EdTech Discovery
Hermes

An instrument for spotting the next edtech opportunity — generated ideas, each traced to the real-world signals behind it.

Updated Jun 24, 2026 · 10 ideas · 1304 signals
Admin mode — curation controls visible. Keep this URL (with token) private.

Signals

The evidence library — the raw signals the pipeline is watching across the education ecosystem. Every idea is built from these.

behavior Tue, 06 Jan 2026 12:06:12 +0000
HN: tutoring

AI tutoring outperforms in-class active learning

Article URL: https://www.nature.com/articles/s41598-025-97652-6 Comments URL: https://news.ycombinator.com/item?id=46511304 Points: 1 # Comments: 0

Source ↗
behavior Tue, 06 Jan 2026 10:00:00 +0000
eSchool News

The rise of remote psychoeducational testing: What school leaders need to know

Special education is at a breaking point. Across the country, more children than ever are being referred for evaluations to determine whether they qualify for special education services.

Source ↗
behavior Tue, 05 Nov 2024 16:03:46 +0000
HN: online learning

WebRL: Training LLM Web Agents via Self-Evolving Online Reinforcement Learning

Article URL: https://arxiv.org/abs/2411.02337 Comments URL: https://news.ycombinator.com/item?id=42052558 Points: 23 # Comments: 1

Source ↗
behavior Tue, 05 May 2026 18:47:02 GMT
EdSurge

Educators: Why Are You Thinking of Leaving the Field?

EdSurge wants to hear from educators who have recently left or plan to leave their jobs for another sector.

Source ↗
behavior Tue, 05 May 2026 15:09:05 +0000
eSchool News

ClassMate by World Book Recognized with Prestigious ISTE Seal

CHICAGO, May 5, 2026 — ClassMate by World Book, the leading platform of trusted content that helps build knowledge through ... Read more

Source ↗
behavior Tue, 05 May 2026 10:00:00 +0000
eSchool News

When AI means something different in every classroom

In many schools, AI is being handled through individual teacher decisions rather than a shared structure. That makes sense in the short term. Teachers are responding in real time, trying to protect their classrooms, their expectations, and their students.

Source ↗
technology Tue, 05 May 2026 09:00:00 +0000
Tech & Learning

Edtech's Big Tobacco Moment Is Here. Schools Can't Afford to Miss the AI Reckoning That Follows

Conversations with Kevin Hogan: Author and educator Andrew Marcinek argues that the Meta lawsuit is the inevitable outcome of 20 years of algorithmic manipulation — and that schools have a narrow window to get AI right before history repeats itself.

Source ↗
behavior Tue, 04 Nov 2025 10:00:00 +0000
eSchool News

Reliable and relatable resources build confident students: The triple Rs of scholastic esports

I know what it feels like to stand in front of a classroom that does not have enough. Not enough computers. Not enough up-to-date software and technical tools. Not enough resources to give every student the experience they deserve. When students notice these gaps, they notice more than the missing tools.

Source ↗
behavior Tue, 03 Mar 2026 10:00:00 +0000
eSchool News

Too many apps, too many missed messages

In our district, families were checking multiple apps just to keep up with school communication. One child’s teacher posted in one platform. Another school used something different. District updates lived somewhere else entirely.

Source ↗
behavior Tue, 03 Jun 2025 17:56:15 +0000
HN: online learning

Which tech jobs can you realistically get by learning online without a degree?

I'm 28 years old, and to be honest, I haven’t done much with my life so far. Recently, I stumbled across programming and cybersecurity online, and the positive aspects of both fields really caught my attention. I’ve always been patient with solving problems, and I actually enjoy figuring things out. It gives me a sense of accomplishment. I'm also fairly tech-savvy, and for the first time in a while, I feel like I might have found something I could be genuinely good at. The thing is, I’m not in a position to go to college or attend any formal institution. I’ve seen stories about people learning online and breaking into tech, but I’ve also read a lot of negative takes. Even graduates sometimes struggle to land jobs. So I’m genuinely curious: if I commit to learning and work really hard, do I realistically have a chance to turn my life around and get into programming or cybersecurity without a degree? Comments URL: https://news.ycombinator.com/item?id=44172736 Points: 3 # Comments: 2

Source ↗
behavior Tue, 03 Feb 2026 17:14:12 +0000
eSchool News

School Specialty Expands Learning Beyond the Screen with New Outdoor Furniture Line

Childcraft expands early learning beyond four walls and screens with durable, sustainable furniture designed for outdoor discovery.

Source ↗
behavior Tue, 03 Feb 2026 15:15:26 +0000
eSchool News

Avantis Education Transforms Classroom Inclusivity with Launch of EduverseTHRIVE at TCEA 2026

Chicago, (February 1, 2026) — Avantis Education, a global leader in virtual and augmented reality (VR/AR) technology for K-12 schools, ... Read more

Source ↗
behavior Tue, 03 Feb 2026 10:00:00 +0000
eSchool News

The digital divide redux: Why AI is the new broadband

Remember the early 2000s, back when high-speed internet felt like a luxury reserved for the tech elite and the lucky few with deep pockets? We called it the Broadband Gap or Equity of Access, and it influenced who got ahead and who got left behind.

Source ↗
behavior Tue, 02 Sep 2025 15:40:51 +0000
HN: tutoring

Erik's Plea in the Free Press: Bring Back Aristocratic Tutoring

Article URL: https://www.theintrinsicperspective.com/p/eriks-plea-in-the-free-press-bring Comments URL: https://news.ycombinator.com/item?id=45104556 Points: 1 # Comments: 0

Source ↗
behavior Tue, 02 Jun 2026 22:25:54 GMT
EdSurge

Report: School IT Officials Worried About AI Adoption, Cybersecurity

School districts are adopting AI policies more than ever, but a lack of resources, funding and expertise has some still concerned.

Source ↗
behavior Tue, 02 Jun 2026 10:00:00 +0000
eSchool News

How districts can build a shared AI structure

In the second week of January, a senior mathematics teacher with 22 years in the classroom raised a hand at the end of a staff meeting and asked a question that changed the way I now design AI literacy work for entire faculties.

Source ↗
behavior Tue, 02 Jun 2026 09:15:00 +0000
Getting Smart

The Conditions That Make Durable Skills Real: How Schools and Systems Build for Agency, Identity, and Vision

How schools build durable skills through authentic work, reflection, relationships, and learner-centered design. The post The Conditions That Make Durable Skills Real: How Schools and Systems Build for Agency, Identity, and Vision appeared first on Getting Smart .

Source ↗
technology Tue, 02 Jun 2026 09:00:00 +0000
Tech & Learning

Devices Down Is The Wrong Goal

Where the AFT's new 10-point plan gets it right, where it falls short, and why “devices down” is not the path to meaningful learning.

Source ↗
technology Tue, 02 Jun 2026 09:00:00 +0000
Tech & Learning

Time to Clean House

Conversations with Kevin Hogan: CoSN Board Member Kris Hagel downloads on the state of edtech in US schools.

Source ↗
behavior Tue, 02 Dec 2025 10:00:00 +0000
eSchool News

How one school reimagined learning spaces–and what others can learn

When Collegedale Academy, a PreK–8 school outside Chattanooga, Tennessee, needed a new elementary building, we faced a choice that many school leaders eventually confront: repair an aging facility or reimagine what learning spaces could be.

Source ↗
behavior Thu, 30 Oct 2025 10:00:00 +0000
eSchool News

How districts can avoid 4 hidden costs of outdated facilities systems

School leaders are under constant pressure to stretch every dollar further, yet many districts are losing money in ways they may not even realize. The culprit? Outdated facilities processes that quietly chip away at resources, frustrate staff, and create ripple effects across learning environments.

Source ↗
behavior Thu, 30 Apr 2026 16:21:04 GMT
EdSurge

Quality Concerns Remain as States Invest More Than Ever in Preschool Programs

A new report found states hit an all-time high for both spending and enrollment, but the quality of the programs remains a concern.

Source ↗
behavior Thu, 30 Apr 2026 10:00:00 +0000
eSchool News

The 4 keys to creating meaningful student-led inquiry

Does the thought of student-led inquiry make you nervous? For some teachers, handing over control of the classroom to their students sounds like an invitation for disaster.

Source ↗
technology Thu, 30 Apr 2026 09:00:00 +0000
Tech & Learning

4 Ways Teachers Are Using AI

Researchers looked at more than 150,000 prompts from more than 4,400 K-12 teachers interacting with AI. Here's what they found.

Source ↗
technology Thu, 29 May 2025 06:34:47 +0000
HN: edtech

Rethinking African edtech: Why AI alone won't be enough

Article URL: https://techcabal.com/2025/05/28/rethinking-african-edtech/ Comments URL: https://news.ycombinator.com/item?id=44123628 Points: 1 # Comments: 0

Source ↗
behavior Thu, 29 Jan 2026 10:00:00 +0000
eSchool News

AI in edtech: The 2026 efficacy imperative

AI has crossed a threshold. In 2026, it is no longer a pilot category or a differentiator you add on. It is part of the operating fabric of education, embedded in how learning experiences are created, how learners practice, how educators respond, and how outcomes are measured. That reality changes the product design standard.

Source ↗
behavior Thu, 28 May 2026 21:26:59 GMT
EdSurge

Why College Degrees Matter in the Age of AI

Technical skills are changing rapidly. A college education teaches students something more durable.

Source ↗
behavior Thu, 28 May 2026 10:00:00 +0000
eSchool News

Finding the “low way”: Reclaiming creativity in schools

When my daughter was little, every time we climbed into the car, she’d look up and ask, “Are we going to take the low way?”

Source ↗
technology Thu, 28 May 2026 09:00:00 +0000
Tech & Learning

What is Vibe Coding? Creating Code with AI Explained

Vibe coding can feel instant, but it is not simply pressing a button and getting a finished app.

Source ↗
behavior Thu, 27 Nov 2025 10:00:00 +0000
eSchool News

Students must intentionally develop durable skills to thrive in an AI-dominated world

As AI increasingly automates technical tasks across industries, students’ long-term career success will rely less on technical skills alone and more on durable skills or professional skills, often referred to as soft skills. These include empathy, resilience, collaboration, and ethical reasoning--skills that machines can’t replicate.

Source ↗
technology Thu, 27 Mar 2025 17:11:06 +0000
HN: edtech

Ask HN: Why aren't there more EdTech AI startups?

We've been promised that AI will introduce personalised tutoring, that it will replace traditional schooling, etc. However, I see fewer and fewer edtech startups these days... Chegg, Udemy, Busuu and many others are on the decline. What's happening to Edtech? Comments URL: https://news.ycombinator.com/item?id=43495666 Points: 2 # Comments: 0

Source ↗
behavior Thu, 27 Feb 2025 13:30:57 +0000
HN: online learning

InkSight: Offline-to-Online Handwriting Conversion by Learning to Read and Write

Article URL: https://charlieleee.github.io/publication/inksight/ Comments URL: https://news.ycombinator.com/item?id=43194202 Points: 3 # Comments: 0

Source ↗
behavior Thu, 26 Mar 2026 10:00:00 +0000
eSchool News

Building AI for kids: A developer’s guide to age-appropriate safety architecture

When I shipped Gramms AI to the App Store, I ran straight into a question that every developer building for kids will eventually face: What does “age-appropriate” actually mean in practice? And how do you build systems that enforce it reliably?

Source ↗
behavior Thu, 26 Mar 2026 08:00:00 GMT
EdSurge

Which Education Jobs Are Growing the Fastest? Mostly Non-Classroom Roles.

Student support and tech professions are projected to make gains while teaching positions shrink.

Source ↗
behavior Thu, 26 Feb 2026 10:00:00 +0000
eSchool News

Strengthening the PK-12 leadership pipeline during a succession crisis

Across the country, districts are confronting a growing PK-12 leadership pipeline crisis. Veteran principals, assistant principals, and district administrators are retiring at increasing rates, yet there is not a sufficiently prepared pool of aspiring leaders ready to step into these roles.

Source ↗
regulation Thu, 25 Jun 2026 01:40:40 +0000
The 74

LAUSD Board Appoints Longtime Administrator Andres Chait as Next Superintendent

The Los Angeles Unified Board voted unanimously to appoint Andres Chait, a longtime district administrator, as superintendent days after his predecessor resigned. “This board’s decision reflects the confidence in Mr. Chait’s leadership, his decades of service to Los Angeles Unified, and his demonstrated ability to guide the district during this period of transition,” said board […]

Source ↗
technology Thu, 25 Jun 2026 00:00:00 -0400
arXiv cs.CL

ESBMC-PLC+: A Unified IEC 61131-3 Formal Verification Framework as a PLCverif Successor

arXiv:2606.23870v2 Announce Type: replace-cross Abstract: PLCverif is the most mature open-source platform for PLC formal verification, developed at CERN and in production use since 2019. Yet it has two fundamental limitations: no support for Ladder Diagram (LD) programs, the dominant PLC notation, and reliance on CBMC as its primary backend, which restricts verification to bounded proofs. The PLCverif authors themselves identified ESBMC as the appropriate backend improvement. Prior work established ESBMC-PLC (a textual LD frontend with k-induction) and ESBMC-GraphPLC (graphical PLCopen XML support); together, they cover LD with unbounded proofs but not Structured Text (ST), and graphical LD with timer/counter function blocks remains unverifiable. This paper presents ESBMC-PLC+, a unified framework that closes both gaps: (1) an ST/SCL frontend via the MATIEC IEC 61131-3 compiler, routing C-compiled ST to ESBMC with nondeterministic input modeling and YAML property injection; (2) functi

Source ↗
technology Thu, 25 Jun 2026 00:00:00 -0400
arXiv cs.CL

Memory Contagion: Cross-Temporal Propagation of Evaluator Bias via Agent Memory

arXiv:2606.23195v2 Announce Type: replace-cross Abstract: Large Language Model (LLM) agents increasingly rely on memory systems to maintain long-term coherence. Recent work shows that agent memories degrade during continuous consolidation. However, existing research assumes memories are derived from unbiased experiences. In this work, we identify and formalize a novel phenomenon: Memory Contagion -- the cross-temporal propagation of evaluator bias through agent memory. We show that when agents are trained or guided by biased evaluators, their experiences become biased; when these trajectories are stored and consolidated into memory, the bias propagates to future agents retrieving from the same memory store, even when consolidation is perfect (oracle). Across two bias types (length preference, authority bias) and four experimental phases, we demonstrate: (1) Memory Contagion occurs for length bias even with perfect consolidation on older models (Gamma_A = 13.18, DeepSeek V4-Chat), while

Source ↗
technology Thu, 25 Jun 2026 00:00:00 -0400
arXiv cs.CL

SingGuard: A Policy-Adaptive Multimodal LLM Guardrail with Dynamic Reasoning

arXiv:2606.22873v2 Announce Type: replace-cross Abstract: Vision-language models (VLMs) are increasingly deployed in consumer, medical, financial, and enterprise applications. This broad deployment expands the safety surface: risks can arise from multimodal question answering, assistant responses, and cross-modal composition, while moderation policies may vary across products, regions, and deployment stages. Most existing guardrails either rely on fixed taxonomies or target only a narrow set of interaction settings, which limits their adaptability when safety rules change at deployment time. We present \textbf{SingGuard}, a policy-adaptive multimodal guardrail model family for safety assessment in multimodal conversations. SingGuard treats the active policy as a runtime input: given natural-language rules, it checks the target content against the active policy rule by rule and predicts both the safety label and the triggered rule. To balance efficiency and interpretability, SingGuard s

Source ↗
technology Thu, 25 Jun 2026 00:00:00 -0400
arXiv cs.CL

VADAOrchestra: Neurosymbolic Orchestration of Adaptive Reasoning Workflows

arXiv:2606.22485v2 Announce Type: replace-cross Abstract: Decision-making in real-world settings rarely follows a fixed script. Instead, it unfolds as a dynamic reasoning process in which the appropriate course of action evolves as new context and data become available. Traditional Business Process Management systems provide rigor, determinism, and auditability, yet they generally struggle to adapt their execution at runtime. Conversely, agentic systems based on Large Language Models (LLMs) bring flexibility to decision-making, but they are inherently opaque, often unreliable, and suffer from significant scalability constraints when operating over large datasets. To combine these complementary paradigms, we introduce VADAOrchestra, a neurosymbolic framework that models complex workflows as evolving reasoning processes. The framework adopts a hybrid approach: given a user query and a collection of data sources, an LLM-based orchestrator incrementally plans and adapts the workflow. This

Source ↗
technology Thu, 25 Jun 2026 00:00:00 -0400
arXiv cs.CL

Toten: A Knowledge-Based System For Structure-Preserving Representation Of Physical Quantities And Technical Notation In Brazilian Portuguese

arXiv:2606.19626v2 Announce Type: replace-cross Abstract: AI pipelines that reason quantitatively over technical text depend on input where physical quantities, numbers, units, and symbolic expressions arrive intact; when these entities fragment at tokenization, errors propagate downstream. Byte-Pair Encoding, optimized for vocabulary compression, is blind to such entities and fragments them into arbitrary subwords -- a problem aggravated in technical Brazilian Portuguese. We present TOTEN, a knowledge-based system whose input representation preserves each technical entity as a whole, typed unit: vocabulary is not derived statistically but classified declaratively under a formal ontology of engineering entities (OEE). The core is the triple : types, principles, and invariants; a classifier mapping raw text into typed regions; and instantiators yielding a self-descriptive representation. Integrity rests on deterministic coupling to three external authorities: Pint (dimensional), Unicode

Source ↗
technology Thu, 25 Jun 2026 00:00:00 -0400
arXiv cs.CL

IndicContextEval: A Benchmark for Evaluating Context Utilisation in Audio Large Language Models Across 8 Indic Languages

arXiv:2606.19157v2 Announce Type: replace-cross Abstract: AudioLLMs enable speech recognition conditioned on textual prompts such as domain descriptions or entity lists. However, it remains unclear whether these models genuinely utilise such context or rely on parametric knowledge learned during pretraining. Existing benchmarks cannot answer this question because they evaluate transcription under fixed prompting conditions and rarely include explicit contextual inputs. We introduce IndicContextEval, a 56-hour multilingual benchmark of natural speech from 555 speakers across 8 Indian languages and 23 professional domains. We design a 7-level prompting framework that progressively introduces contextual signals, including metadata, natural-language descriptions, entity lists in English and native script, and adversarial prompts with incorrect entities. Evaluating five models reveals substantial differences in context utilisation behaviour, highlighting the need for explicit evaluation of

Source ↗
technology Thu, 25 Jun 2026 00:00:00 -0400
arXiv cs.CL

daVinci-kernel: Co-Evolving Skill Selection, Summarization, and Utilization via RL for GPU Kernel Optimization

arXiv:2606.16497v2 Announce Type: replace-cross Abstract: GPU kernel optimization represents a paradigm where functional correctness is assumed and execution efficiency is the objective. We present daVinci-kernel, a reinforcement learning framework that couples skill discovery with skill exploitation through a dynamically evolving skill library. daVinci-kernel jointly trains three agents sharing one LLM backbone: a Skill Selection Agent that retrieves relevant techniques via BM25 and LLM reranking, a Policy Agent that generates multi-turn CUDA/Triton kernels conditioned on selected skills, and a Skill Summary Agent that distills successful rollouts into reusable skills. Candidate skills are added only after execution-based verification confirms reproducible speedups. All three agents share a single LLM backbone, are initialized via a structured SFT cold start on diversity-filtered data, and are then jointly optimized end-to-end with multi-turn REINFORCE and per-agent advantage estimati

Source ↗
technology Thu, 25 Jun 2026 00:00:00 -0400
arXiv cs.CL

MemDreamer: Decoupling Perception and Reasoning for Long Video Understanding via Hierarchical Graph Memory and Agentic Retrieval Mechanism

arXiv:2606.07512v2 Announce Type: replace-cross Abstract: Current Vision-Language Models struggle with hours-long videos because processing full-length visual sequences induces prohibitive token explosion and attention dilution. To overcome this, we introduce MemDreamer to decouple perception and reasoning, shifting long-video understanding into an agentic exploration process. As a plug-and-play framework, it incrementally streams videos to construct a Hierarchical Graph Memory, a top-down three-tier architecture for semantic abstraction, anchored by a foundational graph capturing spatiotemporal and causal relations. During inference, the reasoning model employs agentic tool-augmented retrieval, navigating hierarchies, searching nodes, and traversing logical edges via an Observation-Reason-Action loop. Experiments show MemDreamer achieves SOTA results across four mainstream benchmarks, narrowing the gap with human experts to only 3.7 points. It constrains the reasoning context window t

Source ↗
technology Thu, 25 Jun 2026 00:00:00 -0400
arXiv cs.CL

Continual Knowledge Updating in LLM Systems: Learning Through Multi-Timescale Memory Dynamics

arXiv:2605.05097v3 Announce Type: replace-cross Abstract: LLMs are trained once, then deployed into a world that never stops changing. External memory compensates for this, but most systems manage it explicitly rather than letting it adapt on its own. Biological memory works differently: coupled multi-timescale dynamics make new associations immediately usable, strengthen what repetition confirms, and let the rest fade. We argue that external memory should follow a similar principle. In Memini, this view takes the form of an associative memory that organizes knowledge as a directed graph. Each edge carries two coupled internal variables, one fast and one slow, following the Benna-Fusi model of synaptic consolidation. From this coupling, episodic sensitivity, gradual consolidation, and selective forgetting are expected to emerge as facets of a single mechanism, reframing external memory as a learning substrate that reorganizes through its own dynamics. This workshop article describes an

Source ↗
technology Thu, 25 Jun 2026 00:00:00 -0400
arXiv cs.CL

CoLA: Cross-Modal Low-rank Adaptation for Multimodal Downstream Tasks

arXiv:2604.03314v2 Announce Type: replace-cross Abstract: Foundation models have revolutionized AI, but adapting them efficiently for multimodal tasks, particularly in dual-stream architectures composed of unimodal encoders, such as DINO and BERT, remains a significant challenge. ParameterEfficient Fine-Tuning (PEFT) methods like LowRank Adaptation (LoRA) enable lightweight adaptation, yet they operate in isolation within each modality, limiting their ability in capturing cross-modal interactions. In this paper, we take a step in bridging this gap with Cross-Modal LowRank Adaptation (CoLA), a novel PEFT framework that extends LoRA by introducing a dedicated inter-modal adaptation pathway alongside the standard intra-modal one. This dual-path design enables CoLA to adapt unimodal foundation models to multimodal tasks effectively, without interference between modality-specific and crossmodal learning. We evaluate CoLA across a range of vision-language (RefCOCO, RefCOCO+, RefCOCOg) and au

Source ↗
technology Thu, 25 Jun 2026 00:00:00 -0400
arXiv cs.CL

Speech Codec Probing from Semantic and Phonetic Perspectives

arXiv:2603.10371v2 Announce Type: replace-cross Abstract: Speech tokenizers are essential for connecting speech to large language models (LLMs) in multimodal systems. Speech tokenizers are expected to preserve both semantic and acoustic information for downstream understanding and generation tasks. However, emerging evidence suggests that the term "semantic" in speech processing does not align with linguistic lexical-semantic, leading to a mismatch between speech and text modality. In this paper, we systematically analyze the information encoded by several widely used speech tokenizers, evaluating their lexical-semantic and phonetic content through three tasks. Our results show that current tokenizers primarily capture phonetic rather than lexical-semantic structure, deriving practical implications for the design of next-generation speech tokenization methods. Code is released to public at https://github.com/Alexuan/codec_probing_release.

Source ↗
technology Thu, 25 Jun 2026 00:00:00 -0400
arXiv cs.CL

CLEF HIPE-2026: Evaluating Accurate and Efficient Person-Place Relation Extraction from Multilingual Historical Texts

arXiv:2602.17663v2 Announce Type: replace-cross Abstract: HIPE-2026 is a CLEF evaluation lab dedicated to person-place relation extraction from noisy, multilingual historical texts. Building on the HIPE-2020 and HIPE-2022 campaigns, it extends the series toward semantic relation extraction by targeting the task of identifying person-place associations in multiple languages and time periods. Systems are asked to classify relations of two types -- $at$ ("Has the person ever been at this place?") and $isAt$ ("Is the person located at this place around publication time?") -- requiring reasoning over temporal and geographical cues. The lab introduces a three-fold evaluation profile that jointly assesses accuracy, computational efficiency, and domain generalization. By linking relation extraction to large-scale historical data processing, HIPE-2026 aims to support downstream applications in knowledge-graph construction, historical biography reconstruction, and spatial analysis in digital hum

Source ↗
technology Thu, 25 Jun 2026 00:00:00 -0400
arXiv cs.CL

SPARC: Separating Perception And Reasoning Circuits for Test-time Scaling of VLMs

arXiv:2602.06566v3 Announce Type: replace-cross Abstract: Despite recent successes, test-time scaling -- i.e., dynamically expanding the token budget during inference as needed -- remains brittle for vision-language models (VLMs). Unstructured visual reasoning chains entangle perception and reasoning, leading to long, disorganized contexts where small perceptual mistakes may cascade into completely wrong answers. Reasoning also requires expensive reinforcement learning with hand-crafted rewards. Here, we introduce SPARC (Separating Perception And Reasoning Circuits), a modular framework that explicitly decouples visual perception from reasoning. Inspired by sequential sensory-to-cognitive processing in the brain, SPARC implements a two-stage pipeline where the model first performs explicit visual search to localize question-relevant regions, then conditions its reasoning on those regions to produce the final answer. This separation enables independent test-time scaling with asymmetric

Source ↗
technology Thu, 25 Jun 2026 00:00:00 -0400
arXiv cs.CL

Streaming-dLLM: Accelerating Diffusion LLMs via Suffix Pruning and Dynamic Decoding

arXiv:2601.17917v3 Announce Type: replace-cross Abstract: Diffusion Large Language Models (dLLMs) offer a compelling paradigm for natural language generation, leveraging parallel decoding and bidirectional attention to achieve superior global coherence compared to autoregressive models. While recent works have accelerated inference via KV cache reuse or heuristic decoding, they overlook the intrinsic inefficiencies within the block-wise diffusion process. Specifically, they suffer from spatial redundancy by modeling informative-sparse suffix regions uniformly and temporal inefficiency by applying fixed denoising schedules across all the decoding process. To address this, we propose Streaming-dLLM, a training-free framework that streamlines inference across both spatial and temporal dimensions. Spatially, we introduce attenuation guided suffix modeling to approximate the full context by pruning redundant mask tokens. Temporally, we employ a dynamic confidence aware strategy with an earl

Source ↗
Showing 401–450 of 1304 signals
← Prev Page 9 of 27 Next →