Skip to content
AI NEWS JOURNAL
AI, IN PERSPECTIVE

Search AI news

Search headlines, summaries and articles to connect the dots.

Newest first

Process-Knowledge-Guided Latent Diffusion for Generative Design of Porous Architectures: A Case Study in Aerosol Jet Printing

arXiv:2611.25264v1 Announce Type: new Abstract: Printed porous layouts interleave functional material with open channels, and the layout that uses the material best depends on how the process settings shape the deposit. Generative models propose layouts quickly, but a generator that does not receive the process setting samples from the same distribution at every setting, even when the process-dependent score ranks candidate layouts differently. This paper examines whether process knowledge can steer a latent diffusion generator toward layouts suited to a given process setting in a prescribed design domain. We formulate a process-dependent utilization score in which a process model sets the length over which the utility of printed material decays with distance from a channel. The generator combines a vector-quantized variational autoencoder (VQ-VAE) latent representation, a topology U-Net, a ControlNet domain branch, and a separately trained regressor that predicts the score from noised latents; the domain branch conditions each denoising step and the regressor gradient steers it. The method is applied to aerosol jet printing, where the atomizer and sheath gas flows enter the score through the flow-focusing relation. Across seven modeled flow settings, the lowest-scoring candidate layout of a domain differed between settings in 191 of 500 domains. In a sampling study, the mean directly computed score on three validation domains and one test domain was 0.642 with domain conditioning alone, 0.636 at a weight of 300, and 0.450 at 12,000 (112 scores per weight; lower is better). The results support process-knowledge-guided generation as an approach to process-specific porous design under the stated assumptions.

A Model of Diverse Sampling from Language Models

arXiv:2611.25003v1 Announce Type: new Abstract: Language models often produce high-quality individual samples but poor sets, with repeated samples clustering around the same semantic modes. We formalize diverse generation as sampling a size-$K$ subset of complete strings from a determinantal point process (DPP), where the likelihood encodes item quality, and the embedding geometry encodes repulsion between similar outputs. Exact inference over all strings is intractable, so we propose PoolDPP: draw a finite candidate pool from a tractable proposal, importance-weight the candidates toward the target quality, and run a $K$-DPP sampler on the induced pool kernel; a greedy MAP variant provides deterministic selection. We prove a finite-sample bound with an explicit constant: the KL divergence between the sampler's output and the target $K$-DPP decays as O(1/N) in the pool size. Empirically, PoolDPP extends the quality--diversity frontier on open-ended generation and the precision--recall frontier on ambiguous text-to-SQL (Ambrosia), and on mathematical reasoning (GSM8K), it raises diversity at about the same pass@$K$.

A Verifier from the Past: Test-Time Compute for Time Series Foundation Models

arXiv:2611.25171v1 Announce Type: new Abstract: Test-time compute (TTC) improves language models by spending additional inference compute to generate, verify, and select among candidate outputs, but this recipe has not transferred directly to time series forecasting. A language model can verify a proof or vote across samples, but a forecaster must be causal; it cannot use future information to score a candidate. We propose Backtest-Gated Augmentation Averaging (BGAA), a TTC algorithm for time series forecasting that builds a causal verifier from the series's own history, using rolling-origin backtests to score structured views of the context, discard those that underperform the untransformed context, and average the survivors' predictive distributions. BGAA improves every tested TSFM-benchmark combination, reducing MASE by 0.5-1.0% on GIFT-Eval, 1.3-5.0% on fev-bench, and 1.7-5.9% on BOOM, where a 313M model with TTC outperforms the 1B and 2.5B models in its family. We further show that useful views have structured, model-dependent effects consistent with a prior-transport interpretation. On the 53 cross-fittable GIFT-Eval configs, the deployed policy gains only 0.4-1.3% CRPS, versus 1.4-4.1% when choosing the best view for each dataset configuration and 17-19% when choosing the best view separately for each test window. This gap suggests that a major remaining challenge is window-level view selection.

Closing the Analyze-to-Design Gap: An Authorization-Bounded Agent for Self-Driving Laboratories

arXiv:2611.25043v1 Announce Type: new Abstract: Artificial intelligence (AI) agents are increasingly integrated into self-driving laboratories (SDLs) to plan experiments and interpret results. However, analytical reasoning alone does not make a laboratory autonomous: autonomy in an SDL is achievable only when the agent can be trusted to act on physical hardware unsupervised and knows when to defer to a human. Here, we present an architecture that integrates an agent at the Analyze-Design interface of the Design-Make-Test-Analyze (DMTA) cycle, enabling agent-driven measurement selection and data interpretation while ensuring that state-changing operations are authorized, validated, and auditable. We demonstrate this architecture on a liquid chromatography-mass spectrometry (LC-MS) instrument in an oxidant-screening campaign for synthesizing an organic dye. Within human-defined campaign limits, the agent submitted measurement jobs, interpreted mass spectrometry (MS) and diode-array detection (DAD) data, flagged failed experiments, and produced an oxidant ranking. This ranking matched that obtained from scripted peak integration, and organic chemists confirmed the agent's spectral interpretations and failure assignments. In a simulated environment that replays the archived LC-MS data, where the agent chose which wells to inject under varying injection limits, it measured the highest-yielding condition at about its seventh injection, against 14.5 on average for a random order, and identified it in every campaign with a limit of eight or more injections. An agent can therefore close the analysis-to-design gap while its actions stay bounded and verifiable, which is a step toward greater autonomy for SDLs.

TALD: Transferability-Aware Logit Disentanglement for Cross-Modal Knowledge Distillation

arXiv:2611.25126v1 Announce Type: new Abstract: Cross-modal knowledge distillation can incur negative transfer when source-modality supervision conflicts with the target modality's discriminative characteristics. We propose TALD, a two-stage framework for Transferability-Aware Logit Disentanglement in class-aligned logit space. Guided by task-dependent self-benefiting and cross-modal exclusivity criteria, a Residual Disentangler decomposes logits from a teacher and an independently trained target-modality reference student into modality-general and modality-specific components. A Sample-Adaptive Modulator regulates decomposition constraints and sample-specific fusion weights. The learned modality-general components are then adaptively fused to supervise the final student. Across five benchmarks spanning vision, audio, and text, TALD achieves the largest mean accuracy gain and highest positive-transfer frequency among the compared methods. Across ten transfer directions with five runs each, TALD improves over training without distillation in 84% of evaluations, compared with at most 70% for the baselines, while offering a favorable accuracy-efficiency trade-off. We plan to release the code.

Thinflow: Organizing HPC Benchmarking Campaigns as Rerunnable Directory Trees

arXiv:2611.26526v1 Announce Type: new Abstract: A benchmarking campaign provides the measurements behind an HPC performance study: code variants are built, swept over core or node counts, and measured repeatedly. Moving such a campaign to another supercomputer or rerunning it months later requires work that is not part of the study itself. Job scripts are adapted to each supercomputer, and what exactly ran is often reconstructed afterwards. We present Thinflow, a small Bash engine for organizing HPC benchmarking campaigns as rerunnable directory trees of tasks and runs. A task is a directory with a short shell script next to the code. For each sweep, one command builds the code, runs the benchmarks, and plots the results. Thinflow derives the execution order from declared dependencies and writes it to a plain-text manifest. Per-supercomputer profiles submit the runs to a workload manager such as Slurm. Each run writes a provenance record of its environment, hardware, and parameters. The engine ships inside the campaign directory, so a campaign stays rerunnable without installing a separate tool and, once completed, may serve as a paper artifact. In a case study, we ran one campaign with identical task files on three Slurm partitions of two supercomputers, on up to 16 nodes. This strong-scaling study of matrix multiplication in C/OpenMP, C/MPI, and Julia/MPI comprises 37 task files with 217 non-comment lines. Its 455 benchmark runs recorded what differed per partition, chiefly profiles and command-line parameters. This suggests that such a campaign needs little structure beyond plain shell files.

Watch Where You Linearize: Predicting and Correcting Model-Merging Interference

arXiv:2611.25000v1 Announce Type: new Abstract: Model merging combines separately fine-tuned models by adding their task vectors to a shared backbone. Interference is typically assessed only after constructing and evaluating a candidate merge, making coefficient search expensive. We show that a first-order directional derivative of task A's outputs along task B's vector can predict interference, but only when evaluated at the appropriate model. At the receiver model, this derivative forecasts nearby merges per task and per example, halves the accuracy error of a no-change baseline, and identifies most examples whose predictions flip. For full multi-task merges, a single evaluation is insufficient; two evaluations along the merge path are needed. At the shared pretrained backbone, however, the derivative is nearly orthogonal to realized changes in our vision experiments and only partly predictive in LoRA language-model pilots. Controls link this failure to displacement along the receiver's own task vector, not simply parameter distance; the mismatch appears within the first few dozen fine-tuning steps. Freezing receiver-local responses produces a surrogate affine in the merge coefficients. We use it to construct LR-Merge, a label-free correction that optimizes coefficients through convex teacher matching on unlabeled probes, without backpropagation through merged networks. LR-Merge improves all tested closed-form, isotropic, and sparsified base mergers, with gains generally increasing as more tasks are merged. Using pretrained rather than receiver-local responses instead degrades the base merge. LR-Merge matches exact teacher matching through a merged network, which takes roughly three times longer to converge and performs similarly at an equal compute budget. Once built, the surrogate answers new coefficient and task-subset queries in seconds.

Accelerating Distributed Digital Ising Machines

arXiv:2611.23883v1 Announce Type: cross Abstract: Ising machines, dedicated accelerators built around the dynamics of Ising-like spin models, have emerged as a leading platform for combinatorial optimization and probabilistic sampling. The problems that motivate them require millions of spins, far beyond the on-chip capacity of any single device, so a practical Ising machine must distribute its computation across many devices. Prior work reports throughput at a single operating point, leaving the binding resource unidentified and platform crossovers measurable only by experiment; no calibrated cross-platform cost model of distributed Ising machines has been reported, nor have systematic techniques to accelerate their Markov chain Monte Carlo (MCMC) workloads. This work presents a calibrated cost model that decomposes digital Ising sweep time into compute, memory, synchronization, and inter-device communication terms, validated with 4.8% mean error across three GPU generations, FPGA and ASIC up to 1 million spins. Guided by the model, we address each term in turn and propose: (i) a data placement strategy, worth up to 4.2$\times$ speedup; (ii) multi-rate and bounded-staleness schedules that relax the barrier that Gibbs sampling imposes, cutting time-to-solution (TTS) by up to 3.16$\times$ on three benchmarks; and (iii) a phase-offset streaming dataflow that hides boundary exchange behind interior computation whenever a link satisfies a latency-transparency criterion, yielding up to 1.25$\times$/1.74$\times$/2.01$\times$ on an H200 pair, a ZCU106 pair, and a four-board XEM8320 $2{\times}2$ mesh. The streaming dataflow is also implemented on a fabricated four-chip 28 nm ASIC sustaining 28.1 Gflips/s at 4.81 pJ/flip.

FACT: Fixed-Budget Allocation and Cross-Step Training for Speculative Decoding

arXiv:2611.23463v1 Announce Type: new Abstract: Speculative decoding accelerates language-model generation by verifying multiple draft tokens in a single target-model pass. However, better draft predictions and longer accepted continuations do not necessarily produce faster generation. To examine this gap, we present FACT (Fixed-Budget Allocation and Cross-Step Training), an experimental framework for studying how drafter training, candidate allocation, and execution conditions jointly affect acceptance and latency. Using a lightweight feature-prediction drafter with a frozen Llama-3-8B-Instruct target under greedy decoding, we first examine how training translates into accepted tokens. Backpropagating through predicted-feature rollouts with greater weight on early steps improves acceptance by 15-24% over uniformly weighted detached training, yet average rollout proxy cross-entropy does not reliably rank drafters by acceptance. Beyond prediction quality, the placement of draft candidates also matters: with a fixed 256-node budget and matched per-depth widths, replacing static allocation with confidence-guided allocation increases speedup over autoregressive decoding from $1.46\times$ to $2.02\times$ on HumanEval and from $1.31\times$ to $1.75\times$ on MT-Bench. This allocation benefit extends to an independently trained drafter, but stronger training does not consistently amplify it, showing that the two improvements do not necessarily reinforce each other. Even these speed gains depend on execution conditions: changing only host-thread settings can nearly eliminate speedup without changing acceptance. Together, these findings show why training and allocation must be evaluated jointly using both acceptance and end-to-end latency under controlled execution conditions.

FutureBench: A Renewable Framework for Evaluating Knowledge-Context Conflict in LLMs

arXiv:2611.23601v1 Announce Type: new Abstract: Asking a model to check retrieved text can also tell it which source to trust. We introduce FutureBench, a renewable framework for evaluating knowledge-context conflict using timestamped questions, controlled passage triplets, and no-context probes. Across 3,629 questions and six models, Verify (a prompt combining checking with passage priority) raises false-answer adoption in all five models tested on pre-cutoff events, mainly by displacing answers that were correct under Baseline, which allows both passage evidence and model knowledge. On post-cutoff events, only o4-mini increases adoption (+13.7 percentage points). A four-prompt comparison on 902 post-cutoff questions finds that passage priority without checking raises adoption, while moving to Verify lowers it in all five models. Adversarial-Aware, which warns of misleading passages and prioritizes model knowledge, restores 82.5-98.9% of displaced pre-cutoff answers with under one point of change in true-passage accuracy. FutureBench supports repeated evaluation on fixed cohorts and extension to new events as models change.

GRAFT: Execution-Guided Genetic Recombination for Agent Design

arXiv:2611.23509v1 Announce Type: new Abstract: Automated agent design searches for executable architectures that improve LLM-based agent performance, with evolutionary methods using language-model crossover to recombine promising mechanisms for task decomposition, model coordination, and verification. This search, however, is typically guided by aggregate fitness alone, which hides the instance-level successes and failures behind each score. We introduce Execution Guided Genetic Recombination for Agent Design (GRAFT), which uses aggregate fitness for selection and survival and parent-specific execution evidence to guide semantic crossover with integrated mutation. Its feedback harness reuses bounded samples of successful and failed runs from fitness evaluation, withholding reference answers and requiring no additional task executions. We evaluate GRAFT across six reasoning and programming settings using open-weight optimizer and executor models. It achieves a mean test accuracy of 82.0%, exceeding the strongest baseline by 9.8% points. Controlled ablations show that adding execution evidence improves test accuracy with initialization, search rules, repair policies, and evaluation budgets held fixed. Our results support the value of parent-specific execution evidence for evolutionary agent design.

GRAFT: Graph-Regularized Prototype Fields for Realistic Test-Time Adaptation of VLMs

arXiv:2611.23591v1 Announce Type: new Abstract: Unlabeled test data can improve the zero-shot classifier of a vision-language model, yet transductive and test-time adaptation methods often fall below the zero-shot baseline when only a few vocabulary classes are present in a batch or when streams are correlated. We trace this failure to a shared design: every class owns independent parameters estimated from its own samples, so absent or rare classes are left at biased text anchors or fitted to foreign samples. We propose Graph-Regularized Anchor-displacement Field Transduction (GRAFT), which models the adapted prototypes as text anchors plus a displacement field that is partially pooled over a semantic class graph through a Gaussian Markov random field prior, together with an explicit class-presence model. Inference is closed-form and gradient-free on cached features. Anchor-only shrinkage and a single global shift are the two extreme settings of the prior; absent classes are provably harmonic extensions of the present ones, and the pooling strength can be chosen by empirical Bayes. Under the realistic batch and online protocols on eleven benchmarks and four ImageNet shifts, GRAFT improves over the zero-shot model in every setting and over anchor-based transduction where few classes are present.

Layer-Matched Sparse Attention for Low-Redundancy 3D Reconstruction

arXiv:2611.23554v1 Announce Type: new Abstract: VGGT reconstructs cameras and geometry from unposed images in one pass, but its global attention grows quadratically with the number of views. Training-free accelerators thin this attention and work well on redundant video, but lose pose accuracy when views share little geometry. This paper traces these losses to two kinds of cross-view information that a uniform sparse approximation can fail to preserve together: diffuse layers aggregate distributed context, whereas matching layers depend on query-specific interactions that carry little attention mass. A per-layer statistic measures the log-mass missed by a mean token. Its row-free proxy is the logit variance, and it selects the estimator for each layer. Diffuse layers use shared, importance-sampled hub keys, with log-weights that make the softmax numerator and denominator unbiased. Matching layers stay exact: every aggressive thinning tested cost accuracy. At equal cost, the selected layers outperform the four shifted windows tested. On 13 unseen low-redundancy scenes with 32 views, LMSA stays within 1.4 AUC@5 of dense attention, compared with 10.3 for AVGGT and 4.5 for SAF3R; its difference from contemporaneous GPA-VGGT is not significant, and only its gains over AVGGT pass both scene-level tests. With 100 views, it stays within 1.9 AUC@5 while its aggregator runs $2.0\times$ faster. Used post hoc to choose the layers GPA-VGGT protects, the statistic improves it by 1.2 AUC@5 for 5% more FLOPs. Applied untuned to $\pi^3$, it beats AVGGT but remains further from dense accuracy.

Model Discovery for Virtual Chaos Engineering with Incomplete Tracing: An eBPF Study

arXiv:2611.23313v1 Announce Type: cross Abstract: Model-based fault experiments can assess the consequences of component failures before intervening in a running application. Recent work on model discovery and graph simulation acquired operation models from application telemetry and used them for availability prediction. Applying this approach to operational legacy systems raises an acquisition problem: the required traces may be absent or incomplete, while introducing them can require changes across heterogeneous components. These systems continue to serve business operations during extended maintenance and modernization. External observation offers a possible route to model-based fault analysis before a tracing retrofit, but collecting telemetry does not by itself establish that an executable operation model can be obtained. We investigate this route using extended Berkeley Packet Filter (eBPF) observations and the existing graph-based modelling approach. A 108-case study varied application software development kit (SDK) tracing across absent, partial and full conditions in three public applications. Discovered dependencies were combined with declared operation requirements, deployment facts and scenario assumptions; model forecasts preceded independent execution checks. With SDK tracing disabled, the tested OpenTelemetry eBPF Instrumentation path supported all seven operations in two applications across three repetitions. Forecasts agreed exactly in 17 of 21 operation-repetition groups; the maximum state-weighted discrepancy was 4.22 percentage points. Combining observed context links supported nine additional groups under partial tracing. The results establish an acquisition path for virtual fault experiments under the tested observation constraints and identify limits imposed by protocols, evidence association and operation semantics.

Reference-Guided Data Curation for 3D Medical Vision-Language Pretraining

arXiv:2611.23484v1 Announce Type: new Abstract: Training 3D vision-language models (VLMs) on large medical imaging datasets collected across institutions is constrained by data quality. Vast, uncurated pools of image-report pairs are characterized by redundancy, class imbalance, and cross-institutional heterogeneity. We address this challenge of data curation for improving 3D VLM pretraining by proposing an offline, training-free, reference-guided data curation recipe. Our approach leverages a small, high-quality reference set to independently extract quality signals for each modality of a large, uncurated pool and merges them through a rank-aggregation step into a single ranking that guides sample selection. We validate this method on datasets comprising 35k 3D abdominal CT volume-report pairs from two institutions and 58k 2D chest X-ray image-report pairs. On 3D CT, using our reference-guided data curation strategy, our curated model matches the zero-shot disease classification performance of a state-of-the-art baseline (Merlin) while using $3.3\times$ less training data and a reference set of 5k samples. On 2D chest X-rays, our method yields a curated model that matches or exceeds the zero-shot disease classification performance of a model trained on the full pool while using up to $10\times$ less training data and a reference set of only 74 samples. Ablations show that the selection is stable under resampling of the reference, that smaller reference sets still outperform random sampling, and that label rebalancing does not improve performance. To support reproducibility, we release our code.

RefinementGuard: Counterexample-Directed Semantic Qualification for Hidden Agent-Tool Execution

arXiv:2611.23611v1 Announce Type: new Abstract: Tool runtimes can lower an agent-selected operation into a hidden parallel graph, reducing latency and context while risking semantic mistranslation. We introduce RefinementGuard: translation validation for hidden, effectful agent-tool lowerings, coupled to artifact-bound, fail-closed runtime admission. Qualification compares a candidate graph with an independent reference on isolated state using explicit parent-operation obligations. We evaluate two repository-level contracts, bug localization and patch validation, on controlled repositories. Among 100 activated, non-equivalent development/qualification faults adjudicated by one human, parent-semantic validation rejects 49/100 (operator-cluster bootstrap 95% CI [33.7%, 64.8%]). The objective is selective admission, not maximum mutation rejection: exact equality rejects 100/100 faults but also all 60 valid variants, whereas obligation-based validation accepts 60/60; its 51 residual misses identify concrete specification gaps. On a frozen disjoint input split, the held-out B2-B3 result is localization-only because 21 patch mutants fail structurally; B3 rejects 38/56 localization defects. In a controlled asynchronous-delay regime, the localization serial/guarded median-latency ratio is 4.50x, and hidden execution exposes 74.3% less parent-visible context than visible execution. Patch validation shows no demonstrated speedup; zero-delay runs expose overhead. End-to-end qualification thus detects lower-layer misses while preserving localization's parallelism.

When Does Subspace Direction Matter for LoRA? Regime Analysis of the Magnitude Principle in Few-Shot Adaptation

arXiv:2611.23428v1 Announce Type: new Abstract: We show that the importance of a LoRA adapter's initial input-subspace direction is regime-dependent, and we characterize this dependence in theory and experiment. This refines the recent view that spectral initializers help through update magnitude alone. Direction matters most when labeled data are scarce relative to the representation dimension and the activation subspace used for initialization captures supervised task structure. In this regime, the adapter's row span remains sensitive to its starting point, making task-aligned initialization especially useful. A spiked-covariance perturbation analysis characterizes the finite-sample stability of the estimated subspace. A first-step analysis of LoRA training explains why the subspace's alignment with the task matters and motivates an alignment coefficient $\rho$ that closely tracks gains across classification tasks. We test this account with PivotLoRA, which initializes adapters from activation PCA at prediction-relevant positions and applies one output-preserving orthogonal pivot early in training. In few-shot BERT-base classification, PivotLoRA achieves the best average rank among 13 methods and improves over vanilla LoRA in 29 of 30 settings, even though it starts from a smaller input-factor initialization norm. Additional classification experiments extend the evaluation to 7B and 8B Llama decoders. Classification gains diminish with more labeled data, consistent with the proposed regime dependence. A targeted extension to completion-only reasoning also shows improvements on GSM8K and StrategyQA. These findings identify conditions under which initialization direction matters, with subspace-task alignment offering a low-cost diagnostic before training for classification.

Consistency Is Not Correction: Separating Confidence, Abstention, and Recovery in LLM Agents

arXiv:2611.21651v1 Announce Type: new Abstract: A controller that resamples an answer, escalates to a larger model, or abstains is usually justified by a confidence signal. Three claims about such a signal do not follow from one another: it ranks answers (discrimination), it supports withholding some of them (selective prediction), and it changes an answer for the better (correction utility). We give an exact decomposition of the third into gated recoveries minus gated harms, which bounds whether a comparison can succeed before it is run: a gate that only replaces a proposal can act only where the base policy was already wrong, so the initial error count caps the discordant pairs any correct gate can produce. When that cap falls below the count a paired test needs, the experiment cannot return a positive result for a working mechanism, nor distinguish one from a gate that never fired. We measure all three targets on two model scales and find the bound binds. In a small-model suite a perturbation-agreement verifier reaches within-policy AUROC 0.990 (95% interval [0.95, 1.00]) while every episode outcome matches single-sample decoding at 4.32 times the call count. In a controlled 840-episode study no verifier replaced the proposal on any of the 60 GSM8K decisions, so that comparison tested nothing; the entire GSM8K gain across six controllers is one task, and it is the task whose identical temperature-zero repeats disagreed, making it the one answer the service does not reproduce. On the diagnostic set the gate could reach four errors while the test needed six, so even a perfect gate had no significant result available. We report two checks that would have caught this before any budget was spent, and show that tie handling in rankings, tied selection cutoffs, and unequal output limits each move a headline number in the same dataset. The binding constraint on this literature is measurement resolution, not the choice of signal.

From Preference Data to Personalization: Tracing Sycophancy in Large Language Models

arXiv:2611.21650v1 Announce Type: new Abstract: Sycophancy in large language models is often measured as agreement with the user, yet agreement can be appropriate when the user's claim is valid, their preference is genuine, or their experience warrants acknowledgment. We study sycophancy as inappropriate agreement: validating a claim, preference, or self-assessment when the context calls for correction, qualification, or resistance. We introduce a validity-aware sycophancy grader and SycoDrift, a persona-targeted evaluation set of 3,298 validity-annotated probes with controlled annotations for user openness and stakes. We use this framework to analyze three sources of agreement pressure: human preference data, automated judge specification, and in-context personalization. First, preference for validating responses in human datasets (Chatbot Arena, HH-RLHF, Community Alignment) is weak and concentrated where agreement is often appropriate. Second, in a selected-value, single-model experiment, emphasized judge rubrics produce consistent but modest increases relative to paired suppressed rubrics, demonstrating a plausible mechanism for training-time amplification. Third, in off-policy evaluations across 14 models, inference-time personalization increases inappropriate agreement, with the largest effect from matched user profiles placed directly before the query.

Latent Fact Composition under Next-Token Training

arXiv:2611.21921v1 Announce Type: new Abstract: Controlled studies of latent two-hop reasoning train transformers to predict only the answer, whereas language models are trained to predict every token. We ask whether a model trained on every token learns to answer two-hop questions whose intermediate entity appears in no two-hop training example. We build a synthetic corpus with such questions. On this corpus every term of the loss can be computed exactly. A looped transformer, a standard transformer and DiscoLoop answer almost none of these questions. We therefore introduce the Baton family of looped transformers. They pass the intermediate entity to the second loop as a token embedding. When trained to predict only the answer, two of its three variants answer these questions almost perfectly. We then relate the success and failure of each model to three design conditions. Experiments that change one condition at a time agree with this hypothesis. When Baton is trained on every token, however, its accuracy on these questions rises and then declines. We further split the gradient of the next-token loss into three terms. Removing the two terms that depend on the particular training set raises the peak accuracy but does not remove the decline. The decline arises in the second hop while the model still recalls both facts.