking segment-level harm, calibration without discrimination, calibration drift, prompt injection into routing signals, and operational failures such as tier outages and multi-provider price variation -- that determine whether the savings materialize in production. The analysis is primarily analytical, complemented by a GSM8K evaluation of the cascade framework (Section 5), validated (Section 5.2); extending the evaluation across tasks and model families is future work.

SPIRE: Task-agnostic embodied continual learning through autonomous situation inference and causal motor-primitive reuse

arXiv:2611.14840v1 Announce Type: new Abstract: Embodied continual learning requires controllers to infer changing situations, retain prior behavior and reuse motor skills without task labels or known boundaries. The challenge is deciding when to recover earlier behavior and which skills can support new learning. We introduce Situation-guided Primitive Integration for Recurrent Embodiment (SPIRE), a recurrent-spiking system that infers latent contexts from action-outcome evidence and learns context-specific motor actors online. Mature actors become protected sources that combine with a target-context residual to drive recurrent-spiking motor populations. Across coordinate, three-joint and hidden-force control, SPIRE maintained control, autonomously inferred situations, retained prior behavior and accelerated acquisition of an unseen source composition, reducing mean interactions to criterion by approximately 40% relative to matched fresh controllers. Rewinding compatible source actors impaired early acquisition in every matched intervention, demonstrating causal reuse. Against seven recurrent continual-learning and modular-control comparators, SPIRE achieved 2.4 times the mean prior-learning advantage of the strongest eligible task-free alternative selected for each paired stream. These results demonstrate autonomous recovery and causal reuse of motor skills for new learning in continuous embodied control.

SynthCorp: A Platform for Creating Benchmarks for Data Agents

arXiv:2611.14765v1 Announce Type: new Abstract: Current benchmarks for LLM-driven data agents saturate quickly and cannot target specific capabilities. We present SynthCorp, a platform that creates customizable data estates and benchmarks by simulating how business events produce and modify corporate data over time. Each data estate combines relational databases, a document store, and a data catalog, paired with evaluation tasks and executable gold programs. First, given a scenario and target complexity dimensions, SynthCorp synthesizes reusable event types and compiles each into a program over predefined simulation primitives. Second, SynthCorp schedules event instances on a virtual calendar, interleaving routine and disruption events so that tables, documents, and catalog entries remain consistent. Third, SynthCorp executes the calendar to materialize the data estate and records every operation in a simulation log. Fourth, SynthCorp generates analytical tasks and derives verified gold programs from the log. We use SynthCorp to generate initial evaluation benchmarks comprising seven data estates (up to 1K operational tables) and 256 tasks across five domains. The data estates pass deterministic pipeline checks (such as foreign-key violations, undeclared writes, recurrence violations, and database execution failures) with a near-zero error rate, introduce cross-asset complexities, and cover all requested dimensions. Evaluating three models across four data-context settings shows that the benchmarks separate data contexts and model tiers. Auditing candidate queries against causal provenance in the simulation log verifies gold programs without manual annotation, including on multi-event composite tasks that expose reasoning failures in state-of-the-art agents. We hope the community uses SynthCorp to generate more challenging and useful data estates and benchmarks that foster the development of enterprise data agents.

Termination Lag: Coding Agents Solve Early but Stop Late

arXiv:2611.14960v1 Announce Type: new Abstract: Coding agents can solve a task well before they stop working on it. We characterize \emph{termination lag}--the number of steps between an agent's first correct patch and the end of its run--across five models, two agent frameworks, and three SWE-bench datasets by recording and testing every distinct repository patch generated during execution. Our analysis reveals that 38--82% of successful runs achieve a correct solution on their initial source-code edit. However, agents continue to expend substantial effort after correct patches, with post-solution steps accounting for a median of 35--37% for Qwen3-Coder and 54--60% for Laguna-XS.2 in successful runs on SWE-bench Verified and Multilingual. Agents spend most of these additional steps verifying their work and exploring the repository. In a small fraction of cases (1--8%), this continued activity leads agents to undo their earlier correct patch, leaving the final solution incorrect.

ToolTurns: Benchmarking LLM Agents in Multi-Turn Dialogue

arXiv:2611.14925v1 Announce Type: new Abstract: Tool-using language models are increasingly deployed in conversational settings, yet most large-scale tool-use benchmarks focus on isolated requests. We introduce ToolTurns, a benchmark of 2,620 problems spanning 1,827 tools and 11,004 API endpoints across 47 categories. Each dialogue begins with an intentionally underspecified request and continues through two context-dependent follow-ups, testing clarification, tool use, and retention of conversational context. We also introduce an open-source user simulator that reveals withheld context through dialogue and generates follow-up requests, with a checker--fixer pipeline for improving its turns. Across nine language models, the best agent resolves 69.6% of individual requests but completes only 39.8% of full dialogues. Under this evaluation protocol, every evaluated model receives lower success scores on follow-up requests than on the initial request, and dialogue-quality rubric scores and task success can rank models very differently. Paired turn-level evaluation shows that the checker--fixer reduces detected dialogue defects, and blind human evaluation favors the revised turns. ToolTurns provides a testbed for evaluating whether agents sustain task success across multi-turn dialogue with tools.

UniPipe: Unified Inference Serving and Fine-Tuning for Deadline-Constrained DNN Surrogate Workflows

arXiv:2611.14748v1 Announce Type: new Abstract: DNN surrogates are increasingly used to accelerate scientific workflows that must process high-rate sensor streams in near real time, exemplified by Ptychographic imaging at next-generation X-ray light sources. Such surrogates deliver 100-1000x higher throughput than iterative solvers, enabling faster steering of experiments. To preserve surrogate accuracy during an experiment, as the input distribution evolves, state-of-the-art deployments freeze the pretrained surrogate or run a separate CUDA context that fine-tunes concurrently with serving. However, multi-context concurrency causes significant contention on the GPU, inflating both the mean and the variance of inference latency, triggering deadline misses in workflows where sensor data has to be processed within a hard per-unit deadline. To address this challenge, we propose UniPipe, a single-context system that unifies inference serving and fine-tuning of a DNN surrogate on a single edge GPU. UniPipe is driven by three key insights: both serving and fine-tuning require a forward pass and can be fused into one; the per-sample latency of modern GPUs is non-linear in batch size; and the workload is repetitive across intervals, enabling offline scheduling. Specifically, UniPipe fuses training and inference data in a single forward pass with a selective backward pass on the training subset, and uses a polynomial-time scheduler that exploits the latency curve to compute the batch composition. We evaluate UniPipe on a real-world Ptychographic reconstruction workflow with sensor rates up to 5000 patterns/sec.; UniPipe lowers reconstruction MSE by up to 30% and inference MSE by up to 50% over a pretrained baseline, and reduces the deadline-miss rate by 13%-22% over a multi-context baseline. UniPipe's unified deployment reduces the latency spread from 106% to 57% of the mean, and increases total throughput compared to multi-context baseline.

Validation of a Facial-Video Digital Hydration Biomarker Score for Tracking Hydration State Transitions in a Real-World Fasting Model

arXiv:2611.14993v1 Announce Type: new Abstract: Hydration influences physiology and daily function, yet repeated practical measurement in everyday settings remains limited. We present initial real-world evidence that a facial-video lifestyle digital biomarker score can track the direction of hydration state transitions. Using Ramadan fasting as a model of prolonged fluid abstinence, we assessed whether the score tracks the transition between the fasting (pre-drink) state and the rehydrated (post-drink) state across three validation layers. In the primary cohort (104 participants, 208 videos), the score increased from a mean of 36.6 in the fasting (pre-drink) state to 70.6 in the rehydrated (post-drink) state (change = +33.9, 95% CI [+30.7, +37.1], p < 0.001, Cohen's d = 2.03), with 102 of 104 participants (98.1%) showing an upward shift. In a nested body-weight subgroup (20 participants), body weight increased by +0.74 kg after fluid intake (p < 0.001), confirming fluid uptake and providing supportive physical anchoring. In a nested serum osmolality subgroup (10 participants), osmolality decreased by -5.1 mOsm/kg (p < 0.001), with all post-drink values within the normal clinical range, providing exploratory biological support. An exploratory secondary comparison of water versus electrolyte rehydration showed no statistically significant difference in short-term score change (p = 0.247). This study does not establish a clinical diagnostic test for dehydration. Rather, these findings support the score as a lifestyle digital biomarker for tracking the direction of hydration state change in real-world settings.

What Do Tokenizer Comparisons Measure? Disentangling Granularity, Root Assignment, and Word-Level Readout across Amharic and Hausa

arXiv:2611.14818v1 Announce Type: new Abstract: Morphology-aware tokenizers are expected to help morphologically rich languages, and their reported gains over byte-pair encoding (BPE) are usually credited to their linguistic units. A comparison with BPE cannot support this attribution, since such a tokenizer also segments text more finely (granularity) and, on token-level tasks, changes how much the sub-token readout matters. We separate both factors from the effect of grouping words by their roots (root assignment) in Amharic, a Semitic language whose Ge'ez abugida hides consonantal roots inside syllabic glyphs, with Latin-script Hausa as an orthographic contrast. Our design combines a lossless root-and-pattern tokenizer with controls that change one factor at a time, in two independent implementations. In both, real root assignment improves Amharic named-entity recognition over scrambled roots (+4.9 and +4.7 F1), and root-frequency structure explains little of this gain. In small from-scratch models, the readout choice moves F1 by up to 13.4 points and reverses the sign of every comparison of the root-and-pattern arms with BPE, whereas in two pretrained encoders it moves F1 by less than one point. The granularity effect changes sign across languages and implementations. We recommend decomposing tokenizer effects, and reporting the readout, before a gain is attributed to linguistic structure.

When Is Proof-Step Verification Worth Its Cost? Operating Requirements of Certify-or-Decline Verification for Open-Weight Mathematical Reasoning

arXiv:2611.14794v1 Announce Type: new Abstract: Proof-step verification rewrites a solver's answer into typed steps and uses LLM judges to certify or decline. We evaluate this pipeline on open-weight models across GPQA, AIME-2025, and a MATH-500 subset. Certification raises answer-key-match precision by 31 percentage points on the strong GPQA configuration and 46 on AIME, while answering only 35% and 21% of attempts. On near-ceiling MATH-500, gains are at most 5 points at a loss of 13-27 points of coverage. An output-token-matched self-consistency selector obtains 82.9% precision at 35% coverage, compared with the verifier's pooled 87.6% (paired difference: +4.8 points, 95% interval [-8.2, 18.5]). A diagnostic sample reveals a distinct failure mode: 52 of 88 completed problems end in invalid proof construction without ever reaching a judge. Finally, external-LLM acceptance of rendered certified chains is 61.0-67.6%, compared with 87.6% answer-key match. These audits lack human or formal-kernel ground truth. Open-weight auditor pairs show near-zero chance-corrected agreement, which measures consistency rather than accuracy against proof truth. These results motivate separate reporting of execution failures, selected-answer precision, and independently assessed proof validity. We provide per-configuration results and evaluation details in the supplement.

Agent Architecture Cost-Success Trade-offs: A Controlled Evaluation across Enterprise-Style SOPs

arXiv:2611.12953v1 Announce Type: new Abstract: Enterprises increasingly use large language model agents to execute standard operating procedures (SOPs): long, branching, tool-intensive workflows. Yet there is little controlled evidence for choosing among competing agentic architectures. Holding the model, tasks, and tools fixed and varying only the architecture, we compare four designs (a function-calling single agent, a skill-augmented agent, centralized multi-agent orchestration, and a triage swarm; eight arms) across 13 enterprise SOPs, and introduce a skill-definition quality rubric to build matched high and low-quality skill corpora. Three takeaways emerge. (1)~\textbf{There is no universal best architecture; the right choice is SOP-dependent.} On a given SOP, task success ranges from under $10\%$ on the hardest to near $100\%$ on the easiest; no architecture is statistically separable in aggregate, and added structure only plausibly pays off on complex, branching SOPs (a directional, not yet significant, trend). (2)~\textbf{Added architecture cost must be justified by a task-success gain.} Every coordinated arm (multi-agent or skill-augmented) spends $1.4$--$4.7\times$ the tokens and $2.6$--$5.8\times$ the latency of the single agent for gains that are usually small and rarely significant, making the single agent the efficient default; high-quality skills do lift success (to $73.1\%$ from $67.4\%$) while degraded skills fall below it. (3)~\textbf{Implementation quality outweighs the single- versus multi-agent choice.} A plausible multi-agent implementation loses to the single agent, yet a targeted redesign of it becomes the only arm to significantly beat the baseline ($75.6\%$ vs.\ $67.4\%$).

Auditable Human-Agent Teams for Autonomous X-ray Experimentation

arXiv:2611.12933v1 Announce Type: new Abstract: Closed-loop autonomous experimentation can accelerate scientific discovery, but implementation effort limits accessibility. Agentic systems may lower this barrier, although sustained operation on physical instruments remains largely unexplored. Here, using coding-agent infrastructure grounded in validated scientific methods, we demonstrate autonomous synchrotron X-ray experiments on nanoparticle superlattices under evolving, human-defined objectives. Complete human-agent and agent-agent transcripts enable an audit that reconstructs decision provenance and characterizes the division of cognitive labor. Across three experiments, 61-67% of messages involved interpretation or judgment, while the share of methodological and scientific contributions increased from 46% in the setup-focused first experiment to 56-61% in later experiments. Provenance showed that 59-68% of error corrections involved multiple participants and that about half of all decisions incorporated the input of another participant. These results may guide design principles for multi-agent scientific systems and inform future agent harnesses, model training, and accessible autonomous experimentation.

BioMem: Agentic Memory for Biomedical LLM Agents

arXiv:2611.13426v1 Announce Type: new Abstract: 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.

Branch-and-Commit for Frozen RL Agents: When and Where to Spend Test-Time Compute

arXiv:2611.13003v1 Announce Type: new Abstract: Reinforcement learning agents are often deployed with policies that cannot be retrained or modified. In digital environments whose states can be copied and resumed, an operator has another option: sample several continuations from the current state, evaluate them, and continue from the endpoint of the selected branch. We call this procedure branch-and-commit. We study when it improves return, what produces the improvement, and where its computational budget should be spent. Across four environments, two frozen policies, five continuation verifiers, and a model-free selection rule, we find that the benefit of branching depends on dispersion: the difference between the best candidate return and the mean candidate return. Branching provides no benefit in low-dispersion settings, while increasing return by 12.8% in our main Craftax-Medium setting. Most of this gain comes not from selecting better actions, but from selecting among different stochastic realizations of the same action. Several competent verifiers achieve similar episode returns, suggesting that verifier quality is not the main constraint at our operating budget. Instead, computation is better allocated to states where useful dispersion is likely. On held-out seeds, gating branching using predicted dispersion retains 69% of the improvement from uniform branching while using 48% as many branching calls. These results characterize when branch-and-commit is useful and clarify what its gains represent.

Continuous Angular Control in Text-to-Image Generation via Geometric Token Encoding

arXiv:2611.13172v1 Announce Type: new Abstract: Text-to-image diffusion models can render a photorealistic analog clock but cannot control the time it displays. Across four pretrained models, fewer than 1% of generated clocks place both hands within 6 degrees of the requested time; the models instead collapse onto visually dominant configurations such as the canonical 10:10 pose. The same failure appears when a specific fraction of a circular object is requested, where models return complete or conventionally sliced objects. We observe that these two failures share a geometric structure: both target quantities can be parameterized by angular variables. We introduce a geometric token encoding framework that conditions generation directly on these angles. Each angular component is encoded as a periodic sine-cosine feature and projected into a role-specific token embedding that is injected into the prompt before a frozen text encoder. To spatially ground these tokens without requiring localization input at inference, we introduce Stochastic Localization Relaxation (SLR), which stochastically relaxes attention supervision from component-level to object-level to unconstrained localization during training. For clocks, we additionally introduce a hand-swap preference objective that discourages assigning the correct angles to the wrong hand roles. On analog clocks our method reaches 63-78% Acc@6 degrees across evaluation methods, compared with below 1% for the baselines. On circular-fraction generation, it achieves 89.5% Acc@0.10, while maintaining 88.8% on unseen object categories and 84.9% on unseen target values, compared with roughly 20% for pretrained baselines.

Evidence-Grounded Iterative Reasoning for the Diagnosis of Diffuse Cystic Lung Diseases

arXiv:2611.12981v1 Announce Type: new Abstract: Overlapping imaging manifestations complicate the differential diagnosis of Diffuse Cystic Lung Diseases (DCLDs), while limited annotated data constrain supervised model development. Multimodal Large Language Models (MLLMs) can integrate imaging and clinical information, but their diagnostic reasoning remains limited by incomplete volumetric perception and insufficient grounding in disease-specific evidence. We propose DCLDs-EIR, an evidence-grounded iterative reasoning framework coupling diagnostic hypothesis refinement with multimodal evidence acquisition and verification. Multiview renderings of predicted lung and cyst segmentations characterize global cyst distributions and initialize the differential diagnosis. Competing hypotheses guide axial CT inspection to identify local findings that, together with clinical context, inform retrieval from a disease-specific knowledge graph. Accumulated patient observations and retrieved knowledge are jointly assessed for diagnostic support and refutation to revise the diagnostic state. Updated hypotheses and unresolved distinctions guide subsequent inspection, forming a feedback loop between evidence acquisition and hypothesis verification. Multicenter evaluation demonstrates that DCLDs-EIR consistently improves diagnostic accuracy over baselines and enables traceable differential diagnosis through explicit links between diagnostic hypotheses, patient observations, and disease-specific knowledge.

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).

From Questions to Decisions: Benchmarking How LLM Agents Acquire Decision-Relevant Information from Human Beings

arXiv:2611.13015v1 Announce Type: new Abstract: LLM agents are increasingly asked to make plans for people, such as scheduling drivers or assigning nurses. However, people's requests are often incomplete, and the supporting information for decision-making may be outdated, requiring agents to ask for clarification. Existing benchmarks typically provide fully specified problems or rely on uncalibrated LLM-simulated users, leaving this ability insufficiently tested. We introduce Ask2Decide, a benchmark of 43 planning tasks that evaluates whether agents ask for missing information that matters for their decisions. An operator holds facts absent from the agent's request and records, and optimization solvers determine which facts change the optimal plan. We also include similarly worded facts that leave the optimal plan unchanged to test whether agents distinguish useful information from irrelevant details. We collect 79 interaction sessions from 19 participants and find that, after an incorrect plan, human replies omit the correcting fact far more often than replies from an uncalibrated simulated user (43% versus 7%). Using these interactions, we calibrate a simulated operator that more closely matches human reply patterns. Evaluating four frontier LLMs, we find that their initial plans are infeasible on 58-98% of tasks when information is missing, compared with 5-28% when all facts are provided upfront. Agents uncover facts that change the optimal plan only slightly more often than facts that leave it unchanged (39-52% versus 32-41%). These results show that obtaining decision-relevant information through interaction remains a major bottleneck for LLM planning. This is ongoing work and will be updated. The date of this version is shown on the first page.

Instance-preserving graded semantic reporting for surgical action triplets

arXiv:2611.13009v1 Announce Type: new Abstract: Complete-or-reject reporting of surgical action triplets can discard a correctly identified instrument and action when the target is uncertain. We introduce instance-preserving graded semantic reporting, which separates instance selection from the choice of semantic detail. Each selected instance retains its predicted mask and labels while hierarchical correctness scores guide instrument (I), instrument-verb (IV), or complete instrument-verb-target (IVT) reporting. The utility-increment rule DELTA weighs the expected gain from finer reporting against its error cost. We compare three decision rules on 5,641 test frames from eight surgical videos using fixed predictions and scores. At the 10% development calibration target, DELTA retains correct partial reports for 80.0% of candidates rejected by a matched complete-or-reject baseline (5,329/6,662). Adding IV increases source-normalized weighted correct-report yield by 19.13 to 23.13 per 100 frames across all three rules relative to separately recalibrated I/IVT controls. Weighted reporting risk changes by -0.77 to +1.01 percentage points, with complete correct-triplet counts unchanged or reduced. Graded reporting makes partial surgical predictions available as localized descriptions, with an explicit trade-off between report utility and completeness.

INTENT: Out-of-Sample, Out-of-Time, and Cold-Start Evaluation on User Event Sequences

arXiv:2611.13069v1 Announce Type: new Abstract: Banks now train foundation models on ordered sequences of customer events such as transactions, app sessions, and authorizations. Deployed, such a model predicts forward in time, often for customers and merchants absent from training, yet each is evaluated on private data, so no two can be compared. Public benchmarks for such sequences rarely test a period the model has not seen; recommendation studies usually hold out each user's last events instead. We present INTENT (INTEraction, Novelty, and Transfer evaluation for next-action prediction), a public benchmark that scores a model in seven cells defined by whether the user, the period, and the target item appeared in training. An audit of 59 public dataset families retains five datasets; the most common defect is the lack of a calendar timestamp that orders events. No date lies within the time span of all five datasets, so each is split at its own cutoff, chosen by one rule fixed before any model is evaluated. Summed over the datasets, every cell holds at least 111,769 users. In reference results the cell determines the best model: a transformer leads on seen items after the cutoff but scores zero on new items, a variant that embeds item metadata helps mainly on MicroLens, and repeating the user's last item leads on new items in four of five datasets. This parameter-free control also shows that on three datasets most of the score lost after the cutoff reflects changed behavior rather than the model. The pipeline, the cutoff rule, and the per-cell user counts are publicly available.

Linear Probes to Bayesian Manifolds: Geometry of Arithmetic in Large Language Models

arXiv:2611.12998v1 Announce Type: new Abstract: Linear probes are a standard method to study geometric structures, such as helices and circles, from the activations of large language models (LLMs), and such structure is often taken as evidence that the model computes with it. Probe accuracy alone cannot establish this, because it conflates four separate claims. These are (1) that the concept is decodable from the activations, (2) that they form a specific shape, (3) that the shape belongs to the probed concept rather than to a related concept the model also represents, and (4) that the output of the model depends on the shape. We call these claims decodability, shape, ownership, and causal use, and we test them separately with a four-stage Bayesian pipeline. Stage 2 compares candidate shapes and selects the best one only when its evidence passes a predefined threshold. Stage 3 removes a set of related concepts, specified in advance, and tests whether the shape is owned or inherited. Stage 4 intervenes on the shape to test whether the output of the model depends on it. We apply the pipeline to two-digit addition and multiplication in GPT-J 6B, Llama 3.1 8B, and Pythia 6.9B. Nearly every concept is decodable, but the decoded directions overlap substantially across concepts. Every strongly supported comparison selects a ribbon, which is a helix whose radius changes with the value, or a torus. Of 1,330 cells with an ownership verdict, 19% of shapes are owned and 81% are inherited. Owned and inherited shapes move the output about equally often. Intervening on the shape retains approximately 61% of the effect of intervening on the full probe while using at most five dimensions, against a median of nine for the probe. Our results indicate that a probe verdict alone overstates what a model computes with, and we propose that probe results be reported as verdicts on all four claims rather than as a single accuracy.