https://aijourney.space/perplexity-ai/news/ Changes at Perplexity AI, logged from its own announcements, documentation and price pages, never from another publisher, each with what it means for you. en https://aijourney.space/updates/2026-11-04-output-aware-block-influence-jacobian-lens-weighting-for-depth-pruning/ 2026-11-04-output-aware-block-influence-jacobian-lens-weighting-for-depth-pruning Wed, 04 Nov 2026 00:00:00 GMT What happened. On September 30, 2026, researchers affiliated with Perplexity AI submitted a paper to arXiv titled Output-Aware Block Influence: Jacobian-Lens Weighting for Depth Pruning, introducing a scoring method called J-BI that reweights Block Influence pruning scores using a Jacobian Lens transport map, tested on Qwen3-8B, Llama-3.1-8B-IT, and Gemma-3-12B-IT. Sources. arXiv abstract page for arXiv:2611.00015, submitted 30 September 2026: https://arxiv.org/abs/2611.00015 (read 4 Nov 2026) Why it matters. For teams running large models on tight hardware budgets, a pruning method that holds onto 19 more points of accuracy than the older block-influence ranking changes how much of a model can be cut before an answer engine's replies start to degrade in ways a user would notice. Perplexity AI's own product depends on calling several underlying models quickly, so cheaper inference work like this could eventually widen the gap between what a free tier and a paid tier of its answer engine can afford to run, though this paper is exploratory research rather than a shipped change to the service. https://aijourney.space/updates/2026-11-03-byod-build-your-own-dlm-efficient-autoregressive-to-diffusion-language-m/ 2026-11-03-byod-build-your-own-dlm-efficient-autoregressive-to-diffusion-language-m Tue, 03 Nov 2026 00:00:00 GMT What happened. On October 30, 2026, researchers published a paper on arXiv titled BYOD: Build Your Own DLM, describing a parameter-efficient recipe that converts four autoregressive language models, Gemma 2 9B, Llama 3.1 8B, Qwen2.5 7B and Ministral 8B, into diffusion language models using only LoRA adapters while every base-model parameter stays frozen. Sources. arXiv preprint, BYOD: Build Your Own DLM, submitted October 30, 2026: https://arxiv.org/abs/2610.41574 (read 3 Nov 2026) Why it matters. For teams weighing whether to adopt diffusion generation, a recipe that touches only a small adapter rather than retraining a whole model changes the calculation from a multi-week pretraining budget to a few hours on one rented GPU, which lowers the bar for anyone who wants to test parallel decoding against an existing chat deployment rather than committing to it blind. The gap the paper reports at low step counts also tells a buyer exactly where the tradeoff still bites: faster generation comes with weaker open-ended text quality until more diffusion steps are spent, so a workflow built around short summaries or classification prompts has less to lose than one built around long free-form writing. https://aijourney.space/updates/2026-10-27-mpq-a-message-passing-view-of-post-training-quantization/ 2026-10-27-mpq-a-message-passing-view-of-post-training-quantization Tue, 27 Oct 2026 00:00:00 GMT What happened. On October 23, 2026, researchers posted a paper to arXiv describing Message-Passing Quantization (MPQ), a post-training quantization method that keeps soft posterior information across weight blocks instead of committing each block to a single codeword immediately, and tested it on dense language models from the Llama, Qwen and Mistral families. Sources. arXiv paper, MPQ: A Message-Passing View of Post-Training Quantization, submitted October 23, 2026: https://arxiv.org/abs/2610.32127 (read 27 Oct 2026) Why it matters. For any company running large models cheaply enough to answer millions of queries a day, a compression method that squeezes more accuracy out of the same bit budget changes the hardware bill rather than the headline feature list, and a reader comparing answer engines should ask whether the serving stack behind one has adopted a technique like this before assuming two products built on similarly sized models cost the same to run. https://aijourney.space/updates/2026-10-27-learning-structured-sparsity-for-large-models-bridging-model-compressibi/ 2026-10-27-learning-structured-sparsity-for-large-models-bridging-model-compressibi Tue, 27 Oct 2026 00:00:00 GMT What happened. On October 22, 2026, researchers from ETH Zurich and collaborators posted a paper on arXiv describing S-SALAAD, a training framework that jointly learns model weights alongside structured sparse and low rank components so a single checkpoint can be compressed to several memory budgets without the usual slowdown from unstructured sparsity. Sources. arXiv paper 2610.31897, Learning Structured Sparsity for Large Models, posted October 22, 2026: https://arxiv.org/abs/2610.31897 (read 27 Oct 2026) Why it matters. For a company selling a search box that answers in prose, where the product is the serving layer rather than a model, cutting inference latency at a fixed memory budget changes the bill per query and widens the hardware a local or edge deployment can run on without a separate checkpoint for every device tier. https://aijourney.space/updates/2026-10-27-conceptizer-next-token-prediction-is-not-all-you-need/ 2026-10-27-conceptizer-next-token-prediction-is-not-all-you-need Tue, 27 Oct 2026 00:00:00 GMT What happened. On October 11, 2026, researchers Haoran Zheng and Chen Shani posted a paper on arXiv titled Conceptizer: Next-Token Prediction Is Not All You Need, describing a training framework that adds a meaning-level reward alongside standard next-token prediction, tested across 3 model families at 1B-8B parameters. Sources. arXiv abstract page for Conceptizer: Next-Token Prediction Is Not All You Need, submitted October 11, 2026: https://arxiv.org/abs/2610.33027 (read 27 Oct 2026) Why it matters. For a reader who relies on Perplexity's answer engine, a training method that rewards meaning rather than exact wording could mean fewer oddly phrased or synonym mismatched answers if such a method were adopted, though the paper tests only smaller open research models and makes no claim about any shipped product. It is a reminder that the quality of an answer engine's prose depends on training choices made well before a query ever reaches the box. https://aijourney.space/updates/2026-10-14-edgellm-nano-auditing-polar-guided-key-cache-compression-and-consistency/ 2026-10-14-edgellm-nano-auditing-polar-guided-key-cache-compression-and-consistency Wed, 14 Oct 2026 00:00:00 GMT What happened. On October 10, 2026, researchers posted EdgeLLM-Nano to arXiv, a measurement study that audits KV-cache compression and weight pruning on a nano-scale 8.39M parameter, 33.6 MB decoder-only language model built as a testbed for on-device deployment, finding that a polar-guided key-cache quantizer and a consistency-anchored progressive pruner both underperform simpler baselines once the compressed cache is actually used during generation. Sources. arXiv paper on EdgeLLM-Nano, a KV-cache compression audit, submitted October 10, 2026: https://arxiv.org/abs/2610.14475 (read 14 Oct 2026) Why it matters. For anyone choosing a compression method before shipping a model onto a phone or a small board, the finding means the standard perplexity score can tell you every method is tied when they are not, so a team picking a key-cache quantizer off a leaderboard sorted by perplexity alone could ship the worse option without ever seeing the gap. It also means compressing the cache does not automatically save power at inference time, which changes the calculus for anyone trading memory for battery life. https://aijourney.space/updates/2026-10-14-next-concept-modeling-fast-parallel-coarse-to-fine-text-generation/ 2026-10-14-next-concept-modeling-fast-parallel-coarse-to-fine-text-generation Wed, 14 Oct 2026 00:00:00 GMT What happened. On October 9, 2026, researchers including Hermann Kumbong, Simran Arora, Azalia Mirhoseini and Christopher Re posted a paper to arXiv describing Next-Concept Modeling (NCM), a two stage framework that pairs a multiscale residual VQ-VAE encoder with a transformer that generates one level of codes at a time, predicting every code within a level in parallel rather than token by token. Sources. arXiv paper, Next-Concept Modeling: Fast, Parallel Coarse-to-Fine Text Generation, submitted October 9, 2026: https://arxiv.org/abs/2610.14162 (read 14 Oct 2026) Why it matters. For anyone weighing how a chat answer engine gets built underneath, the paper argues that the biggest brake on diffusion language models, namely degraded output once sampling steps are cut, can be loosened by generating coarse structure first and filling detail afterward, which changes the tradeoff between a cheaper inference bill and answer quality that a service like Perplexity has to balance on every query it serves. https://aijourney.space/updates/2026-10-12-byzantine-resilient-federated-large-language-model-via-hypergraph-messag/ 2026-10-12-byzantine-resilient-federated-large-language-model-via-hypergraph-messag Mon, 12 Oct 2026 00:00:00 GMT What happened. On October 8, 2026, researchers posted a paper to arXiv describing a defense against Byzantine attacks on federated large language models, where malicious model updates from participating devices or agents can corrupt the shared model's factual consistency during decentralized training without anyone touching the private data behind it. Sources. arXiv preprint 2610.12711, Byzantine-Resilient Federated Large Language Model via Hypergraph Message Passing, submitted October 8, 2026: https://arxiv.org/abs/2610.12711 (read 12 Oct 2026) Why it matters. For a service whose whole product is an answer engine that reads widely and writes back prose, a cheap way to launder false claims into a shared model during training is a direct threat to whether its citations can be trusted, so a defense like this one lands on the choice of how much decentralized training such products can safely accept before output quality is quietly degraded by poisoned contributors. https://aijourney.space/updates/2026-10-12-an-input-dependent-readout-phase-for-resonant-sequence-models/ 2026-10-12-an-input-dependent-readout-phase-for-resonant-sequence-models Mon, 12 Oct 2026 00:00:00 GMT What happened. On September 26, 2026, a paper titled An Input-Dependent Readout Phase for Resonant Sequence Models was submitted to arXiv, describing a change to ResonatorLM, a fixed-size state sequence model that reads a bank of damped complex oscillators through a learned phase per head. The authors make that readout phase a function of the current token rather than a fixed setting, which adds content dependence to the read step while keeping the O(T log T) convolutional training path exact and adding 0.2 percent more parameters. Sources. arXiv paper 2610.12496, An Input-Dependent Readout Phase for Resonant Sequence Models, submitted September 26, 2026: https://arxiv.org/abs/2610.12496 (read 12 Oct 2026) Why it matters. For anyone weighing a fixed-state sequence model against attention for long context work, the result changes the training budget calculus: a model trained on a shorter sequence no longer needs a costlier state transition just to stay usable once deployed at roughly four times that length, which lowers the retraining bill a team faces when it decides to extend a context window rather than rebuild the whole architecture. That tradeoff matters because it separates the cost of training from the cost of serving, letting a team keep the cheap convolutional path at inference while only paying the small parameter tax once, at training time. https://aijourney.space/updates/2026-10-09-tiny-scale-chinese-bert-pretraining-a-controlled-comparison-of-mlm-wwm-a/ 2026-10-09-tiny-scale-chinese-bert-pretraining-a-controlled-comparison-of-mlm-wwm-a Fri, 09 Oct 2026 00:00:00 GMT What happened. On October 6, 2026, a single-author paper posted to arXiv compared three pretraining strategies, MLM, WWM and MacBERT-style masking, on an 8.7 million parameter Chinese BERT model trained from scratch on 1.29 million Wikipedia sentences, measuring each with perplexity and four other intrinsic tests. Sources. arXiv paper 2610.08879, Tiny-Scale Chinese BERT Pretraining, submitted October 6, 2026: https://arxiv.org/abs/2610.08879 (read 9 Oct 2026) Why it matters. For teams choosing a pretraining recipe for models meant to run on phones or at the edge, this result argues against copying masking choices from systems ten times larger, since the ranking that holds at base scale reversed entirely once this model shrank under nine million parameters, so a workflow built on big model defaults could quietly end up picking the worst available option, raising the real cost of deployment chosen on reputation rather than measurement. https://aijourney.space/updates/2026-10-07-onconotebert-a-foundation-representation-model-for-natural-language-proc/ 2026-10-07-onconotebert-a-foundation-representation-model-for-natural-language-proc Wed, 07 Oct 2026 00:00:00 GMT What happened. On October 2, 2026, researchers posted a paper describing OncoNoteBERT, an oncology-specific BERT-style encoder trained from scratch on a governed UK outpatient corpus of 290,026 notes from 21,564 patients treated for lung and head-and-neck cancer, and compared it against RadBERT, PathologyBERT and a continued-pretraining variant called OncoNote-RadBERT. Sources. OncoNoteBERT paper on arXiv, submitted October 2, 2026: https://arxiv.org/abs/2610.03829 (read 7 Oct 2026) Why it matters. The result tells a hospital IT team shopping for a clinical note reader that borrowing a biomedical encoder off the shelf is a real gamble, since two existing tools scored badly on this corpus before any retraining happened, and building a bespoke tokenizer changed what the model got right in ways a loss curve alone did not show. For anyone weighing whether to buy a general answer engine or fund a narrower in-house model, the paper is evidence that the choice of vocabulary matters as much as the choice of architecture, and that a cheaper continued-pretraining route can win on one metric while a from-scratch model wins on another. https://aijourney.space/updates/2026-10-03-attention-manifolds-steering-or-blocking-language-models-by-editing-lear/ 2026-10-03-attention-manifolds-steering-or-blocking-language-models-by-editing-lear Sat, 03 Oct 2026 00:00:00 GMT What happened. On September 24, 2026, researcher Naveen Mysore posted a paper on arXiv describing attention manifolds, learned B-spline surfaces that modulate what each attention head extracts rather than only how much it attends, tested on LLaMA 3.2-1B-Instruct and 3B-Instruct. Sources. arXiv paper 2610.00257, Attention Manifolds, submitted September 24, 2026: https://arxiv.org/abs/2610.00257 (read 3 Oct 2026) Why it matters. For a product built around answering questions in prose rather than returning links, a cheap way to nudge a model away from a wrong or vague sentence without retraining it changes the calculus on how often an operator can afford to patch output quality before shipping an answer. https://aijourney.space/updates/2026-09-30-do-llms-understand-context-a-knowledge-graph-based-evaluation-framework/ 2026-09-30-do-llms-understand-context-a-knowledge-graph-based-evaluation-framework Wed, 30 Sep 2026 00:00:00 GMT What happened. On September 24, 2026, researchers posted a paper on arXiv proposing a knowledge graph based framework, called S3KG, to evaluate whether large language models understand context rather than simply pattern match, tested across nine question answering benchmarks. Sources. arXiv preprint, Do LLMs Understand Context? A Knowledge Graph-Based Evaluation Framework, posted September 24, 2026: https://arxiv.org/abs/2609.30484 (read 30 Sep 2026) Why it matters. For a reader who pays for an answer engine such as Perplexity's Pro plan, a metric that catches ungrounded triplet level errors offers a sharper way to judge whether a subscription's answers are trustworthy rather than merely fluent, since surface metrics like BLEU have historically hidden this gap. https://aijourney.space/updates/2026-09-22-tinycenn-lm-quality-gated-conversion-of-pretrained-attention-with-cenn-i/ 2026-09-22-tinycenn-lm-quality-gated-conversion-of-pretrained-attention-with-cenn-i Tue, 22 Sep 2026 00:00:00 GMT What happened. On September 17, 2026, researchers posted a paper on arXiv describing TinyCeNN-LM, a method that replaces attention layers in a pretrained language model with cellular-recurrent layers, but only after a quality gate checks that the swap does not damage the model's internal representations or its next-word accuracy, rejecting any layer that fails the check and keeping the original attention in place instead. Sources. arXiv paper 2609.21139, TinyCeNN-LM: Quality-Gated Conversion of Pretrained Attention with CeNN-Inspired Cellular-Recurrent Layers, submitted September 17, 2026: https://arxiv.org/abs/2609.21139 (read 22 Sep 2026) Why it matters. For anyone deciding whether to run a smaller local model instead of paying for a hosted service, this work is a reminder that cutting compute inside a language model usually means testing every layer separately rather than trusting a single global swap, so a team weighing a cheaper self-hosted setup against an API bill should budget time for that layer-by-layer validation work rather than assuming a published speedup number will hold on their own model. https://aijourney.space/updates/2026-09-16-layerroute-adaptive-layer-skipping-with-lora-preserved-quality-for-effic/ 2026-09-16-layerroute-adaptive-layer-skipping-with-lora-preserved-quality-for-effic Wed, 16 Sep 2026 00:00:00 GMT What happened. A paper titled LayerRoute, submitted to arXiv on September 12, 2026, describes a method for skipping transformer layers during inference without retraining a model from scratch. The author added a small per-layer router and LoRA adapters to Qwen2.5-0.5B-Instruct, a 24-block open model, then trained the router and adapters together so the router learns which layers a given input can skip. The paper reports results across ten independently seeded training runs rather than a single one, which is the detail that makes the claim checkable. Sources. LayerRoute paper on arXiv, submitted September 12, 2026: https://arxiv.org/abs/2609.13682 (read 16 Sep 2026) Why it matters. Inference cost is the daily bill for any product that answers questions by running a language model on every query, which is the business Perplexity AI is in. A router that reliably skips a fixed block of middle layers without hurting output quality points at a cheaper way to serve the same model, though this result comes from a 0.5-billion-parameter model and has not been shown at the scale production answer engines actually run. https://aijourney.space/updates/2026-09-16-nepali-legal-expertise-through-generative-and-extractive-pre-trained-tra/ 2026-09-16-nepali-legal-expertise-through-generative-and-extractive-pre-trained-tra Wed, 16 Sep 2026 00:00:00 GMT What happened. A paper submitted on 8 August 2026 and indexed on arXiv on 16 September 2026 describes NepLEGiT, a small language model pre-trained from scratch on a curated corpus of Nepali legal text covering constitutional law, civil and criminal codes, and administrative regulations. The authors, working independently of any major AI lab, built a decoder based transformer with a 6 layer, 6 head, 384 dimensional architecture and trained it with warmup cosine decay scheduling, gradient accumulation, and mixed precision arithmetic. Sources. arXiv paper abstract page, NepLEGiT (submitted 8 August 2026, this version 16 September 2026): https://arxiv.org/abs/2609.16010 (read 16 Sep 2026) Why it matters. The work is a reminder that useful language coverage does not always require a frontier lab or a general purpose model. Answer engines such as Perplexity route queries to whichever model or index answers best, and a narrow, well curated legal model in an underserved language is the kind of component that could plug into that routing rather than compete with it, since a small model trained on domain text can outperform a general model on that domain without the cost of a large pretraining run. https://aijourney.space/updates/2026-09-16-the-functionalizer-lossless-functional-decomposition-for-subword-tokeniz/ 2026-09-16-the-functionalizer-lossless-functional-decomposition-for-subword-tokeniz Wed, 16 Sep 2026 00:00:00 GMT What happened. Researchers Connor Makowski and Willem Guter posted the Functionalizer paper to arXiv, submitted July 7, 2026, describing a lossless pre-tokenizer that separates casing, diacritics and repeated characters from a word's base token using reversible opcodes. Sources. arXiv preprint, The Functionalizer paper: https://arxiv.org/abs/2609.15991 (read 16 Sep 2026) Why it matters. For any company running an answer engine such as Perplexity, tokenizer vocabulary size directly shapes model cost and context budget, so a scheme claiming smaller vocabularies without losing text coverage is a lever worth testing before it reaches production.