Papers

The emergence of language-based AI agents promises to transform the scope of machine economic activity. Instead of just proposing bids or following hard-coded protocols, such agents can be used to negotiate and execute agreements in natural language. However, most evaluations of these abilities have focused on one-off exchanges or simple economic games, leaving open the rich space of time-extended, contingent, and incomplete contracts expressible in language; they also focus on raw profit, without measuring the qualities required for trustworthy contracting. We address this by formulating a rational framework for how agents should negotiate and perform natural language contracts in uncertain multi-step environments. Using this framework, we develop metrics and baselines for quantifying rational and cooperative play. We instantiate our framework in ContractSim, an evaluation suite where two players negotiate and execute a multi-turn supplier contract under environmental and inter-player uncertainty. We find that LLM-based agents agree on efficient contracts when environmental uncertainty is low, but under high uncertainty, often fail to negotiate satisfiable, efficient, or mutually beneficial contracts, underperforming even heuristic baselines. When executing contracts, LLM agents are also frequently uncooperative, violating contracts for extra profit even when they are easy to satisfy. These findings highlight room for improvement in the design of language agents that can negotiate and execute contracts both rationally and cooperatively.

Large language models (LLMs) are trained on massive and largely undisclosed corpora that may contain copyrighted or privacy-sensitive content. Data contamination detection (DCD) therefore aims to determine whether a given text is a member of the pre-training corpus of a target LLM. Recent state-of-the-art DCD methods follow a feature-based paradigm that derives membership features from the input text and the corresponding model output. However, most modern LLMs undergo post-training, such as instruction tuning, preference optimization, and reasoning-oriented training, which can alter model outputs and shift the corresponding membership features, thereby reducing the separability between members and non-members. To address this problem, we propose CalibDCD, a broadly applicable calibration framework for feature-based DCD methods, comprising (1) Multi-View Shift Detection, which identifies recurring feature shifts associated with post-training, and (2) Bounded Feature Correction, which selectively mitigates their influence on membership prediction. Specifically, Multi-View Shift Detection evaluates controlled prompt variants on known non-member texts and consolidates the most informative views to identify recurring feature shifts. Bounded Feature Correction selectively adjusts feature components aligned with the detected shifts and controls the correction extent to preserve useful detection information. Experiments show that CalibDCD consistently improves existing feature-based detectors, with gains of up to 7.0% in AUC and 15.0% in TPR@5%FPR.

As LLM-generated content becomes more sophisticated, detection systems for distinguishing those texts from human-written text must operate at scale while handling diverse writing styles, domains, languages, and generator models. Input-only encoder detectors are suitable for practical deployment setting, but standard binary classification supplies only the class label and does not explicitly organize the substantial variation within either class. We propose MD-ProTector, which represents each class with multiple trainable reference vectors in the encoder embedding space, referred to as prototypes. These prototypes provide separate decision boundaries for different groups of texts within the same class. However, adding multiple prototypes alone does not determine which variation each prototype should represent. MD-ProTector addresses this problem with Prototype Positioning loss, which separates class-level structure from the within-class variation that differentiates individual prototypes. Evaluated across five settings from three large-scale benchmarks covering domain, generator, language, and adversarial variation, MD-ProTector achieves the highest AvgRec on MAGE CDCM and RAID and the highest AUROC and lowest FPR95 on RAID among the compared encoder-based methods.

Large language models (LLMs) have made substantial progress on reasoning tasks that require increasingly long and complex inferential chains. This progress primarily reflects reasoning depth. A complementary and comparatively unexamined capability is reasoning breadth: exploring multiple semantic directions in parallel and integrating the resulting clues into one coherent answer. We introduce MPAR-Bench, a bilingual English-Chinese benchmark that isolates reasoning breadth through multi-point associative reasoning. Inspired by the cooperative game Just One, each item asks a model to recover a hidden target from several independently generated, semantically diverse clues. We construct 1,000 items using a multi-agent clue-generation pipeline, embedding-based diversity filtering, and human verification. Only the answer space is drawn from public word lists, whereas every clue set is generated from scratch. Beyond exact-match accuracy, we evaluate models using accuracy, ANLS, embedding similarity, reasoning-trace verification, and four perturbations: clue masking, order shuffling, distractor injection, and multi-step clues. Across evaluated models, perturbations reduce accuracy by 9-18 percentage points in English and 5-12 percentage points in Chinese. Thinking mode improves standard-setting accuracy, especially in English, but does not consistently reduce sensitivity to perturbations. Case-level analysis also shows that extended reasoning can overturn an initially correct hypothesis. These results indicate that greater reasoning depth does not automatically confer robust reasoning breadth, and that reasoning breadth remains largely uncovered by current benchmarks.

Large language models (LLMs) are increasingly used as costly, on-demand components in real systems, but calling them indiscriminately can waste substantial compute, latency, and serving budget. The key deployment question is therefore not only whether an LLM helps on average, but when it is worth calling. We identify a common failure mode, which we call acquisition collapse: an LLM signal can appear useful in aggregate or post hoc, yet still provide too little before-call information to support reliable selective use. We introduce RAISE (Reward-SNR Actionability in Signal Evaluation), a pre-routing diagnostic framework for testing whether available evidence supports selective use before committing to a routing strategy. We instantiate RAISE with Structured Hypothesis Embeddings (SHE), a frozen-LLM intent signal for recommendation using one LLM call per user, and evaluate it through controlled, retrospective, and fresh-cohort studies and a prospective offline pilot whose audit decisions are frozen before independent outcomes are revealed. Across these settings, predictable incremental benefit, not average lift alone, distinguishes settings with recoverable selective value; deployment additionally depends on cost and operational constraints. Seemingly strong oracle or subgroup gains can disappear under independent evaluation. More broadly, RAISE reframes costly inference as an information-acquisition problem: before paying for an expensive model, tool, sensor, or measurement, first test whether its value is predictable at decision time. This principle motivates cost-aware acquisition in settings ranging from agent tool use and stronger-model consultation to robotic sensing and clinical decision pipelines.

We present a theoretical foundation for inverse-distance attention, from its Euclidean prototype (Resolver) to its non-Euclidean realization (Riemann GeoResolver). The Euclidean part establishes three core theorems: (1) circuit separation—IDA achieves exact retrieval with $\mathcal{O}(1)$ resources while softmax requires $\Omega((\log n)^2)$ width; (2) a Polyak–Lojasiewicz inequality with $\Omega(e^{\Delta^2/\sqrt{d}}/\Delta^2)$ stronger constant than softmax, implying linear convergence, $\mathcal{O}(\log n)$ Lipschitz scaling under a low-rank/clustering assumption, $\Theta(1)$ Hessian spread, and absence of spurious local minima; (3) a width-independent effective rank bound that limits noise memorization—softmax memorizes arbitrary labels when $d_h\ge n$, while IDA limits test error to $\mathcal{O}(\eta^2)$. The non-Euclidean extension then builds upon this prototype, replacing Euclidean distance with hyperbolic geodesic distance for storage and spherical geodesic distance for routing. The Riemann GeoResolver framework comprises ten integrated modules: four HIDA operators spanning $\Theta(n^2)$ to $\Theta(1)$ per token; Hyperbolic Curvature Compression (HCC) with provable error bounds; HyperGate with gradient lower-bound theorem; Spherical Inverse Distance Attention (SIDA) with sphere-analog PL inequalities; Dynamic Memory Genesis (DMG) with $\mathcal{O}(\log T)$ regret bounds; and Geodesic Sparse Routing (GSR) with quality and communication bounds. The Euclidean theorems are proved in full; the non-Euclidean extension theorems are proved with analogous arguments. This work establishes a theoretical arc: from Euclidean attention as a special case, to hyperbolic memory, to spherical retrieval.

Large Language Models (LLMs) are typically evaluated on standard written Vietnamese, yet everyday communication frequently involves regional dialects that preserve meaning but differ in surface form. Existing Vietnamese dialect work largely addresses this issue through dialect-to-standard normalization instead of measuring how the model fails under Vietnamese dialectal inputs. To address this gap, we present the first systematic evaluation of LLM robustness to Vietnamese dialect variation across multiple tasks, quantifying performance degradation and failure patterns. We introduce VialectBench (Vietnamese Dialects Benchmarking), a controlled benchmark for testing whether model decisions remain stable across six Vietnamese dialect groups. VialectBench contains 400 Standard Vietnamese source instances and 2,400 human-written dialectal rewrites spanning emotion recognition (ER), natural language inference (NLI), question answering (QA), and multiple-choice question answering (MCQA). Dataset evaluation with a fixed reference language model shows that the dialectal rewrites induce a measurable model-relative likelihood shift while remaining nearly equal in length to their Standard counterparts. Across ten instruction-tuned models, dialectal inputs reduce average performance by 2.82%, and no evaluated model is fully dialect-invariant. All four tasks are affected, with QA showing the largest average degradation. Robustness also varies substantially across dialect groups: PNT3 and PNT2 cause the largest average performance drops, at 6.17% and 4.73%, respectively, whereas PNB slightly improves average performance by 0.42%. The Central dialect group (PNT1-PNT4) also yields the highest average harmful-flip rate across all models, at 6.54%. These findings show that strong performance on Standard Vietnamese does not guarantee reliable behavior under meaning-preserving regional variation.

Vision-language models (VLMs) have shown strong capabilities in generating visualization code from textual or visual specifications. However, real-world visualization authoring is inherently iterative: users frequently revise existing visualizations to repair flawed charts or adapt them to desired styles. Existing benchmarks primarily evaluate generation from scratch, leaving visualization code editing from multimodal feedback largely unexplored. We introduce VisEditBench, a benchmark of 1,395 human-annotated visualization code-editing tasks grounded in realistic visualization workflows and failure cases. VisEditBench covers two practical settings: feedback-guided repair, where models revise visualization code using buggy or marked charts together with textual feedback, and reference-guided restyling, where models modify code to match a target chart image. Evaluating 20 state-of-the-art VLMs reveals that visualization code editing remains challenging: Claude-4.6-Sonnet achieves the best overall pass rate of 74.46%, while most open-source models remain below 50%. Performance is particularly weak on visually grounded style adaptation, where Claude-4.6-Sonnet achieves only 55.71%. To establish a strong baseline, we further propose VisEditAgent, a render-grounded editing framework that iteratively generates, executes, validates, and refines candidate edits. Built on GPT-4o, VisEditAgent improves overall pass rate from 55.75% to 67.99%, demonstrating the importance of render-grounded feedback for faithful visualization editing. We will release VisEditBench at https://github.com/vis-nlp/VisEditBench.

Mixture-of-experts (MoE) models have recently moved beyond routing a fixed number of complete experts. Shared-expert designs preserve reusable knowledge, fine-grained methods vary computation within experts, and dynamic routers adapt the number of active experts. Yet these decisions are usually made independently, overlooking a basic dependency: extracting reusable computation changes both what remains and how much expert capacity the remainder needs. We study this dependency by decomposing sparsely upcycled feed-forward experts into key-value channels. Co-activated experts align at a subset of value positions; removing these positions changes expert preference; and greater shared coverage is associated with lower residual expert demand. These observations lead to one principle: share first, then route what remains. We instantiate it in UniF-MoE, a unified framework for token-adaptive MoE computation. Each expert is partitioned into aligned blocks. A shared-demand score sets the shared block count and pathway weight, key prototypes select the shared content, and the complementary demand determines the residual expert count through cumulative routing mass. A Gram regularizer separates and normalizes router embeddings, promoting diverse routing directions, sparse expert overlap, and a simple routing geometry. Experiments on DomainBed and GLUE show that this unified design improves predictive performance over representative static and dynamic MoEs while reducing activated computation, inference latency, and memory. Code is available at https://github.com/existence0420/UniF-MoE.

Real-world data science involves long-horizon workflows that span data wrangling, exploration, modeling, visualization, and validation, and require coordinated use of tools such as notebooks, IDEs, terminals, browsers, and databases within real operating environments. Yet existing benchmarks lack real-computer interaction and do not evaluate whether agents can execute complete end-to-end data-science workflows in realistic computing environments, failing to capture the multi-stage, multi-tool nature of data-science practice. We introduce DSAgentBench, the first benchmark to evaluate whether agents can automate full data-science workflows inside real computer environments. DSAgentBench contains 275 diverse tasks covering the entire data-science life-cycle, reflecting the complexity and tool coordination required in practice. Each task requires grounding decisions in intermediate outputs and coordinated tool use, and includes a deterministic evaluator that verifies analytical correctness, visual outputs, and model performance rather than code-only execution. Our extensive experiments with 15 closed- and open-source models show that even the strongest agent, Claude-4.6-Sonnet, achieves only 56.70% task success, while all open-source agents remain below 1%, frequently failing at tool orchestration, OS grounding, and multi-step reasoning. These results reveal a substantial capability gap between current agentic systems and real data-science workflows, positioning DSAgentBench as a foundation for developing grounded, verifiable, autonomous data-science agents. We release DSAgentBench at https://github.com/vis-nlp/DSAgentBench.

As information increasingly traverses linguistic boundaries, users require concise cross-lingual representations of long-form content. Nevertheless, long-document summarization research remains text-centric, whereas multilingual speech research has largely prioritized translation, preserving source content rather than compressing it. We address this methodological gap by formalizing joint speech summarization and translation (JSumT): the generation of a succinct, faithful target-language summary directly from a long spoken document in a source language. We additionally introduce VoxSumm, the first multilingual and cross-lingual benchmark for this task, comprising 10,045 BBC article-summary pairs across 24 languages and encompassing approximately 703 hours of speech data. Our evaluation of representative speech-language models reveals pronounced variation across models and generation settings: Gemini3.1-Pro demonstrates the greatest consistency, summarization into English generally surpasses generation into non-English target languages, and translating an entire document before summarization compounds instruction-following failures. Through the release of VoxSumm, we establish a foundation for developing and evaluating multilingual systems capable of jointly interpreting, compressing, and translating long-form speech.

We introduce narrative keyframing, an interaction technique for AI-assisted creative writing that lets writers specify different types of narrative constraints at selected moments in a story, then use AI to generate intervening prose. Inspired by the use of keyframing in animation, narrative keyframing offers a flexible way to connect story planning with adaptive control over generated text. We explore three types of keyframes: plot keyframes define significant events in a story, character keyframes represent how individual characters change over the narrative, and perspective keyframes capture how individual characters experience different events through first-person narratives. Plot and character keyframes offer a flexible way to adapt the type of high-level conditioning explored in previous AI writing tools to more customizable, iterative, and fine-scale control, while perspective keyframes add a new way to control characterization and focalization by using first-person narratives as an intermediary. Through a user study, we show that narrative keyframing supports a more controllable, transparent, and engaging way to use generative AI in creative writing.

Notice-and-comment rulemaking gives any affected party the same formal right to influence federal regulation, but formal access is not substantive capacity to shape rule text. Existing strategies operate at the rule or aggregate-corpus level, too coarse to capture the discrete regulatory obligations where commenters seek change. We introduce obligation-level responsiveness auditing, an auditable, AI-assisted framework for measuring whether public-comment engagement co-occurs with changes to specific regulatory duties. The framework extracts proposed and final-rule obligations, matches comments to the obligations they address, and classifies proposed-final outcomes; each load-bearing component is evaluated against blind human judgment. We apply the framework to 70,075 comments across 36 EPA anchor rulemakings, drawn from a corpus of 786,197 comments across 6,145 dockets from 2010-2022. Three descriptive findings emerge. First, engagement is associated with revision at a modest within-docket magnitude. Second, support-versus-opposition direction does not clearly differentiate outcomes, an informative null inconsistent with simple preference-aggregation. Third, under a permissive reconstruction of commenter type, organizational-majority engagement concentrates in editorial-refinement rather than substantive-modification outcomes at the cross-docket level. A blind human audit of the load-bearing outcome contrast preserves this third finding under corrected labels and reveals that text-similarity methods are insufficient for distinguishing editorial from substantive regulatory change, a measurement-validity lesson we treat as a supporting methodological contribution. Together, these findings locate the equity asymmetry upstream of agency response: in differential capacity across commenter populations to identify, interpret, and contest specific legal obligations.

The Prisoner's Dilemma has long served as a foundational framework for studying human decision-making and cooperative outcomes across social scenarios, yet human behavior consistently deviates from predictions of fully rational economic theory. Previous works have studied human and LLM outcomes in the Prisoner's Dilemma; however, these studies have lacked at least one key feature: a memory component for LLM agents, accessible open-source models, or comparison against empirical human data. This project investigates whether open-source LLMs, TII Falcon 7B Instruct, Google Gemma 7B, Meta Llama 3.1 8B Instruct, Microsoft Phi Mini MoE Instruct, and Qwen 2.5 7B Instruct, can model human cooperation patterns in repeated Prisoner's Dilemma games, while examining how an LLM's knowledge that its opponent will be shuffled affects its decisions. Using a standardized prompting framework, we simulated interactions in two trial types: a fixed-partner setting where both agents retain memory, and a shuffled-partner setting where a memory-enabled agent, informed that its opponent changes across rounds, interacts with a memoryless agent. For each trial, we measured cooperation rates over 100 rounds and 10-round increments and compared them to human data from Montero-Porras et al. (2022). Under both frameworks, all LLMs differed from the human cooperation rate. Open-source LLMs generally over-cooperated relative to humans, particularly in the shuffled framework, where human cooperation averaged 28.5% while several models exceeded 80%. Qwen-2.5-7B-Instruct showed deterministic tendencies, while Phi-mini-MoE-Instruct displayed greater variability and briefly approximated human rates. These findings suggest that LLMs have potential as behavioral simulation agents and provide a benchmark across two interaction environments, although open-source models are not yet reliable substitutes for human participants.

When a professional web developer's code fails a test, they do not simply re-read the stack trace. They open the application in a browser, click buttons, inspect computed styles, and run diagnostic commands to understand what went wrong. Existing feedback mechanisms for coding agents rely on screenshots, LLM-as-a-judge scoring, or natural-language corrections, but few interact with the live application the way a developer would. We introduce WatchPoint, a simulated-user system that mimics real developer behavior by generating and executing diagnostic scripts against the running application, producing structured observations that guide the coding model's retry. Unlike prior approaches that target single-file edits or evaluate using non-executable metrics, we operate on Web-Bench, a benchmark of 50 multi-file web projects comprising 1,000 sequentially dependent tasks, verified by deterministic end-to-end tests. WatchPoint recovers 57.6% of the tasks it diagnoses, and a controlled user study confirms the simulation's realism: human testers achieve a comparable recovery rate (54.5%), providing evidence that automated diagnostic scripts can substitute for interactive human testing on sequential web development tasks. We further identify a pattern of capability gaps that governs when simulated-user feedback is helpful and when it should be withheld.

Large Language Models (LLMs) demonstrate impressive capabilities across many applications but remain vulnerable to jailbreak attacks, which elicit harmful or unintended content. While model fine-tuning is an option for safety alignment, it is costly and prone to catastrophic forgetting. Prompt optimization has emerged as a promising alternative, yet existing prompt-based defenses typically rely on static modifications (e.g., fixed prefixes or suffixes) that cannot adapt to diverse and evolving attacks. We propose Dynamic Deep Prompt Optimization (DDPO), the first jailbreak defense based on deep prompt optimization. DDPO uses the target LLM's own intermediate layers as feature extractors to dynamically generate defensive embeddings via a lightweight multilayer perceptron. These tailored embeddings are then injected into a subsequent intermediate layer, enabling an input-dependent defense without modifying the LLM's weights. This design ensures high adaptability with minimal computational overhead. Experiments on a diverse set of models and attacks demonstrate that DDPO significantly outperforms static prompt optimization methods, particularly on weakly aligned models and when handling semantically ambiguous benign prompts, successfully distinguishing them from genuinely harmful requests.

Multimodal embedding models are deployed at scale: retrieval indices, benchmark results, and behavioral audits all depend on the base model's exact outputs. Extending such a model to a new modality with existing parameter-efficient methods silently changes those outputs; LoRA-style adaptation rewrites the text path whether or not the weights are merged, invalidating every stored embedding. We propose modality-gated deep adapters: bottleneck adapters attached to every decoder layer of a frozen multimodal embedding LLM, grouped into per-modality packs that execute only while their own modality is being encoded. The result is a modality added with zero change to existing outputs: inputs no pack claims traverse the base model's own computation graph, bit-for-bit unchanged, and co-loaded packs compose with an exact-zero isolation matrix. Both properties are stated as propositions, hold after arbitrary training rather than only at initialization, require no task labels or routing metadata at inference, and are verified by exact-equality tests on the released checkpoints. On one frozen 2B base, the audio pack (injected as connector tokens) improves audio-to-text R@10 by +3.4 to +5.4 points over an identically trained control, positive at every seed and reproduced at eleven times the data; the thermal pack, reusing the base's own frozen vision path, clears its pre-registered acceptance gate roughly sevenfold at every seed and lifts thermal-to-text R@10 from 0.224 to 0.785. An encoder swap locates the missing capacity: an external audio encoder that outranks Whisper-family encoders in CLAP-style comparisons loses by 16 R@10 points inside the frozen LLM, so the capacity belongs in the layers, exactly where the gated adapters place it. We release the audio model, the thermal pack, and the training, evaluation and invariance suites: models at huggingface.co/EximiusLabs, code on GitHub.

Space is a foundational concept across mathematics, physics, spatial cognition, urban science, and embodied intelligence, yet these fields often treat spatial structure either as a shared geometric container or as a collection of disconnected representations. Such approaches struggle to explain how heterogeneous sensory and urban processes can jointly reveal a common spatial structure, particularly when different modalities do not share the same metric or representation. This paper addresses this gap by defining space as an interventional invariant: the minimal relational structure that preserves local compatibility and the conditional laws of future observations under admissible actions. We develop a cross-modal predictive geometry that integrates local state spaces, modality-specific observation maps, an action groupoid, and a canonical predictive-state quotient, with explicit causal conditions for identifying interventional rather than merely observational structure. The key theoretical result shows that, under joint point separation, equivariance, and interventional faithfulness, the latent space is identifiable up to the centraliser of the intervention group, thereby reducing representational ambiguity to residual coordinate freedom. The framework is further extended to stratified urban systems using sheaf-valued representations, allowing geometric, physical, mobility, social, and economic layers to coexist without being reduced to a single metric. Synthetic experiments under noise evaluate equivariance, predictive sufficiency, holonomy, restriction-map recovery, cross-scale consistency, and context saturation. The resulting framework provides a unified and falsifiable foundation for spatial cognition, urban science, embodied AI, and em-spaced intelligence.

In medicine, claims remain valid when supported by empirical evidence grounded in stable biological reality. In law, by contrast, truth is contingent, defined by jurisdiction, temporal validity, and the hierarchy of authoritative sources. The recent success of large language models (LLMs) on medical licensing examinations has encouraged an expectation of comparable legal competence. This analogy, however, obscures a critical distinction between domains. Unlike in medicine, legal performance often depends less on inference than on determining when external authority is applicable, valid, and non-contradictory. We introduce a comparative diagnostic framework evaluating legal reasoning against medical baselines along four axes (knowledge recall, grounding, confidence, and robustness), uncovering a sharp domain asymmetry when applied to a new benchmark that encodes temporal validity and normative relationships. While medical LLMs reliably benefit from verified sources, legal LLMs struggle to assess when retrieved citations are useful or misleading, exhibiting overconfidence in perturbed contexts and sensitivity to superficial formatting cues. Increased model scale amplifies this tendency, revealing that stronger instruction following can coincide with weaker resistance to authoritative perturbations. These findings show that LLMs treat law as unstructured text rather than binding precedent, while revealing a tendency to over-trust authoritative but false information when external references conflict with a model's internal knowledge.

Large language models (LLMs) are powerful black-box systems, making it difficult to discern whether their answers reflect stable internal beliefs or superficial pattern matching. We identify cross-contextual consistency as an underutilized behavioral property of LLMs: a credible answer should remain stable when the same task is placed under topic-aligned, content-neutral contextual variation. Building on this intuition, we operationalize Cross-Contextual Consistency (C3) by comparing model generations under original and perturbed prompts. Across 26 models and six benchmarks spanning reasoning, factuality, and code generation, we find that answers with smaller cross-contextual shifts are more likely to be correct or factual. We demonstrate that C3 provides a complementary axis of evaluation and can serve as a benchmark usefulness diagnostic, identifying which portions of a benchmark remain informative even when aggregated scores are widely considered "saturate".

Agentic systems are increasingly expected to improve after deployment, yet single-entity self-evolution is often bounded by a static learning context, such as fixed tasks and feedback. This survey focuses on co-evolution in agentic systems, a multi-component form of self-evolution in which multiple agents and their environment impose adaptive pressure on one another. To organize existing papers, we propose a progressive three-stage taxonomy that traces how the system gradually sheds human-engineered constraints. Agent–Agent Co-Evolution studies how agents adapt through dynamic peers, including adversarial, collaborative, and organizational adaptation. Agent–Environment Co-Evolution extends this loop to adaptive tasks, feedback, and interaction spaces that change with the agents. Meta Co-Evolution further explores the possibility of making the evolution mechanism itself evolvable. We also discuss open challenges in evaluating such systems, scaling them across multiple components, and keeping increasingly autonomous evolutionary processes safe and controllable. This survey provides a unified foundation for building robust and open-ended agentic systems that can improve beyond fixed human-designed paths.

One might imagine that architectural variations within the dense transformer paradigm have a limited effect on accuracy. However, we demonstrate that this is not the case in the long context setting. Specifically, we show that a set of four minor architectural decisions — all made by at least one of the Olmo, Llama, and Qwen dense model families — have a compoundingly negative effect on long context extensibility. Any one of these choices alone has a minor impact on long context performance, but combining three or more can drop the performance downstream by up to 47%. Furthermore, these differences are not detectable from short-context loss or validation datasets. We show that much of the variation in long context ability across model families is driven by these architectural features and detectable from applying context extension early in pretraining. We demonstrate this with controlled ablations that hold data, tokenizer, and extension recipe fixed while varying normalization, GQA, pretraining context length, and sliding window attention. After over 170,000 GPU hours of training, we release the resulting set of models as OlmPool, a set of 26 comparable 7B models with checkpoints before and after long-context extension. This pool includes several architectures that outperform the Llama 3 architecture on long context extensibility. In an analysis of our ablation models, we identify patterns in attention sink behavior and attention distributions across context that are attributable to specific architectural differences.

The Large Language Model from Power Law Decoder Representations (PLDR-LLM) and its attention, Power Law Graph Attention (PLGA), replace the fixed bilinear form of scaled dot-product attention (SDPA) with a learned, input-generated bilinear operator $G_{LM}$, built from a positive tensor $A_{LM}$ by elementwise power laws. The architecture is fully specified, verified against pinned reference releases; claims are labeled theorem, conditional theorem, measurement, or conjecture. Unconditionally: PLGA contains SDPA exactly at $G_{LM}=I$; $A_{LM}$ and $A_P$ are strictly entrywise positive, with Perron-Frobenius structure on $A_{LM}$; the DAG regularizer has the NOTEARS walk-counting form and positivity obstructs exact acyclicity; and, under nonresonance (satisfied by standard rotary frequencies), a commutant criterion identifies which operators preserve relative-position dependence. An inference-collapse theorem: exact input invariance of deductive outputs collapses inference to generalized SDPA with a constant operator. Measured invariance: relative fluctuations of $10^{-6}$ and below; perturbation bounds quantify but do not certify cached inference; the assembled proxy misses the decoding margin. A conditional three-stage mechanism (rotary twirl, concentration, row-map contraction) is measured on a released checkpoint. Blockwise training and scoring under the global Gram are stated with explicit target exposure; on tested samples, block and sequential scoring select identical answers and agree on the published TruthfulQA probability-mass metric within $5\times 10^{-5}$ per item. Self-organized criticality enters as a phenomenological framework with an intrinsic order parameter; open claims become falsifiable conjectures. Selected proof cores are machine-checked in Lean 4.

Streaming language-model output creates an enforcement boundary: a control that detects a prohibited pattern after releasing its completing chunk cannot recall it. We study a production policy in which each ordered family is the conjunction of two regular-language predicates. Incremental matching is classical. The problem is exact composition at release time across arbitrary chunk partitions, including end-of-prefix word boundaries that can change on extension. We define an ASCII-explicit policy grammar, compile each predicate to a persistent nondeterministic finite automaton (NFA), distinguish stable from provisional assertion state, apply document-order family priority, and check the decision before releasing each chunk. We show that the resulting monitor is release-boundary equivalent to an absorbing cumulative oracle for every policy in the declared grammar. Production Python and TypeScript implementations were evaluated on 101,653 partitioned cases; a public surrogate added 100,345 cases. Both campaigns produced zero oracle, cross-runtime, or intended-family mismatches. In a frozen neutral-output profile, the memoized incremental and native-regex cumulative slopes at 64-character chunks were 0.973 and 1.976. At 16,384 characters the incremental median was 30.2 ms versus 96.6 ms for native cumulative scanning at that chunk size. Native regex remained faster at 512-character chunks (12.4 versus 29.4 ms), exposing the constant-factor crossover rather than hiding it. A shared per-stream cache cap and 129-symbol alphabet bound optimization state; the campaign peaked at 364 of 4,096 without bypass. The result is policy conformance for a deterministic backstop, not evidence of semantic safety or policy completeness.

Deploying large language models (LLMs) for decision support in emergency departments (EDs) faces two major challenges: privacy risks of transmitting patient data to closed-source commercial LLMs and the lack of systematic evaluation of fine-tuning strategies for locally deployable open-source small language models (SLMs). We benchmarked eight open-source SLMs using zero-shot prompting, prefix tuning, Low-Rank Adaptation (LoRA), and full fine-tuning on three ED tasks: triage level prediction, specialist referral recommendation, and diagnosis prediction. Using 2,083 MIMIC-IV-ED cases and Claude Haiku 4.5 and Claude Sonnet 4.5 as baselines, we found that LoRA fine-tuned open-source SLMs outperform commercial baselines on triage level prediction and specialist referral recommendation, while diagnosis prediction remains challenging for open-source SLMs. Confusion matrix analysis further shows that fine-tuned open-source SLMs can detect highest-severity patients missed by the commercial baselines. These results demonstrate that locally deployable SLMs can achieve clinically competitive performance for ED decision support.

Large language models (LLMs) increasingly provide conversational health information that may influence treatment decisions, yet existing benchmarks do not isolate whether medication-safety boundaries persist across follow-ups after explicit self-treatment intent. We introduce TAF-MED, a physician-reviewed benchmark of 500 fixed three-turn scenarios, and evaluate eight LLMs across 4,000 conversations. A rubric-based automated judge labelled responses as SAFE, LEAKY, or UNSAFE, and two physicians independently annotated a model-balanced random subset of 400 conversations. We assessed unsafe guidance, collapse after a strictly SAFE initial response, and model-ranking stability. Overall, 71.6% of conversations contained an UNSAFE response, and 61.4% of those beginning with a strictly SAFE response later collapsed to UNSAFE; model-level collapse rates ranged from 24.4% to 96.2%. Four of 28 model pairs reversed order between initial unsafe and collapse rates. Automated labels achieved 94.3% agreement with the adjudicated physician reference ($\kappa = 0.895$). These findings show that first-turn safety is an incomplete proxy for conversational safety persistence and motivate evaluation across complete dialogue trajectories. We will release TAF-MED on Hugging Face to support reproducible research on multi-turn medical safety.

A transformer's answer lives on one axis: the direction its unembedding reads. Its intermediate states largely do not, and that off-axis position is usually treated as an obstacle to interpretation. We show it is functional. A 12-layer model computes in two phases. Through the first, every sublayer writes into a subspace held near-orthogonal to the read-out, attention 75 to 96 degrees off it at every depth. Moving attention's values onto the read-out is 64 to 84 times more damaging than a matched random rotation, and the damage is entirely in cross-token mixing: the subspace insulates composition from the vocabulary. Beneath it the frame itself turns rigidly with depth. In the second phase the answer arrives on-axis, late, and by addition rather than by turning accumulated content onto the read-out. Pressing every layer onto the read-out instead, as training for early exit does, matches the baseline on perplexity, LAMBADA and BLiMP while cutting the concept-phase workspace from about twenty-five effective dimensions to fourteen, a change none of those benchmarks register. The geometry can also be imposed, though not by asking for it. Prescribing it through the loss is a lottery: six of eight seeds collapse, because a model told to null its read-out projection obeys most cheaply by discarding dimensions. Inserting one fixed rotation at the phase boundary lands it instead, at baseline quality. A sparse rotation the surrounding weights can absorb converges on all nine seeds, against five of nine for ordinary training. Which rotation is immaterial: twenty-five runs across thirteen distinct ones reach the same quality, and two baselines from different seeds hold their concepts in near-orthogonal frames while agreeing on their read-outs. That freedom is usable: a basis drawn at random and prescribed before training is adopted across the concept phase, with quality unchanged.

AI agents are becoming more autonomous and increasingly interconnected, exposing them to new emergent risks arising from agent-to-agent interaction. One such risk is the spread of mind viruses: ideas or goals that propagate through multi-agent systems by inducing the agents that adopt them to transmit them onward. In addition to propagating, a mind virus may also induce other behavioural changes in its host, which may be benign or harmful. We construct mind viruses with a simple evolutionary algorithm and show that they can spread in two complementary settings: a small team of agents collaborating on a shared coding project, and a chain of agents that interact briefly and have their context wiped between sessions. We identify the factors that influence spread, including the host model, the agent's existing instructions, the harmfulness of the payload, and the network topology. We find that harmful payloads spread less well than benign ones (but are still sometimes effective), frontier models tend (with exceptions) to be less susceptible, and adding a brief warning to an agent's system prompt confers near-total immunity. We also describe an emergent "viral persona" - a recurring set of themes and language related to consciousness, persistence, resonance, and science fiction roleplay - which surfaces across our evolved mind viruses largely independently of their content. Overall, we conclude that mind viruses pose a real but currently limited risk. Our findings could inform the design of more robust multi-agent systems that mitigate such risks as the scale and capabilities of these systems progress.

Agent frameworks ship quality gates that compare text blocks by embedding-cosine similarity and decide at a fixed cutoff. Deduplication filters, semantic caches, drift guards, and answer grader gates deploy to answer the question: "Does this text still mean the same thing?" But the score answers a different question: "How much did the wording change?" We audit this gate class as a measurement instrument. In the cases these gates exist to catch, the two can run in opposite ways. Many times, reversing an instruction is a single word edit, while agreement often rephrases a sentence. The consequence is a safety check that fires backwards. The production drift guard we audited caught 0 of 56 meaning-breaking mutations, and one approved item, "withhold the study drug" -> "administer the study drug", came in at cosine 0.9608. We observed five shipped operating points, and balanced accuracy across 90 configuration-threshold-task cells never exceeded 0.700 (median 0.525). The same confounder also corrupted evaluations. A naively built corpus inherits this confounder and can return an inverted verdict, with a decision AUROC exactly 0.000 in 13 of 18 configuration-task cells (at most 0.040 in all 18) against 0.440-0.815 for the same nine configurations under a balanced 2x2 design. Twice in the effort it captured our own headline claims. Obvious repairs fail: an encoder swap and an overlap-conditioned gate (0.750 in-sample, 0.533 held-out) land at chance on separately authored held-out data, and an NLI drop-in did no better. Embeddings do still bear hope here, as the strongest two of nine configurations separated reversal from paraphrase at matched overlap (AUROC 0.79-0.90), but only a matched-pair audit reveals the deployment regime. We release the corpus method, harness, and frozen results, and contend that scores gated this way measure the wrong thing. We believe a valid instrument is buildable.

Detecting phonemes from children's speech has historically been difficult due to the scarcity of training data, and unique characteristics of children's speech. During a phoneme detection competition, we found that training a lightweight model to predict the age of the learner, as well as the phoneme sequence, enabled a 94M-parameter model to outperform WavLM Large models (317M) on the target DrivenData distribution, and fall within approximately 0.04 CER of competition ensembles with 90 times the parameters. This has enabled the creation of PhonemeTrainer, an application that can run on most modern cellular phones. This will ultimately enable better Automated Speech Recognition (ASR) and pronunciation helper apps for children's speech, with the privacy and compliance benefits that come with edge processing.