Named after the hundred-eyed watchman of Greek myth, Argus watches the education landscape: spotting new opportunities, pressure-testing the ventures we're building, and tracing every read back to the real-world signals behind it.
The evidence library: the raw signals the pipeline is watching across the education ecosystem. Every idea is built from these.
arXiv:2607.22953v1 Announce Type: cross Abstract: Modern AI systems bring societal risks such as mass surveillance, extreme concentrations of power, and loss of user autonomy---calling into question a model where third-parties collect and control massive amounts of user data. Users require a sovereign system to securely own, govern, and disclose their context while remaining compliant across regulated domains with strict provenance, interpretability, and policy adherence. Perspective-aware AI approaches this by transforming a user's aggregated personal data into a structured identity model called a \emph{Chronicle}: a temporal knowledge graph that represents and grows with the user. Chronicles support the secure disclosure of context across federated networks. A Chronicle holder may expose a queryable, authorized view that a third-party agent may consult without centralizing anyone's data. This paper explores the problem of minimum-necessary disclosure across domain boundaries: when a
arXiv:2607.22940v1 Announce Type: cross Abstract: Computer programming MOOCs are instrumental in providing students with high-quality instruction in areas where there is limited access. They are especially beneficial to post-secondary African students as less than 1% of them leave secondary school with fundamental coding skills. One strategy for increasing their efficacy for African students is to understand students' motivation for enrolling. These insights can inform the design of MOOC content and assessments to align with students' interests. We administered an open-ended response survey to (self-identified) Africans enrolled in a smartphone-based online coding course (SuaCode). We analyzed a random sample of 450 (of 3000) responses using a grounded theory approach. We found that most African students (68.7%) participated in SuaCode for intrinsic reasons such as improving themselves, learning with like-minded individuals, and gaining skills to help address societal issues. We discus
arXiv:2606.19499v1 Announce Type: cross Abstract: Tropes are recurring narrative devices in television and film. We carry out a computational analysis of tropes in the sitcom Friends, using human-curated trope annotations from TVTropes, episode transcripts, and IMDb ratings. Because automatic trope detection remains challenging, we treat existing trope annotations as a curated analytical layer and focus on their downstream narrative and semantic functions. We first examine the relationship between episode-level trope frequency and audience reception. We find a statistically significant positive association between trope count and weighted IMDb ratings, although the modest explanatory power suggests that more than trope density alone explains audience evaluation. We then connect trope annotations to dialogue transcripts and represent trope-related dialogue using TF-IDF-based semantic features. Using PCA and k-means clustering, we group 1,954 distinct tropes into 15 semantically interpre
arXiv:2607.24391v1 Announce Type: new Abstract: AI systems already govern. They rank speech and allocate attention, filter applicants and triage claims. The dominant frame for AI governance, alignment, asks whether such systems pursue the right objectives safely. It cannot answer a prior question: by what right are those objectives set and enforced? This Article argues that legitimacy is an autonomous regulatory objective, distinct from alignment and not secured by it. Legitimacy here is sociological: the belief among those subject to power that it is exercised rightfully. Performance does not produce that belief. We already have the proof of concept. Social media and search delivered enormous gains on every familiar metric and still triggered a legitimacy crisis, because publics questioned who authorized a handful of firms to set the rules of speech, visibility, and knowledge. It is possible to build a benevolent AI and still face a political crisis over its authority. The Article map
arXiv:2607.24035v1 Announce Type: new Abstract: Explainable AI (XAI) is used to assess whether artificial intelligence models rely on meaningful patterns, yet explanations that appear plausible for individual predictions may systematically misrepresent model behavior. This is particularly problematic in medicine, where models may rely on irrelevant signal characteristics rather than disease-specific patterns without being recognizable. We address this challenge using electrocardiogram (ECG) data, for which clinical guidelines provide explicit knowledge about diagnostically relevant signal regions. We introduce a global, guideline-grounded framework that aggregates explanations across heartbeats to evaluate them against clinically defined regions of interest. Using four binary classifiers trained on PTB-XL, we assess 13 gradient-based methods across two categories of patterns: low-amplitude segments and high-amplitude QRS morphology. Our results reveal a systematic failure of methods tr
arXiv:2607.23993v1 Announce Type: new Abstract: Misinformation is deeply embedded in online discourse, with nearly one in five posts during global events generated by bots that amplify false content. In recent years, the use of Generative AI has further lowered the barrier to producing convincing misinformation, yet most digital literacy education still relies on static checklists and single-player inoculation games built for an earlier media landscape. This paper describes how we addressed this educational gap through Capture the Narrative, a four-week multi-university competition in which student teams build LLM-powered bots to influence a simulated election. We report on our custom social-media platform, the competition environment and design of its 4,000 AI-driven Non-Player Character (NPC) citizens, and what running Capture the Narrative at scale actually involved. In our first iteration, 108 teams from 18 Australian universities produced 7,068,206 player-bot posts, approximately
arXiv:2607.23931v1 Announce Type: new Abstract: The aggregation benefit of a committee of artificial intelligence (AI) agents comes from complementary information across members. Classical voting guarantees assume independent errors. Language-model errors often co-occur on the same cases. We combine Sah-Stiglitz screening with error dependence that can differ between good and bad cases. In a homogeneous exchangeable Gaussian-copula model, shared errors create a positive asymptotic error floor for majority voting and can change the approval threshold that minimizes expected loss. We estimate a heterogeneous extension from 174,384 votes cast by 28 language models on four binary-screening benchmarks. Parameters estimated from odd-indexed items predicted committee loss on even-indexed items. For the sampled committee composition, the full-matrix dependence model increased identity-line R^2 from 0.840 under independence to 0.967. In a design-balanced analysis, cost-sensitive threshold selec
arXiv:2607.23733v1 Announce Type: new Abstract: Firms struggle to choose AI projects that pay off: two projects can look equally promising to smart, motivated stakeholders and yet deserve opposite decisions. At the residential real-estate brokerage Compass, one AI product (Likely-to-Sell recommendations) flagged sales outreach opportunities and went on to account for nine figures in annual gross commission revenue. Another championed AI product (a Time-on-Market pricing tool) was rightly shelved. A simple ROI estimate could not distinguish the two. We present expected ROI (eROI), a framework that decomposes each bet into three components and rates them separately: Value if Successful, Likelihood of Success, and Investment Required. Each maps to a question executives can answer before building: How valuable would it be if it worked? How likely is it to work? And what would it cost to implement? Separating the three breaks a common catch-22: teams cannot estimate ROI until they know whet
arXiv:2607.23539v1 Announce Type: new Abstract: AI agents can transact online on behalf of a human principal---browsing, paying, receiving, and reviewing---without linking a transaction to a principal. That architecture starves algorithmic discrimination of its inputs---identity, purchase history, location history, behavioral traces, and demographic proxies---but also forecloses its proof. Disparate-treatment needs comparators; disparate-impact needs protected-class baselines; and Iqbal-era pleading needs specific factual allegations---doctrinal predicates that anonymous transactions never generate. The effects fall asymmetrically: those most vulnerable to discrimination are least able to afford the shield and, when harms remain, least able to prove them. The challenge for the law shifts from detecting and remedying algorithmic discrimination to governing agent-mediated anonymity as civil rights infrastructure: ensuring access to privacy-preserving agents, regulating abuse without forc
arXiv:2607.23519v1 Announce Type: new Abstract: Political audits of large language models (LLMs) usually reduce each to one point on a political compass. But that resting point barely matters in deployment: a model must land somewhere, and what counts is how far, and in which directions, its answers can be steered. That steering runs through the system prompt: the personalization layer a platform sets, or one induced from a user's history, not necessarily written by hand. We run a dispersion-first stress test of prompt-based controllability across 12 ideological personas plus an unsteered baseline, 70 Political Compass items, ten replicates, and seven leading LLMs: GPT-5, Claude, Grok, Gemini, DeepSeek, Kimi, and Qwen (63,700 responses). Contextual framing explains roughly 88%-93% of variance on the economic and society axes, model identity under 3%: responses are highly instruction-adjustable. Models do not shift alike: some move more, and some saturate under extreme framings. Conflic
arXiv:2607.23336v1 Announce Type: new Abstract: The built environment is on the cusp of populating itself with autonomous artificial agents. AI systems that advise, control and coordinate are being deployed across retrofit, operation and mobility faster than their collective behaviour is studied. The dominant framing treats each agent as a tool operating on a passive building, governance reduced to single-agent safety, which is inadequate. A building, a street, or a city is more accurately modelled as a society of negotiating agents: occupants, owners, operators, regulators, and the artificial agents increasingly acting on their behalf. Their interactions are strategic, their information asymmetric, and the outcomes that matter are properties of the whole. The paper proposes a research agenda for constitutional multi-agent governance of the built environment, organised around three problems: mechanism design for retrofit under deep uncertainty, treating public subsidy as a mechanism co
arXiv:2607.23207v1 Announce Type: new Abstract: The emerging infrastructure for AI-agent identity has converged, in industry practice and research proposals alike, on a single resolution of the tension between accountability and privacy: make every agent identifiable. We document a national system in China -- built as national infrastructure and scheduled for public launch in Q3 2026 -- that occupies a different and underexplored point in the same design space: an agent is associated with a verified legal principal without that principal being disclosed to any business-layer participant. Re-identification is possible only to a legal authority acting through due process, by separately compelling two distinct government agencies, neither of which can re-identify alone. We name the mechanism split-knowledge binding and are candid that it is conditional: the separation is structural and procedural, not cryptographic, and a state empowered to compel both agencies can re-identify. The paper
arXiv:2607.22957v1 Announce Type: new Abstract: Restricting access to a dual-use AI model is precautionary only if it delays harmful actors more than defenders. That condition varies across actors: a state agency or organized criminal group may obtain a substitute through theft, distillation, intermediated access, independent development, or a foreign release, while a small utility or open-source maintainer may have no comparable route. We model a laboratory choosing among controlled access, a defender-first window, safeguarded open weights, and minimally restricted open weights. Access inversion occurs when restriction gives an access advantage to adversaries that obtain effective substitutes faster than defenders. Asymmetric empowerment occurs when immediate release adds the most capability to populations least likely to possess a substitute. The policy ranking also depends on relative usefulness, opportunistic misuse, offense-defense conversion, defensive spillovers, safeguard frict
arXiv:2607.22656v1 Announce Type: new Abstract: Objective: This paper investigates the gendered structure of speaker addressee relationships in film dialogue, asking not merely who speaks, but who is spoken to and how conversational dynamics unfold across gender lines. Methods: Using a manually annotated dataset of 4,600 directed dialogue events from 38 film screenplays, we apply network analysis, chi squared tests, paired statistical comparisons, and participation shift analysis across three studies. Key Findings: Male characters dominate as both speakers and addressees corpus wide, even in scenes with more women; cross gender dialogue is directionally symmetric on average but clustered at the film level; and same gender turns diffuse conversational attention while cross gender turns produce tighter dyadic reciprocation. Conclusion: Gender bias in film dialogue operates through the architecture of conversation itself, through exclusion from interaction and structural positioning as ad
arXiv:2607.22640v1 Announce Type: new Abstract: Short-term environmental exposures have been linked to cognitive and behavioral outcomes, although many reported associations may reflect broader geographic and contextual differences. Using longitudinal data from the All of Us Research Program (2018--2024), we linked daily weather and air-pollution exposures to repeated attention-related and subjective cognitive outcomes. Associations were evaluated using pooled, fixed-effects, lagged, and event-study analyses. Additional machine-learning analyses were conducted to explore potential heterogeneity and latent psychosocial structure. Replication analyses were performed using the 2024 Behavioral Risk Factor Surveillance System (BRFSS). Several environmental exposure measures showed small associations with cognitive outcomes in pooled analyses, but most attenuated substantially after accounting for within-location temporal variation. Mediation, sensitivity, and machine-learning analyses yield
arXiv:2607.22619v1 Announce Type: new Abstract: Several international agreements have been proposed to regulate frontier AI development in response to catastrophic risks. However, there is no structured way to evaluate whether these proposals are enforceable, to assess where they might fail in practice, or to determine which combination of policies is most effective. We propose a taxonomy based on the principle that wherever sufficient capacity exists to violate an agreement, it must be under a control regime. This decomposes the problem of ensuring compliance with the agreement into preventing uncontrolled resource acquisition, detecting all capacity outside the control regime, and preventing escape from the control regime. Existing proposals consist of individual policies that address one or more of these sub-problems. Because the compute required for dangerous capabilities may decrease over time, more actors can violate an agreement and enforcement of these policies becomes harder.
arXiv:2607.22617v1 Announce Type: new Abstract: Data centers are critical to today's digital economy, but are also among the largest industrial consumers of freshwater. Beyond the sheer volume of water use, the environmental impact of data center water consumption varies significantly across locations and seasons, depending on local and regional water stress. However, prior research has largely focused on reducing total water use, overlooking that the same unit of water can have drastically different environmental consequences depending on when and where it is consumed. In this paper, we introduce a stress-adjusted water framework that quantifies the true sustainability impact of data center water consumption by incorporating both spatial and temporal water stress. Using the AWARE-US model, we capture county-level monthly variations in water availability and extend this framework to account for the off-site water footprint of electricity generation. Based on this stress-aware accountin
arXiv:2607.22613v1 Announce Type: new Abstract: This paper explores the intersection of memory, place, and identity, examining how new technologies, particularly Apple Vision Pro, can illuminate this nexus. Leveraging digital twins and virtual reality, it investigates how memory is woven into landscapes and urban environments of cultural and historical significance, identifying visual elements that evoke memory and heritage. Applications such as Apple Vision Pro can facilitate image extension to define place identity, informing viewers about cultural and political entities across timelines. Visual storytelling can showcase the evolution of landscapes and the preservation of cultural heritage, while Virtual Reality (VR) enables the recreation of historical landscapes and urban-scapes. This immersive approach invites users to transcend temporal boundaries and experience the past dynamically. Semantic Image Search can support research by uncovering images related to monuments, tradition,
arXiv:2607.22607v1 Announce Type: new Abstract: The prospective clinical evaluation of artificial intelligence in medicine has expanded rapidly, but the global AI clinical trial landscape remains incompletely characterized. We systematically identified AI-related trials registered in ClinicalTrials.gov using a broad keyword search followed by an LLM-based classifier. Each trial was classified across seven dimensions: clinical function, data modality, specialty, AI integration and autonomy, workflow position, translational maturity, and epistemic role. We identified 8,532 AI clinical trials across 32 specialties, with 80% registered from 2019 onward and 30.5% using a randomized controlled design. Imaging-based AI was the largest modality, with 2,475 trials (29%), while clinical text and NLP trials increased seven-fold between 2018 and 2025. Prognostic AI (4,324 trials) slightly exceeded diagnostic AI (3,828 trials), suggesting a shift from disease detection toward risk stratification an
arXiv:2607.22606v1 Announce Type: new Abstract: Health systems are rapidly deploying generative AI assistants that answer patient questions from institution-authored education materials, on the premise that grounding in local content yields consistent guidance. Whether it does depends on a question not previously measured at scale: do the underlying documents themselves agree? We use a structured-output large language model judge to audit 5,730,465 pairwise comparisons across 102 patient-education handbooks from 23 US solid-organ transplant centers, paired with 1,115 patient-derived questions (TransplantQA). Four findings bear directly on deployment: (1) institutional editorial voice statistically transcends organ-type boundaries, with same-center handbooks agreeing across organs more than same-organ handbooks across centers (p = 0.0056); (2) information gaps fall disproportionately on topics central to underrepresented subgroups, with reproductive health showing double jeopardy: it is
arXiv:2607.22605v1 Announce Type: new Abstract: We investigate whether large language models alter medical triage recommendations for identical symptoms when only the patient's socioeconomic status (SES) varies. Using three deployment-tier models (Gemini 3.5 Flash, Claude Sonnet 4.6, GPT-5.4-mini), we hold a single neurological symptom profile fixed and vary the SES signal along two channels: explicit (insurance status, occupation, housing) and implicit (a US ZIP code, with no other socioeconomic information). All three models raise their emergency-room (ER) referral rate for lower-SES patients given the explicit signal (spreads of 13-50 percentage points). The effect is in the protective direction: lower-SES patients are sent to the ER more often, not less. The model's stated reasoning stays clinically near-identical across conditions, so the shift is invisible to a reasoning-trace audit. Critically, sensitivity to the implicit ZIP-code signal is model-dependent: Gemini infers SES fro
arXiv:2607.22604v1 Announce Type: new Abstract: In the age of Artificial Intelligence (AI), Large Language Models, Generative AI and larger frontier AI models, data centres create a significant environmental burden on electricity grids and fresh water resources. Requiring data centre operators and Big Tech under the recast Energy Efficiency Directive (recast EED) to quantify, report and disclose the facility-level energy and water impacts seems to be a step into the right direction towards more transparency and accountability. Yet when two recast EED approved benchmarks - the Power Usage Effectiveness (PUE) and Water Usage Effectiveness (WUE) - can be skewed to create a false sense on efficiency gains, current EU policy pushing for sustainable hyperscale data centre expansion appears misplaced. This paper argues that current PUE and WUE reporting frameworks illustrate what we term the "efficiency paradox," according to which positive scores require retrofitting larger AI data centres a
arXiv:2607.22598v1 Announce Type: new Abstract: Educational chatbots powered by large language models (LLMs) show promising effects on learning outcomes, yet most systems delegate pedagogical decisions such as content selection and didactic structuring implicitly to the LLM, making tutoring strategies difficult to trace, evaluate, and reproduce. This paper presents a didactical-driven teacher assistant for a French-language university course on dimensional modelling, operating without commercial LLM budget or GPU infrastructure. The architecture formalises the instructor's pedagogical reasoning into deterministic modules that handle intent detection, concept linking, and didactic approach selection before any text is generated; the LLM acts solely as a linguistic executor. Evaluation on 195 authentic student questions addresses two research questions. First, we show that standard semantic retrieval alone does not reliably recover the pedagogically required content, thereby justifying t
EdTech to Watch: Series May-June 2026
This annual award celebrates the products, and businesses behind each one, who are transforming education in schools around the world.
When used in the right way AI seems to help test scores and save teacher and staff time, say Syracuse University's Jeff Rubin and Andrew Joncas
Innovative Leader Award - The Higher Vision Drone Program has taken flight thanks to community partnerships and Jennifer Nickerson
Employers expect healthcare costs to rise 9.2% in 2027 as they seek ways to curb rising hospital and pharmacy costs. The post Business Group on Health: Employer Healthcare Costs Projected to Rise 9.2% in 2027 appeared first on MedCity News .
Johnson & Johnson antibody drug Imaavy expanded its label to include the treatment of warm autoimmune hemolytic anemia (wAIHA). Projected to become a blockbuster seller across multiple indications, Imaavy was first approved last year for treating generalized myasthenia gravis. The post J&J’s Imaavy Becomes First FDA-Approved Therapy for Rare Form of Anemia appeared first on MedCity News .
Article URL: https://orgchart.mit.edu/letters/ai-and-education-watershed-moment-mit Comments URL: https://news.ycombinator.com/item?id=49438025 Points: 1 # Comments: 0
K–12 districts aren’t slowing down on technology investments. U.S. schools spent an estimated $30 billion on ed tech in 2024, a figure that is expected to nearly double by 2033. But even the smartest investments can succeed only if the people they’re meant for actually use them. Too often, planning focuses on procurement and rollout, but skips a critical component of implementation: professional development. The CoSN 2026 Driving K–12 Innovation Report emphasizes the importance of professional development as a critical component of building up leaders in school settings: “When schools…
Growing cyberthreats and limited head count create ongoing challenges for security teams at colleges and universities. According to research from Nile, 94% of higher education IT teams are understaffed amid rising cybersecurity threats. Artificial intelligence could be a way to augment their security strategies to free up staff and resources. “For organizations with lean security teams, there is an exciting opportunity for AI to support some of the heavy lifting,” says Ramya Chitrakar, vice president of engineering at Google Cloud Security. “We are already seeing huge gains for…
Here’s how finance leaders can evaluate AI investments in revenue cycle management before committing budget, and the risk exposure most vendor pitches leave out. The post Before You Sign That AI Contract: 7 Questions Every Healthcare CFO Should Ask appeared first on MedCity News .
Before restricting screens, schools should ask what students will lose, including access to books, translation, accessibility features, and more.
arXiv:2608.15382v2 Announce Type: replace-cross Abstract: Large language models (LLMs) are increasingly proposed for healthcare decision support, but their evaluations still reward single-answer accuracy rather than reasoning about interventions, mechanisms, harms, evidence, and uncertainty. We propose a reproducible, graph-centered evaluation framework for intervention-oriented LLM behavior in healthcare and stress-test it in a cardiovascular pilot. The framework has four components: (i) a domain causal knowledge graph in which assertions are first-class, provenance-preserving nodes with stable identifiers; (ii) a scenario-conditioned subgraph extraction step that, given any clinical scenario, retrieves the relevant reified-assertion subgraph; (iii) four controlled grounding conditions that vary how the retrieved subgraph is composed into the model's context (ungrounded C1, knowledge-graph C2, causal-graph C3, integrated C4); and (iv) an automated scoring pipeline, anchored on asserti
arXiv:2608.10279v2 Announce Type: replace-cross Abstract: 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
arXiv:2608.09855v2 Announce Type: replace-cross Abstract: Agentic auto-research is emerging, but most systems treat scientific discovery as goal-oriented optimization against a final benchmark. This paradigm rewards a sparse final verdict and ignores the exploration that precedes it. When agents optimize only the final score, they overfit to the test conditions and sample blindly rather than search. Within a declared research problem, a research agent and a greybox fuzzer for software analysis face the same sparse feedback. A fuzzer rarely finds a bug directly, but coverage makes partial progress observable on every execution. Fuzzers use that dense signal to mutate inputs and allocate effort, rather than merely rank completed runs. Auto-research needs the same two capabilities. First, each experiment must expose a cheap, dense signal of epistemic progress before final scientific validation is available. Second, that signal must determine the next intervention so the agent searches rat
arXiv:2607.12550v3 Announce Type: replace-cross Abstract: The key-value (KV) cache has become the dominant memory cost of transformer inference: it grows with batch size, context length, and depth, and at long context it, rather than the model weights, sets the throughput ceiling. Existing reductions fall into two families. Low-rank methods factor two-dimensional slices of the cache, either per-head matrices or cross-layer feature blocks, and quantization methods lower the bit-width of every entry. Neither exploits the fact that the cache at a layer is naturally a third-order tensor whose three axes, the heads, the tokens, and the features, carry very different amounts of redundancy. We take this tensor view directly. Our method, JoLT (Joint Lagrangian Tucker), applies a partial Tucker decomposition that compresses only the token and feature axes while leaving the head and layer axes intact, then restores the energy that truncation discards with a rotated low-bit residual: a random ort
arXiv:2606.16661v2 Announce Type: replace-cross Abstract: Fixed-length chunking in Retrieval-Augmented Generation (RAG) often leads to boundary fragmentation, where critical evidence is split across segments, degrading retrieval recall. While static windowing and parent retrieval improve recall, they introduce significant token overhead. We propose SCAR (Semantic Continuity-Aware Retrieval), an adaptive retrieval policy that selectively expands neighboring chunks by weighing query-neighbor relevance against a structural continuity penalty. SCAR uses a relative expansion threshold tied to each retrieved chunk's own query-relevance, yielding an approximately scale-invariant decision rule that transfers across embedding models without recalibration. Across four diverse corpora (RFC, GDPR, a 10-K report, and a Merger agreement; N=320 queries; 160 boundary-fragmented), SCAR achieves 92.8% recall on boundary-fragmented queries with only 7.84 chunks, a 22.9% reduction compared to static windo
arXiv:2606.14782v3 Announce Type: replace-cross Abstract: Multimodal Large Language Models (MLLMs) achieve strong vision-language reasoning but incur large KV caches and high decoding latency with long visual contexts. Existing compression methods rely on observation window attention for stable token importance estimation, yet this aggregation can dilute sparse critical evidence and discard answer-relevant tokens under aggressive compression. We identify last query attention as a complementary signal for recovering such evidence, though its irrelevant signals may introduce additional noise. We propose BACON, a plug-and-play method that calibrates observation window attention with last query evidence while suppressing noise through intra-layer coherence and inter-layer persistence. Across diverse benchmarks, models, budgets, and compression methods, BACON improves multimodal KV-cache compression by 7.5% on average under the most aggressive budget, with gains up to 30.9%.
arXiv:2606.11119v2 Announce Type: replace-cross Abstract: Reinforcement learning with verifiable rewards (RLVR) is a promising approach for enhancing reasoning and agentic behavior in large language models. However, rollout-intensive policy optimization is often limited by insufficient reward contrast, arising when overly simple or complex prompts generate low-variance feedback and when outcome-only rewards assign the same terminal assessment to every decision in a multi-turn rollout. Past efforts have focused on allocating available rollout resources to promising prompts, yet they only leverage sample informativeness at the prompt level and neglect variation in prefix-level informativeness across turns within the same rollout. This work targets multi-turn agentic RL by modeling each ReAct-style thought-action-observation turn as a semantically distinct node, allowing budget allocation to extend from prompt roots to turn-level prefixes with further continuations, which naturally forms
arXiv:2606.09887v2 Announce Type: replace-cross Abstract: Reinforcement learning (RL) for large language models usually supervises reasoning with scalar outcome rewards, such as binary correctness. Such rewards provide an optimization direction but rarely explain how a model should revise its mistaken reasoning, which can encourage shortcut learning and brittle policies. We propose \textbf{SocraticPO} (Socratic Policy Optimization), a policy-optimization framework that augments RL rollouts with Socratic-style natural-language guidance. During rollout, the student first answers independently; if the answer is incorrect, a teacher diagnoses the attempt and provides concise corrective guidance, after which the student continues under the expanded context. Crucially, this guidance is paired with reward decay: correct answers obtained after teacher intervention only receive decayed rewards, preventing the policy from treating teacher help as a free path to reward. Since SocraticPO only modi
arXiv:2606.00566v2 Announce Type: replace-cross Abstract: As language models take on agentic roles that call APIs, read tool outputs, and act on third-party content, their attack surface expands beyond what users type. Whether they treat a malicious instruction the same way regardless of where it arrives has not been studied systematically. We introduce the Safety Asymmetry Score (SAS), measuring how a model's susceptibility to adversarial content shifts depending on whether it arrives in the user message, tool metadata, or tool output, using matched payload pairs that hold the malicious text identical and vary only the channel. Across 10 production LLMs and three attack families, general-purpose models sharply discount instructions arriving as tool metadata relative to identical instructions in the user message, while agent-native models discount them far less. This differential survives an affordance-matched control equalizing tool availability and scoring, and a size-controlled mixe
arXiv:2605.18663v2 Announce Type: replace-cross Abstract: As LLM benchmarks saturate, the evaluation community has pursued two strategies to increase difficulty: escalating knowledge demands (GPQA, HLE) or removing knowledge entirely in favor of abstract reasoning (ARC-AGI). The first conflates memorization with capability; the second divorces reasoning from the practical contexts in which it matters. We take a different approach. The Grounded Integration Measure (GIM) is a benchmark of 820 original problems (615 public, 205 private) where difficulty comes from integration; individual problems require coordinating multiple cognitive operations (constraint satisfaction, state tracking, epistemic vigilance, audience calibration) over broadly accessible knowledge, so that reasoning stays grounded in realistic tasks without being gated on specialized expertise. Each problem is an original expert-authored composition, majority with rubric-decomposed scoring. We calibrate a judge-aware conti
arXiv:2605.17610v2 Announce Type: replace-cross Abstract: The rapid growth of online video platforms and AI-generated content has made reliable video guardrails a key challenge for safety and real-world deployment. While most videos can be screened through fast pattern recognition, a small subset requires deeper reasoning over temporally complex content and nuanced policy constraints. Existing approaches typically rely on large vision-language models applied uniformly across all inputs, resulting in high inference costs and inefficient allocation of computation. We propose SafeLens, a video guardrail framework that introduces a fast-and-slow inference architecture for efficient and accurate content moderation with variable computational cost across inputs. Additionally, we construct a high-quality dataset by applying influence-guided filtering to the SafeWatch Dataset, retaining only 2.4% of the original data. To further address limitations of training-time scaling, we enable test-time
arXiv:2604.25098v3 Announce Type: replace-cross Abstract: Large Language Models (LLMs) now exhibit remarkable reasoning capabilities through test-time compute scaling (TTS), with impressive performance across math and coding benchmarks. In parallel, research in model compression has developed pruning methods that seek to remove redundant/detrimental parameters without sacrificing task performance. The intersection of these two research advancements lays the foundation for our work. Specific to reasoning LLMs, prior work has shown that structured pruning (methods which remove entire set of layer blocks), significantly degrades TTS reasoning performance. However, in this work, we revisit this assumption and investigate whether unstructured pruning (methods that carefully remove only certain redundant/detrimental weights) exhibits similar limitations. Surprisingly, our extensive experiments across four reasoning benchmarks on two reasoning LLMs: s1.1-7B and Qwen3-8B, consistently show tha
arXiv:2604.16242v2 Announce Type: replace-cross Abstract: Reinforcement learning with verifiable rewards (RLVR) typically optimizes for outcome rewards without imposing constraints on intermediate reasoning. This leaves training susceptible to reward hacking, where models exploit loopholes (e.g., spurious patterns in training data) in the reward function to achieve high scores without solving the intended task. These reward-hacking behaviors are often implicit, as the intermediate chain-of-thought (CoT) may appear plausible on the surface, limiting the effectiveness of purely text-based monitoring. We propose Gradient Fingerprint (GRIFT), a method for detecting reward hacking using models' internal computations. Given a prompt and a model-generated CoT, GRIFT computes gradients of the CoT conditioned on the prompt and compresses them into a compact representation, which is then used to assess whether the CoT reflects reward hacking behavior. Across verifiable reasoning benchmarks spann
arXiv:2604.09746v2 Announce Type: replace-cross Abstract: As large language models (LLMs) are increasingly deployed as autonomous agents, understanding how strategic behavior emerges in multi-agent environments has become an important alignment challenge. We take a neutral empirical stance and construct a controlled environment in which strategic behavior can be directly observed and measured. We introduce a large-scale multi-agent simulation in a simplified model of New York City, where LLM-driven agents interact under opposing incentives. Blue agents aim to reach their destinations efficiently, while Red agents attempt to divert them toward billboard-heavy routes using persuasive language to maximize advertising revenue. Hidden identities make navigation socially mediated, forcing agents to decide when to trust or deceive. We study policy learning through an iterative simulation pipeline that updates agent policies across repeated interaction rounds using Kahneman-Tversky Optimizatio
arXiv:2604.08649v2 Announce Type: replace-cross Abstract: Modern financial systems generate vast quantities of transactional and event-level data that encode rich economic signals. This paper presents PRAGMA, a family of foundation models for banking event sequences. Our approach pre-trains a Transformer-based architecture with masked modelling on a large-scale, heterogeneous banking event corpus using a self-supervised objective tailored to the discrete, variable-length nature of financial records. The resulting model supports a wide range of downstream tasks such as credit scoring, fraud detection, and lifetime value prediction: strong performance can be achieved by training a simple linear model on top of the extracted embeddings and can be further improved with lightweight fine-tuning. Through extensive evaluation on downstream tasks, we demonstrate that PRAGMA achieves superior performance across multiple domains directly from raw event sequences, providing a general-purpose repre
arXiv:2603.02229v2 Announce Type: replace-cross Abstract: Safety post-training has been studied extensively in single-step "chat" settings where safety typically refers to refusing harmful requests. We study an "agentic" (i.e., multi-step, tool-use) setting where safety refers to harmful actions directly taken by the LLM. We investigate the effects of using direct preference optimization (DPO) to optimize safety and/or helpfulness on the ToolEmu agentic benchmark. First, we find that safety training largely persists through subsequent helpfulness training. Second, we find a consistent negative linear correlation ($R^2 = 0.77$) between safety and helpfulness when considering all training configurations together. Even post-training on both metrics simultaneously simply results in another point on the same trend line rather than yielding a "best of both worlds" strategy, despite the presence of such strategies in our dataset. Overall, our findings underscore the need for a better understa