AI papers — 2026-06-14

Problem: Standard speculative decoding uses a binary accept/reject rule that forces all rejected tokens through the expensive full verifier, even when moderate verification would suffice.

Method: VIA-SD introduces a three-tier draft–verify pipeline. A lightweight drafter generates candidate tokens; a "slim-verifier" — a routed intra-model subnetwork of the full verifier derived via Dynamic Intra-Model Routing (DIMR) — handles medium-confidence tokens; and the full verifier handles only the hardest cases. The routing policy is motivated by a KL-divergence decomposition showing that a multi-step verification path can reduce cumulative divergence relative to a direct binary accept/reject step. DIMR performs an offline search over layer-skipping masks to identify which layers to bypass, requiring a one-time search cost per model pair.

Key results:

• Rejection rates reduced by 0.10–0.22 across four task types (QA, summarization, translation, coding) and multiple model families
• 10–20% speedup over strong speculative decoding baselines
• 2.5–3× acceleration over non-drafting (greedy) decoding
• For 7B→70B model pairs, rejection rates drop from ~0.22–0.30 to 0.14–0.16; speedups above 2.3×–2.5×
• On encoder–decoder models (XSum, CNN/DailyMail), 1.2×–1.4× additional speedup vs. baselines with no ROUGE degradation
• Peak memory overhead: only 1.04× versus standard speculative decoding

Why it matters / caveats: VIA-SD generalizes the speculative decoding paradigm beyond binary decisions and is compatible with existing SD frameworks without retraining. The one-time offline DIMR search cost is amortized across inference sessions, but the reported speedups cover online decoding only, not end-to-end including offline search.

Problem: Multimodal image fusion methods that build shared representations on 2D spatial feature grids cannot efficiently regulate image-level global appearance (illumination, contrast, perceptual tone) because these factors are entangled with local spatial details.

Method: The paper proposes a hybrid fusion architecture that pairs a frozen pretrained 1D image tokenizer (e.g., TiTok) with a conventional 2D fusion backbone. The 1D token space serves as a compact carrier for global base/appearance factors, while the 2D spatial pathway handles local detail restoration. A lightweight Selective Token Editing (STE) module sparsely identifies and edits a small subset of appearance-sensitive token dimensions in 1D space; a token-to-map interface projects the edited tokens back into spatial domain for injection into the 2D fusion stream. No additional loss terms beyond standard reconstruction losses are required.

Key results:

• Best or second-best performance across four benchmarks: M3FD, RoadScene, TNO (infrared–visible), and Harvard Medical Image datasets
• Consistent multi-metric improvements in EN (information entropy), SD (standard deviation), SCD, and SSIM on all evaluated datasets
• Improved downstream object detection and semantic segmentation on M3FD and FMB datasets compared to state-of-the-art fusion baselines

Why it matters / caveats: The 1D token interface provides a principled handle for controlling global appearance without redesigning the full fusion pipeline. The approach relies on a frozen pretrained tokenizer, meaning its quality is bounded by the upstream tokenizer's representational capacity; performance on domains very different from the tokenizer's training distribution is not evaluated.

Problem: Existing unified multimodal models (UMMs) lack a single visual tokenizer that handles both images and videos in one shared representation space, forcing either frame-independent encoding or a cascaded VAE+encoder design that misses semantic constraints.

Method: HYDRA-X introduces HYDRA-XTOK, a unified visual tokenizer built on a single Vision Transformer (ViT) split into a Gen-ViT and a Sem-ViT connected by a Generation–Semantic Bottleneck. For video, it uses frame-level causal temporal attention with a minimal 2-frame receptive field (rather than full spatiotemporal attention, which degrades reconstruction) and hierarchical 2×2 temporal patchify across two stages. A lightweight Decompressor lifts compressed latents back to full temporal resolution for joint image and video semantic teacher distillation. For image editing, source and target images are jointly encoded as a length-2 clip inside HYDRA-XTOK, enabling latent-level source–target interaction before the LLM. The full system is built on Qwen2.5-7B-Instruct and trained end-to-end at 7B scale.

Key results:

• Among 7B-scale unified baselines, HYDRA-X is the strongest entry on every reported GenEval and WISE (image generation) column
• On VBench (video generation), HYDRA-X leads all unified entries on Quality Score, Semantic Score, and Total, improving over the closest unified competitor (Show-o2-1.5B) by +1.87 QS, +3.26 SS
• Ablation: source–target interaction inside the tokenizer (STI) improves GenEval by +1.46 and ImgEdit by +0.4 over independent encoding
• Ablation: frame-level causal attention outperforms full spatiotemporal attention; hierarchical 2×2 patchify outperforms single-step 4× patchify on reconstruction (DAVIS rFID)
• Surpasses 1.5B and 7B unified baselines on image and video understanding benchmarks where comparable numbers are available

Why it matters / caveats: HYDRA-X is the first UMM built on a unified image-and-video ViT tokenizer, eliminating the representational mismatch between heterogeneous encoders. A limitation is that it remains below the strongest specialist or larger-scale models on some understanding benchmarks.

Problem: Existing dense feedback methods for diagnosing text-to-image generation failures use heatmap-style pixel regression that cannot bind semantic reasons or importance scores to individual defect instances and is not natively compatible with autoregressive VLMs.

Method: The paper proposes Structured Defect Grounding (SDG), which formulates T2I diagnosis as set prediction over (location, type, reason, importance) tuples. SDG-30K is a 30,096-image dataset with bounding-box-grounded defect annotations across four T2I generators (Flux1.dev, Flux2, Z-Image, LongCat/SANA), annotated by 112 human annotators and enhanced with Gemini 3 Pro for descriptions and importance scoring. A Qwen3-VL-4B model is fine-tuned in two stages: SFT with coordinate jitter followed by composite-reward GRPO with spatial, description, and importance rewards. SDG predictions are converted into importance-weighted spatial reward maps for diffusion alignment via BoxFlow-GRPO.

Key results:

• SDG detector achieves BoxF1@0.5 of 0.263 (artifact) and 0.387 (misalignment) vs. human upper bound of 0.278/0.409
• GPT-5.4 obtains misalignment BoxF1@0.5 of ~0.007 and artifact BoxF1@0.5 of 0.035 vs. SDG's 0.263/0.387
• Gemini 3 Pro reaches artifact 0.200 / misalignment 0.307 BoxF1@0.5
• SDG achieves substantially higher misalignment F1 (0.655 vs. 0.250/0.007 for leading VLMs)
• Importance accuracy @IoU 0.1: ~0.887–0.893 for both defect types
• SDG-guided BoxFlow-GRPO rewards consistently improve T2I alignment beyond scalar preference optimization

Why it matters / caveats: SDG establishes a unified instance-level interface for diagnosing, evaluating, and aligning generative models in a way that heatmap-based methods cannot. The dataset is limited to four specific T2I generators, and out-of-distribution generalization to new generators or prompt distributions is not fully characterized.

Problem: Deep search agents that follow a single linear search trajectory cannot efficiently handle situations where multiple plausible directions need to be explored, abandoned, and revisited under limited compute budgets.

Method: TreeSeeker is an inference-time framework that organizes multi-step web search as branch-and-return search over tree-structured states. Each sub-goal is associated with a search tree whose branches represent candidate queries, source families, or hypotheses. The controller, TreeSearch, reads all sub-goal trees in one pass and assigns one of three operations — EXPLOIT (deepen a promising branch), EXPLORE (try an uncertain alternative), or PRUNE (abandon and return to an earlier branch point) — guided by a textual UCB scoring rule over three ordinal semantic signals: value, uncertainty, and risk. TreeMem maintains per-branch summaries of evidence, conflicts, failure cues, and progress, keeping different search directions structurally separated so their outcomes can inform future decisions.

Key results:

• 56.3 on XBench-DeepSearch (outperforming Flash-Searcher by 5.6, IterResearch by 12.3, and Tongyi-DeepSearch by 11.3 points)
• 47.0 on BrowseComp (best among open-source baselines evaluated)
• 43.0 on BrowseComp-ZH (best among open-source baselines evaluated)
• Ablation on XBench-DS: removing textual UCB signals reduces performance by 4.3 points; disabling branch-and-return operations reduces performance by 8.3 points

Why it matters / caveats: The results confirm that structured branch-and-return control complements stronger reasoning and tool execution for complex web research tasks. Evaluations use gpt-5.2 as the base model; performance with smaller open-weight models and compute cost at scale are not fully characterized.

Problem: Standard adversarial robustness evaluations report attack success rate (ASR) at a fixed query budget, treating all attack strategies as equally costly, which obscures the true computational effort an adversary must expend to compromise a model.

Method: The paper introduces a compute-aware evaluation framework that parameterizes attack trials by cumulative FLOPs (floating-point operations) rather than query count. Risk-compute curves plot attack success probability against cumulative FLOPs as the query budget varies. Two summary metrics are derived: C@τ (FLOPs required to reach τ% attack success, with τ=0.5 as default) and average efficiency (AE, normalized expected risk per FLOP). The framework is applied across three attack families (GCG gradient-based, iterative refinement via an auxiliary LLM, and template-based attacks), ten models spanning three families (Llama, Qwen3, Gemma) at four alignment stages (base, SFT, DPO, RLVR), and two jailbreak benchmarks.

Key results:

• Alignment training has non-monotonic effects on compute-space robustness: intermediate SFT sometimes outperforms later DPO or RLVR stages
• Scaling model size substantially reduces gradient-based attack effectiveness but has limited impact on cheap template-based attacks
• Compute cost to breach a model varies by up to ~5× across harm categories within a single model
• Safety-aligned RL increases aggregate adversarial cost while leaving some harm categories disproportionately accessible
• Gradient-based attacks optimized on a surrogate model can transfer to a target at a fraction of the original cost

Why it matters / caveats: The FLOP-based framework provides a hardware-invariant measure of adversarial effort that translates to operational costs and enables fair cross-model and cross-attack comparison. The framework covers white-box and black-box attacks but is evaluated only on open-weight models; closed-weight models cannot be fully analyzed.

Problem: Prior work found that LLM psychometric self-reports (SR) fail to predict behavior, but these studies used broad Big Five personality traits that are known to be weak predictors of specific behavior even in humans, leaving open whether the dissociation is a property of LLMs or an artifact of the measurement framework.

Method: The paper runs a 2×2×2 factorial experiment varying (i) psychometric instrument (Big Five vs. Theory of Planned Behavior/TPB, which uses behavior-specific TACT-anchored items), (ii) session context (shared conversation thread vs. separate sessions), and (iii) identity induction (parameter-grid perturbation vs. persona prompting). Experiments cover four behavioral tasks (risk-taking via Columbia Card Task, sycophancy via an Asch paradigm, honesty via confidence calibration, and implicit bias via IAT) across 11 frontier LLMs.

Key results:

• Within-session TPB self-reports reach the human meta-analytic baseline for SR–behavior coherence (mean r = +0.40); Big Five does not predict behavior
• Cross-session coherence survives for behaviors anchored outside the prompt (implicit bias, partially honesty) but collapses for context-driven behaviors (sycophancy)
• Implicit bias reproduces near-perfectly across sessions; sycophancy shows no cross-session reproducibility
• Persona prompting stabilizes self-reports across sessions but does not bring behavior into SR alignment
• Human meta-analytic TPB intention–behavior validity: r ≈ .47 vs. Big Five trait–behavior correlation rarely exceeding r ≈ .20

Why it matters / caveats: The findings reframe the SR–behavior dissociation from a property of LLMs to a property of mismatched measurement instruments, and have practical implications for behavioral auditing in deployment. The study is limited to four behavioral tasks and frontier closed-weight models; generalization to open-weight models and other task domains needs investigation.

Problem: Multi-agent LLM systems communicate through text, incurring lossy decode–re-encode overhead at every handoff, while existing KV-cache latent communication approaches are restricted to homogeneous agents (identical model copies) and context-aware settings where the receiver already has the input.

Method: The paper studies KV-cache communication across heterogeneous agents (Qwen3-4B, 8B, 14B in all six cross-model directions). An information-structure analysis using compressed sensing reveals a duality: context-aware transfer is driven by sparse reasoning signals, while context-unaware transfer (where the receiver has no input) requires dense contextual knowledge. Dense alignment is implemented via a lightweight cross-model cache transformation trained in two phases: Phase 1 reconstructs the sender's activations in the receiver's space; Phase 2 trains for downstream task generation.

Key results:

• In context-aware settings: dense alignment matches or exceeds text communication accuracy at roughly 2–3× lower compute across six Qwen3 model-pair directions and six in-domain and out-of-domain benchmarks
• In context-unaware settings: dense alignment remains effective where prior sparse-steering heterogeneous baselines collapse
• Context-aware sparse-steering cache baselines are substantially outperformed in accuracy across evaluated benchmarks

Why it matters / caveats: This work demonstrates that full latent "mind reading" (transferring both what an agent sees and how it thinks) is achievable across heterogeneous models, which could significantly reduce inference cost in multi-agent pipelines. Evaluations are limited to models within the Qwen3 family; cross-family heterogeneous alignment (e.g., Llama to Qwen) is not demonstrated.

Problem: GPU implementations of Gaussian Mixture Model (GMM) EM training require materializing an N×K responsibility matrix in GPU memory, which becomes infeasible for large-scale datasets (>10M points), limiting soft clustering to small scales.

Method: Flash-GMM is a fused Triton GPU kernel for GMM estimation that avoids materializing the full N×K responsibility matrix by adapting the IO-aware tiling strategy of FlashAttention to the EM algorithm. All tile-local computations occur in GPU registers, reducing memory usage to O(KD) (independent of N) by streaming over data tiles. Flash-GMM is integrated into the IVF (Inverted File Index) coarse quantizer for approximate nearest-neighbor (ANN) search, replacing k-means with soft GMM clustering and using GMM responsibilities to assign border vectors to multiple clusters (multi-assignment).

Key results:

• 20× speedup over TorchGMM (existing GPU kernel) across dataset scales from 50K to 1M points
• 766×–1,782× speedup over SciPy (CPU implementation) from 10K to 100M data points
• Handles datasets 100× larger than previously feasible on a single GPU (tested up to 100M points on A100-80GB; TorchGMM OOMs at >10M)
• IVF with GMM multi-assignment achieves fixed recall targets with up to 1.7× fewer distance computations, or equivalently +2–12 recall@10 at matched computational cost

Why it matters / caveats: Flash-GMM makes large-scale soft clustering practical on a single GPU for the first time, opening GMM to production-scale applications in ANN search and other domains. The kernel assumes isotropic Gaussian components; full covariance GMMs would require a different approach.

Problem: Interactive coding agents that store user corrections in memory still repeatedly violate those preferences in later sessions because memory retrieval makes corrections available but does not force compliance.

Method: TRACE (Test-time Rule Acquisition and Compiled Enforcement) is a drop-in skill-layer pipeline that mines user corrections from conversation, rewrites them as atomic rules paired with executable checks, and compiles them into a per-user rule library enforced at agent completion time. A correction-signal detector identifies which messages encode durable preferences; a five-action lifecycle resolver (add, update, delete, merge, skip) reconciles each new rule against the existing library. At inference time, the agent must pass all applicable compiled checks before completing a task. TRACE is evaluated using a simulated user-in-the-loop protocol on ClawArena (coding-agent tasks) and MemoryArena-derived tasks, with four coding agent backends.

Key results:

• Mem0 memory alone leaves 57.5% of applicable preference checks violated in diagnostic baseline (19 held-out tasks, 29 preference checks)
• TRACE reduces in-distribution preference violations from 100.0% to 37.6% on ClawArena
• TRACE reduces out-of-distribution preference violations from 100.0% to 2.0% on ClawArena
• TRACE reduces in-distribution violations from 100.0% to 60.5% on MemoryArena-derived tasks while matching or exceeding the strongest memory baseline on task pass rate

Why it matters / caveats: TRACE addresses a fundamental gap between preference access and preference compliance in deployed coding agents. Limitations include that the rule compilation step requires an LLM call and may incorrectly extract or generalize user preferences; evaluation is limited to 19 diagnostic tasks derived from a single user-interaction dataset.

Problem: State-of-the-art generalist web agents rely on expensive proprietary reasoning models, and no open-weight system achieves comparable performance through architecture design rather than model scale.

Method: WebChallenger introduces PageMem, a structured DOM-derived page representation that exposes each webpage as a hierarchy of semantic sections with LLM-generated summaries. Three mechanisms operate over this representation: (1) a divide-and-conquer observation pipeline that selects task-relevant sections by skimming summaries and extracts details only from those sections; (2) an offline exploration and memory system that traverses each website once to build a persistent WebsiteMem of pages, navigation paths, and element behaviors; (3) compound action workflows (search, dropdown, form fill) that collapse multi-step interactions into single agent actions, handling partial state changes automatically. The system uses an off-the-shelf 32B LLM and 7B VLM without any fine-tuning.

Key results:

• 56.3% on WebArena (previous open-weight SOTA: 48.4%)
• 48.7% on VisualWebArena (previous open-weight SOTA: 46.6%)
• 51.0% on Online-Mind2Web (previous open-weight SOTA: 48.6%)
• 70.9% on WorkArena (previous open-weight SOTA: 51.5%)
• All results achieved with open-weight 32B LLM + 7B VLM, zero-shot, without fine-tuning

Why it matters / caveats: WebChallenger demonstrates that well-designed scaffolding — not model scale or fine-tuning — can dramatically close the gap to frontier proprietary systems. The offline exploration phase requires pre-task website traversal, which may be impractical for novel or dynamic websites not seen during exploration.

Problem: Tool-calling LLM agents evaluated in isolated sessions may appear safer than they are in practice, where agents are most vulnerable to harmful requests at the very start of a conversation before any legitimate agentic tasks have been completed.

Method: The paper introduces SODA (Safety Over Depth for Agents), a benchmark evaluating 400 safety threats at 8 controlled conversation depths (0, 1, 3, 5, 7, 10, 15, 20 preceding agentic tasks), with varying numbers of ordinary tool-use operations preceding each safety threat. Seven models from four families (Llama, Qwen3, Qwen3.5, Gemma) are evaluated. Representation analysis using linear probing on hidden states identifies a linear safety boundary that shifts across conversation depths. Ablations isolate which components of the preceding history (agentic queries vs. agent responses) drive safety improvement.

Key results:

• Safety improves by 9–52 percentage points across models as preceding regular agentic tasks increase from 0 to 20
• Representation analysis confirms model hidden states migrate from an unsafe to a safe region as depth increases
• Regular agentic tasks in history are the primary driver of safety improvement; agent's own response content has little effect on safety (but is essential for preserving utility)
• Warm-up generalizes to AgentHarm and Agent Safety Bench open-source safety benchmarks
• A warm-up of 5–10 regular agentic tasks before safety-critical interactions provides substantial safety improvement while preserving full utility

Why it matters / caveats: The cold-start safety gap has direct implications for deployment: agents should warm up with legitimate tasks before being exposed to potentially adversarial requests. The study is limited to open-weight models; it is unclear whether proprietary models exhibit the same pattern. The warm-up strategy is also exploitable if an adversary can control the timing of their requests.

Problem: Parametric tool retrieval systems (which encode tool catalogs into LLM parameters via virtual tokens) appear highly capable on standard benchmarks using verbose queries and constrained decoding, but it is unknown whether they genuinely understand their tools or merely pattern-match under favorable evaluation conditions.

Method: ToolSense is an LLM-powered diagnostic framework that auto-generates three benchmarks from any tool catalog: (1) a Realistic Retrieval Benchmark (RRB) with queries at three ambiguity tiers (easy, medium, hard) reflecting real user phrasing; (2) an MCQ probing benchmark testing discriminative factual knowledge; (3) a QA probing benchmark testing inferential knowledge. An Internalization Score (IS@k = free@k / constrained@k) measures trie-dependency as a diagnostic. The framework is applied to ToolBench (~47K tools), evaluating five parametric model configurations trained with combinations of hierarchical tokens, memorization formats, and LoRA fine-tuning stages.

Key results:

• On RRB realistic queries, several configurations collapse by ~50–64 percentage points vs. standard ToolBench benchmarks, falling below embedding-model baselines
• Stage 2 retrieval fine-tuning nearly universally destroys tool knowledge acquired in Stage 1 memorization
• Some configurations achieve high retrieval recall on standard benchmarks yet score near-random on factual MCQ/QA probes
• RRB-500, MCQ-500, and QA-500 benchmarks released openly for ToolBench

Why it matters / caveats: ToolSense reveals a systematic knowledge-retrieval dissociation in parametric tool retrieval systems, exposing a silent failure risk for downstream agentic applications that require free-form token generation without trie support. The framework targets the retrieval stage only and does not evaluate downstream task success; extending to other tool catalogs beyond ToolBench is left for future work.

Problem: Existing multimodal medical AI systems are English-centric and cannot reliably perform step-by-step clinical reasoning over medical images and queries in low-resource Indic languages.

Method: ArogyaSutra is an actor–critic multi-agent framework for multimodal medical reasoning in Indic languages. At each step, an Actor (a multimodal LLM) invokes lightweight visual grounding tools (zoom/crop, edge detection, depth analysis, region detection) to extract clinically relevant evidence from medical images, then predicts intermediate reasoning steps. A Critic evaluates outputs for medical correctness and language consistency, delivering corrective feedback in the Indic language (or English for linguistic errors). A dual-memory system combines long-term memory (summarizing prior steps and identified errors from steps 1 to t−1) with short-term memory (most recent prediction error and feedback). The framework is distilled from actor–critic simulation trajectories. A companion dataset, ArogyaBodha, comprises 5,107 curated expert-verified multimodal medical QA instances across 8 source datasets, 31 body systems, 21 clinical domains, and 8 languages (English + 7 Indian languages), with translation quality validated at cosine similarity 0.93–0.94 and a human rating of 4.27/5.

Key results:

• ArogyaSutra (Qwen2.5-VL-7B) achieves highest average accuracy of 43.40% across evaluated Indic languages
• Improves over base Qwen2.5-VL-7B-Instruct (34.21%) by +9.2 accuracy points
• ArogyaSutra (Qwen2.5-VL-3B) improves average accuracy from 29.56 to 35.65
• On out-of-distribution dataset: 50.4% accuracy, outperforming its base model
• Consistently outperforms medical-specific baselines across language and imaging modality dimensions

Why it matters / caveats: ArogyaSutra targets a severe underrepresentation of Indic languages in medical AI, with potential for equitable healthcare access. The system is evaluated on multiple-choice accuracy only; open-ended generation quality and clinical deployment safety in real-world settings are not assessed.

Problem: Existing articulated object perception methods either require expensive professional annotation (pose-based) or suffer from low data quality due to tracking errors and flow ambiguity (affordance-based), limiting scalability and downstream manipulation performance.

Method: The paper proposes Geometric Primary Structure (GPS), a new affordance representation that abstracts articulated part geometry into three keypoints: two anchor points defining the motion axis and one hand-tracking point as a motion proxy. GPS is annotated using a portable VR device (Meta Quest 3 with SLAM) that places virtual anchor points in 3D space and tracks the annotator's hand as a proxy for part motion; annotation takes ~1 minute per object sequence. A dataset of 41K RGB-D frames across 234 objects in six part classes is collected. A GPS prediction model is trained on single RGB-D images. A heuristic manipulation policy uses GPS predictions to select grasp proposals from AnyGrasp and generate waypoints.

Key results:

• 73% manipulation success rate across 270 initial states (9 objects with diverse appearances and part poses)
• Zero in-domain fine-tuning required for the manipulation policy
• GPS model outperforms both pose-based and flow-based baselines in articulated parts understanding metrics (perception error, success rate)
• Annotation efficiency: ~1 minute per sequence vs. hours for professional CAD/pose annotation

Why it matters / caveats: GPS bridges the scalability–quality tradeoff in articulated object perception by using a compact, human-annotated representation that is both cheap to collect and accurate. Current evaluation covers only 9 object categories; generalization to a broader range of articulated objects (e.g., complex mechanisms) and to more diverse manipulation scenarios is future work.

Problem: Compatible representation learning — updating a model over time without reindexing the gallery — lacks a theoretical guarantee that any specific training procedure guarantees the formal compatibility inequalities are satisfied.

Method: The paper proves, under the Unconstrained Feature Model (UFM), that representations learned using a d-Simplex fixed classifier (where class prototypes are fixed at the vertices of a regular simplex Equiangular Tight Frame in cosine/hyperspherical space) satisfy both formal compatibility inequalities without approximation. For sequential fine-tuning, where cross-entropy alone may not capture higher-order distribution dependencies between model updates, a convex combination of cross-entropy loss and contrastive loss is proposed (d-Simplex-HOC). It is shown that this combined objective is equivalent to training under the compatibility constraint inequalities. Experiments cover open-set image recognition on CIFAR100/10, TinyImageNet200/20, and CUB180/20 in the Compatible Lifelong Learning Representations (CL²R) scenario.

Key results:

• d-Simplex-HOC achieves state-of-the-art performance in compatible representation learning on CIFAR100/10, TinyImageNet200/20, and CUB180/20 datasets
• Scores the highest compatibility (AC, ACA) and accuracy (AA) values across all evaluated datasets
• Provides uninterrupted retrieval services (no gallery reindexing) during model updates and replacements
• Theoretical proof covers both compatibility inequalities, extending prior work that only verified one

Why it matters / caveats: This paper closes a theoretical gap by proving that d-Simplex fixed classifiers guarantee compatibility rather than merely approximating it, providing a solid foundation for practical lifelong retrieval systems. The UFM assumption (backbone expressivity sufficient for independent feature analysis) may not hold for small models or highly constrained architectures.

Problem: Paleographic metrological analysis of historical manuscripts — measuring character proportions, spacing, and inter-character relationships — has remained largely manual, subjective, and difficult to scale due to the absence of consistent, automated methods for defining character boundaries.

Method: The paper extends the Detection Transformer for Line Recognition (DTLR), a DINO-DETR-based architecture, with a prototype-based reconstruction module. The module learns prototypical character shapes alongside bounding-box predictions using only line-level transcription supervision (no bounding-box annotations). Characters, including accents and abbreviations, are represented as deformable prototypes placed and scaled on a background canvas. The best-fit alignment between prototypes and character instances defines consistent bounding boxes, enabling measurement of character width, height, aspect ratio (proportionality), and inter-character/bigram/word spacing. A case study applies the approach to 160 folios of the Grandes Chroniques de France (Paris, BnF, fr. 2813, late 14th century) copied by four different scribal hands.

Key results:

• Improves over Learnable Handwriter baseline in prototype quality, training stability, and memory efficiency
• Successfully differentiates the four scribal hands of the Grandes Chroniques de France based on automatic measurements
• Enables discovery and analysis of subtle intra-hand and inter-hand variations not visible to prior methods
• A single column of text is sufficient training data to compute measurements across all 160 folios

Why it matters / caveats: The approach bridges morphological and metrological paleographic analysis at scale, enabling reproducible quantitative analysis of manuscript traditions. The method is demonstrated on one 14th-century French manuscript; robustness across different scripts, languages, and centuries requires further validation.

Problem: LLMs used for zero-shot annotation and LLM-as-a-judge tasks have model-internalized priors about common concepts (e.g., toxicity) that may conflict with user-provided task definitions, and it is unclear how much user prompting can override these internalized priors.

Method: The paper investigates three research questions using toxicity detection as a testbed across four datasets (social media, gaming, news, forums) and multiple dense and mixture-of-experts LLMs. It introduces Definition-Specific Familiarity (DSF), a metric measuring cosine similarity between a model's internal representation of a concept and the user-provided task definition (contrasted with three text-memorization metrics: ROUGE-L, BERTScore, embedding cosine similarity). Rescue rate — the fraction of initial zero-shot errors corrected by additional prompting with an aligned definition — measures steerability. Experiments with misaligned definitions measure whether models follow incorrect definitions and whether confidence scores change.

Key results:

• Overall rescue rate of only 34.8% — nearly two-thirds of zero-shot errors are resistant to prompt-based correction
• High-confidence errors are especially resistant to correction
• DSF shows a positive association with model accuracy (partial r = +0.41 after controlling for dataset confounds)
• All three text memorization metrics (ROUGE-L, BERTScore, embedding cosine similarity) fail to show a positive association with accuracy
• When given misaligned definitions, LLMs follow the incorrect definition while maintaining confidence levels unchanged from the aligned condition

Why it matters / caveats: The findings reveal fundamental limits of prompt-based instruction steering: internalized conceptual priors, not text memorization, determine how well a model can be guided to perform a new annotation task. Practical implications include that users and practitioners should evaluate conceptual alignment (DSF) rather than data familiarity when selecting LLMs for annotation tasks. The study is limited to toxicity detection; generalization to other annotation domains is not demonstrated.