arXiv:2606.27459v1 Announce Type: new Abstract: We consider the cubic nonlinear Schr\"odinger (NLS) equation on two-dimensional flat tori with varying aspect ratios. In this formulation, the choice of aspect ratio governs the Fourier resonance structure, so rational and irrational geometries can exhibit different high-frequency cascade behaviors. We present a geometry-conditioned Fourier neural operator (FNO) for the cubic defocusing NLS equation, where the input consists of the real and imaginary parts of the solution together with the aspect-ratio parameter \(\omega^2\). The model is trained to approximate the one-step solution operator and is evaluated on unseen trajectories generated from random-phase initial data using Fourier pseudospectral method. Our numerical experiments show that the learned operator captures the main solution dynamics on both tori and reproduces the distinct Sobolev norm behavior of the two geometries, with stronger \(H^2\)-growth on the rational torus and more constrained behavior on the irrational torus, consistent with the findings of \cite{hrabski2021energy}. We perform ablation studies to examine the roles of retained Fourier modes, activation functions, Fourier-layer depth, and explicit geometry conditioning. The results indicate that including $\omega^2$ improves long-time predictive accuracy, especially for the rational geometry, and supports the use of geometry-aware neural operators for learning spectral-transfer phenomena in nonlinear dispersive partial differential equations.

arXiv:2606.27510v1 Announce Type: new Abstract: Activation patching is the primary tool in mechanistic interpretability. It attributes causal responsibility for a model behavior to each of its individual components by estimating its natural indirect effect (NIE). Re-deriving the activation patching estimand from causal mediation analysis, we find that the NIE does not solely capture the causal effect through the specific component. It also contains interaction effects (INT) that measure how much the component's causal effect itself depends on the state of other components in the model. A natural response may be to try to eliminate INT by adjusting the estimator or unit of analysis, but each of these potential remedies has predictable failure modes. We demonstrate these failure modes in the GPT-2 IOI circuit; components whose causal importance is conditional on the state of other components are either invisible or artificially inflated, and INT variance explains the previously documented instability of faithfulness scores. We prove that INT scales with the distance between clean and patched component activations, is negligible when the model is locally affine, and decomposes combinatorially into pairwise and higher-order group interactions. Despite its inevitability, INT is not a nuisance to be eliminated, but rather a diagnostic for interpretability studies. Its individual and group-level magnitude and sign signal when causal conclusions are prompt-dependent, and when greedy NIE-based component ranking will miss mechanisms only discoverable through combinatorial search.

arXiv:2606.27378v1 Announce Type: new Abstract: We introduce an axiomatic evaluation framework for latent thought representations in LLMs, comprising metrics that are independent of downstream benchmark scores and reveal representational failures that benchmark accuracy masks. Existing evaluations conflate representation quality with model capacity. Therefore, failures cannot be attributed to the representation rather than to the model that processes it. We formalize four functional axioms (Causality, Minimality, Separability, and Stability) and define a quantitative measure for each, computed directly on the representation independently of downstream accuracy. We audit open-weight LLMs across 23 reasoning tasks (e.g., Spatial Reasoning, Factual QA). We find that no candidate satisfies all four axioms simultaneously, that the representations distinguish task type reliably but cannot distinguish between two questions within the same task, and that the representations encode little information beyond what is already present in the input embedding. The failure is consistent across dense, reasoning-distilled, and RL-trained model families, indicating that the gap is structural rather than a property of model size or training procedure.

arXiv:2606.27448v1 Announce Type: new Abstract: This paper studies the problem of regret minimization in Markovian bandits with \emph{non-observable states} and possibly \emph{constrained} decision epochs. The focus is restricted to a ``pure'' regret benchmark, that compares the performance of the learning algorithm to the best \emph{pure policy} which -- akin to optimal policies of stochastic bandits -- picks the optimal arm from start to finish without ever switching. We introduce a generalization of rested Markovian bandits, \emph{self-degrading Markovian bandits}, for which pure policies are always asymptotically optimal.We show that without prior knowledge on the underlying bandit, the regret of algorithms that switch arms rarely necessarily scales super-logarithmically for every bandit, i.e., as $\omega(\log(T))$, where $T$ is the learning horizon. Despite the unreachability of the logarithmic regime, we design UCB-NOM, an optimistic algorithm inspired by UCB, of which the regret is nearly logarithmic. Lastly, we show that given prior knowledge on the Markovian bandit in the form of a bound on the bias functions of its arm, a proper instantiation of UCB-NOM achieves $O(\log(T))$ regret. We further show that this prior knowledge allows for a $O(\sqrt{T \log(T)})$ worst-case regret bound for UCB-NOM. Notably, our regret bounds do not depend on the number of states of the underlying Markov chains. Our findings suggest that the non-observability of states is a mild inconvenience in self-degrading Markovian bandits.

arXiv:2606.27440v1 Announce Type: new Abstract: Foundation models for structural biology have achieved remarkable performance in predicting biomolecular structure and show promise for the design of proteins and small molecules. Yet understanding which internal features drive their outputs remains challenging. Standard sparse autoencoders (SAEs), effective on transformer-style sequence embeddings, do not transfer cleanly to pairformer-like architectures: naively operating on pairwise representations yields a quadratic blow-up of features and obscures concepts distributed jointly across sequence and pair representations. We introduce PairSAE, which summarizes pairwise tensors via an N-mode SVD into token-wise interaction roles, then uses a sparse autoencoder to learn a shared set of token-level features that decode into both sequence and pair representations. Evaluated on Boltz-2 activations for PLINDER protein-ligand complexes, PairSAE yields interpretable features that align with UniProt annotations and predict Boltz-2 affinity values. These results indicate that PairSAE links the latent space of foundation models for structural biology to interpretable structural concepts, clarifying what the model "knows" while avoiding pairformer-induced pitfalls that limit conventional SAEs.

arXiv:2606.27379v1 Announce Type: new Abstract: Large language models increasingly face demands to "forget" training data, knowledge, or behaviors due to regulatory deletion obligations, copyright/licensing disputes, and safety or product-policy requirements. This position paper argues that machine unlearning is overused as a term in LLM research and should be reserved for dataset-defined deletion: removing the training influence of a precisely specified forget set such that the resulting model is approximately indistinguishable from retraining without that data. We contend that many tasks currently labeled "unlearning" (e.g., refusal for harmful requests, entity/knowledge removal, or targeted suppression) pursue different, often policy-dependent objectives and therefore require different terminology and baselines (e.g., alignment, suppression, editing, obfuscation). We further argue that this confusion is not cosmetic: because papers make different implicit guarantees under the same label, metrics and benchmarks are frequently reused outside their intended scope, rewarding surface-level non-disclosure (e.g., low ROUGE/forget accuracy) even when retraining-equivalence is not tested and derived capabilities remain. We conclude by calling for stricter terminology tied to explicit guarantees and reference models, and for evaluations that match the claimed objective.

arXiv:2606.27380v1 Announce Type: new Abstract: Automated coaching for oral presentations sits at the intersection of computer-assisted pronunciation training (CAPT), prosody modeling, and speech synthesis, yet no prior work has systematically surveyed and compared existing systems along these dimensions. This survey reviews and categorizes automated presentation coaching systems, spanning pronunciation tutors, fluency and prosody coaches, multimodal trainers, and conference Q&A practice tools. We introduce a five-dimensional task taxonomy - covering segmental pronunciation, lexical stress, suprasegmental prosody, pacing, and content faithfulness - and explicitly map surveyed systems onto it to reveal coverage gaps. We further review the core technical methods these systems employ: TTS-based exemplar generation and diagnostic methods for pronunciation, prosody, and fluency assessment. Key open challenges include the scarcity of annotated presentation corpora, achieving accent-fair feedback across diverse L1 backgrounds, and delivering low-latency diagnostics for real-time rehearsal.

arXiv:2606.27550v1 Announce Type: new Abstract: Multi-token prediction has been shown to increase data density during training, improve downstream text-generation quality, and serves as the defacto approach for self-speculative decoding. Existing foundation and open source models that use MTP heads commit to a static tree-based attention topology throughout the entire generation sequence, meaning the speculation depth, and thus the compute required during verification, stays constant regardless of the context. This is fundamentally misaligned with the entropy patterns of natural language where low-entropy regions often support reliable multi-step drafting, while high-entropy regions require more conservative speculation. To address this, we propose Entropy-guided Multi-Token Prediction (EntMTP), a training-free scheduler that toggles between tree-based attention topologies from a set of task-specific pareto-optimal trees conditioned on a running estimate of local generation entropy. By matching speculation depth to context predictability, EntMTP maximizes expected accepted-token throughput across the full distribution of generated text without sacrificing generation quality. When evaluated across Humaneval, ShareGPT, GSM8k, and Litbench benchmarks, EntMTP consistently achieves a 1.15x speedup against Hydra and peak speedup of 1.36x against Medusa baselines respectively.

arXiv:2606.27446v1 Announce Type: new Abstract: This paper describes team HSA_CORAL's submission to the FinCausal 2026 shared task on extracting cause-effect relations from financial narratives via extractive question answering in English and Spanish. We compare three modeling families: (i) encoder-only token tagging with multilingual BERT, (ii) encoder-decoder generation with multilingual BART, and (iii) decoder-only LLMs (Llama 3.1 and GPT variants) using prompt refinement, few-shot demonstrations, and supervised fine-tuning. Across settings, prompting and few-shot examples yield competitive performance, while supervised fine-tuning provides the largest gains. Our best system, GPT-4.1 Mini fine-tuned on combined English and Spanish training data, achieves a tied highest score on the English subtask (score 4.8140) and ranks third on Spanish (score 4.7753) under the shared task's LLM-as-a-judge metric. Overall, the results highlight the value of task-specific adaptation and multilingual fine-tuning for cross-lingual transfer in financial causality QA.

arXiv:2606.27438v1 Announce Type: new Abstract: Since its initial release in 2020, Darts has become a widely used open-source Python library for time series analysis. A series of foundation models have recently claimed accuracy improvements in zero-shot forecasting, promising a paradigm shift from training custom models to harnessing pre-trained general-purpose forecasters. Foundation models, however, are often released as isolated packages with fragmented interfaces and limited interoperability with common tooling, making joint evaluation and integration within complete pipelines difficult. In Darts, we developed a unified $\texttt{FoundationModel}$ class collection (Chronos-2, TimesFM 2.5, TiRex, PatchTST-FM) that provides standardized, full-cycle forecasting interfaces with minimal external dependencies for integrating foundation models into the ecosystem. Existing Darts pipelines can now use foundation models with only a name change; new pipelines can use them for zero-shot or fine-tuned forecasting, uncertainty estimation, and backtesting, combined with data processing and evaluation tooling, all within a unified framework.

arXiv:2606.27412v1 Announce Type: new Abstract: 3D Scene Graph Generation (3DSGG) represents 3D scenes as structured object-relation-object graphs, providing a compact relational abstraction for spatial understanding. In embodied intelligence settings, the same 3D scene may be observed by agents from viewpoints that differ by yaw rotations. However, current 3DSGG models often fail to produce relation predictions that follow the expected transformation behavior under such viewpoint shifts. This behavior reveals an empirical mismatch related to predicate-level transformation heterogeneity: directional predicates such as left, front, right, and behind should transform with the observation frame, whereas most contact, support, and semantic predicates such as standing on and attached to should remain stable. To reduce this mismatch, we propose Transformation-Aware Decoupling (TAD), a viewpoint-robust 3DSGG framework that decouples relation reasoning according to predicate transformation behavior and is supported by viewpoint-stable object representations. TAD decomposes relation reasoning into two parts: one learns cues that should stay stable across viewpoints, while the other learns directional cues that should change with the observation frame. The two parts are merged for standard multi-label predicate prediction. Transformation-specific descriptors and group-aware auxiliary supervision encourage the two branches to capture complementary relation cues. Extensive experiments on 3DSSG show that TAD achieves state-of-the-art robustness under yaw viewpoint changes without training-time rotation augmentation, while maintaining competitive performance under the standard benchmark. The project page is available at https://tad-predicate.github.io/.

Reuters: Shanghai-based Momenta, whose driver-assist tech is used by Toyota, Mercedes, and Audi in China, files for an IPO in Hong Kong, seeking to raise up to $751.1M  —  Chinese self-driving technology firm Momenta Global launched its Hong Kong initial public offering on Monday …

Emma D'Arcy as Rhaenyra Targaryen in "House of the Dragon" season three. Ollie Upton/HBO Rhaenyra Targaryen has finally claimed the Iron Throne in "House of the Dragon" season three. "Game of Thrones" fans may remember that Rhaenyra's death is mentioned in the show's third season. Here's how Rhaenyra dies in the original book, "Fire & Blood," and possibly "House of the Dragon." Warning: Major spoilers for the book "Fire & Blood" and "House of the Dragon" season three. In the "House of the Dragon" season two finale, Prince Daemon has a vision of his wife, Rhaenyra Targaryen, on the Iron Throne. Daemon comes to believe that Rhaenyra is the chosen one, the "Prince That Was Promised," who will unite the realm against darkness, a belief that motivates Rhaenyra herself to seize power in season three, even through grief, treason, and doubt. But her story won't be so triumphant if the series follows its source material. Although many fans will be thrilled to see Queen Rhaenyra finally materialize after she was named her father's heir all the way back in season one, actor Emma D'Arcy said the show's third season explores "a movement towards tyranny." "I think in Rhaenyra's case," D'Arcy told Entertainment Weekly, "it's a kind of religious fanaticism, actually, that starts to radicalize her position." Meanwhile, Rhaenyra's half-brothers — the primary challengers to her crown — are still at large: Aegon II (Tom Glynn-Carney), aka Aegon the Usurper, quietly fled King's Landing after he was critically injured in battle at Rook's Rest, while Aemond (Ewan Mitchell) has flown to the Riverlands with his massive dragon Vhagar to claim Harrenhal castle. Here's what happens to Rhaenyra in the original book, George R. R. Martin's "Fire & Blood." Rhaenyra becomes Queen of Westeros, but both the throne and the smallfolk reject her Emma D'Arcy as Rhaenyra Targaryen in "House of the Dragon." Ollie Upton/HBO As in season three of "House of the Dragon," Rhaenyra presses her advantage once Aemond sets forth to retake Harrenhal, leaving King's Landing unprotected. Only one dragonrider remains in the city: Queen Helaena, bonded to Dreamfyre, described in the book as a "broken woman" who's unable to fight or fly. (In the show, she's perfectly lucid but refuses to use her dragon to burn people.) Rhaenyra and Daemon fly to King's Landing on their own dragons, Syrax and Caraxes, flanked by her newly recruited dragonriders. They find the knights of the City Watch have remained loyal to Daemon, so they're easily able to coax a surrender from the Dowager Queen Alicent. "The sight of the queen's dragons in the sky above took the heart out of the opposition, and King Aegon's remaining loyalists hid or fled or bent the knee," the book reads, although Aegon himself is nowhere to be found when Rhaenyra arrives in the Red Keep. But Aegon's escape is only the beginning of Queen Rhaenyra's problems. When she finally ascends her father's throne, the ancient seat made of melted swords, it cuts her legs and hands. Onlookers watch as blood drips from her skin. In Westerosi lore, this is considered an ill omen, as if the throne itself is rejecting her claim. Soon enough, the smallfolk turn on Rhaenyra. She is forced to drastically raise taxes, since three-quarters of the crown's gold had been shipped away by Aegon's master of coin (and Aegon spent whatever was left). There isn't enough food in the city, either, so the criminals of King's Landing are publicly executed. Their heads are displayed on spikes above the city gates, and the rest of their corpses are fed to the queen's dragons. People begin spreading scandalous rumors about Rhaenyra's cruelty, and some even come to see dragons as evil, unnatural creatures, not as symbols of House Targaryen's power and divinity. Eventually, a series of disasters and betrayals brings a premature end to Rhaenyra's reign. Hugh Hammer and Ulf White, two of the dragonseeds that Rhaenyra recruited for her army, double-cross the queen and scheme to seize power for themselves. Daemon dies fighting Aemond. Helaena, beloved by the smallfolk, dies by suicide. King's Landing erupts in riots, leading to the deaths of multiple dragons (including the queen's own dragon, Syrax) and Rhaenyra's thirdborn son, Prince Joffrey. Dragonless and vulnerable, Rhaenyra has no choice but to flee. She uses the last of her resources to escape the city with her last living son, Aegon the Younger, and retreat to Dragonstone, her former stronghold. (Her fifth and youngest son, Viserys II, is presumed dead after the Battle of the Gullet.) Rhaenyra returns to Dragonstone, where she is eaten by her brother's dragon Tom Glynn-Carney as Aegon Targaryen in "House of the Dragon." Theo Whiteman/HBO When Rhaenyra reaches Dragonstone, her brother Aegon is already there. Aegon's dragon, Sunfyre, had also found his way back to the island after recuperating in the woods next to Rook's Rest. Aegon and Sunfyre had taken control of Dragonstone and slain Rhaenyra's supporters, though both were greatly weakened in the process. The book says that by the time Rhaenyra returned, Sunfyre was on the brink of death. "Dear brother," Rhaenyra says when she sees Aegon's scarred face. "I had hoped that you were dead." "After you," Aegon answers. "You are the elder." Sealing his victory, Aegon feeds Rhaenyra to his dragon while her son, Aegon the Younger, watches in horror. Rhaenyra's fate is revealed in 'Game of Thrones' season three Rhaenyra's descendant, Daenerys Targaryen, appears in Daemon's vision in "House of the Dragon." HBO "House of the Dragon" is a prequel to "Game of Thrones," set about 170 years before the main series begins. That means "Game of Thrones" fans already know a few spoilers. By the time we meet famous characters like Ned Stark, Cersei Lannister, and King Robert Baratheon, dragons have been extinct for generations; House Targaryen has recently been ousted from power, and there are only a few members of the family left alive. Some fans may even remember a more specific spoiler, courtesy of Cersei's eldest son, Joffrey Lannister. In "Game of Thrones" season three, episode four, "And Now His Watch Is Ended," Joffrey is giving his future queen, Margaery Tyrell, a tour of the Sept of Baelor in King's Landing. "Rhaenyra Targaryen was murdered by her brother, or rather, his dragon," Joffrey recalls. "It ate her while her son watched. What's left of her is buried in the crypts right down there." Ayomikun Adekaiyero contributed to a previous version of this story. Read the original article on Business Insider

Emma D'Arcy and Rhys Ifans as Rhaenyra Targaryen and Otto Hightower in "House of the Dragon." Ollie Upton/HBO HBO's "House of the Dragon" is adapted from "Fire & Blood," a fictional history of House Targaryen. An early death in season three is notably different from its vague description in the book. In earlier seasons, the deaths of characters like Laenor and Rhea were portrayed in unexpected ways. Warning: Major spoilers through "House of the Dragon" season three, episode two. "House of the Dragon" has returned for its third season, continuing its tradition of expanding on the source material. HBO's flagship "Game of Thrones" spinoff is based on George R. R. Martin's "Fire & Blood," a fictional history of House Targaryen. The book is written as an overview of major events during the dynasty's reign. Unlike Martin's main series ("A Song of Ice and Fire," which inspired "Game of Thrones"), "Fire & Blood" is not told from the major characters' points of view but is recounted in retrospect by maesters, members of the royal court, and other third-party observers. Thus, "Fire & Blood" readers don't get definitive causes or clear motivations for why characters take certain actions — and, in some cases, the book may be missing key in-universe information. As the "Fire & Blood" narrator frequently notes, not even eyewitness accounts are entirely reliable. His sources often disagree, each saddled with unique baggage, blind spots, and biases. This means the TV writers hired for HBO's adaptation had a lot more leeway with the story and an opportunity to create a consistent canon of events. So let's take a closer look at some of the pivotal "House of the Dragon" deaths so far, and how they compare to the details found in "Fire & Blood." Aemma's final moments of life are made more brutal by Viserys opting for a C-section. Paddy Considine and Sian Brooke as King Viserys and Queen Aemma in "House of the Dragon" season one. Ollie Upton/HBO In the show: Director Miguel Sapochnik says the scene that killed Queen Aemma (Sian Brooke) was designed to show how King Viserys (Paddy Considine) "effectively murdered" his beloved wife. When Aemma goes into labor, her baby is facing the wrong direction, and the maesters can't get it to turn. This means Aemma will certainly die, but there's a small chance they can save the baby by cutting it out of her. The maester offers the choice to Viserys, who picks the surgery without consulting Aemma or telling her what's about to happen. Instead, she's dragged down on the bed and restrained while the maester cuts open her belly. She bleeds to death as the baby is "born." In the book: Death by childbirth is fairly commonplace in Westeros, and it's treated as such in the text. "Fire & Blood" provides no details or context to Aemma's death, and doesn't mention Viserys being a part of the labor and delivery process at all: "Queen Aemma was brought to bed in Maegor's Holdfast and died whilst giving birth to the son that Viserys Targaryen had desired for so long. The boy (named Baelon, after the king's father) survived her only by a day, leaving king and court bereft." Rhea's death in the book is regarded as a "hunting accident" instead of cold-blooded murder. Daemon and Rhea in "House of the Dragon." HBO In the show: In season one, Prince Daemon Targaryen (Matt Smith) is married to Lady Rhea Royce (Rachel Redford), though neither is particularly happy about it. Hoping to marry his royal niece instead, Daemon sneaks up on Rhea during a hunting trip and spooks her horse. After she's thrown, Rhea is paralyzed from the neck down and unable to fight back. She lobs insults at Daemon until he picks up a rock and bashes in her skull. In the book: According to the historical record, Rhea's death was "a tragic mishap." Daemon is not suspected of being involved: "The 'bronze bitch' of Runestone, Lady Rhea Royce, fell from her horse whilst hawking and cracked her skull upon a stone. She lingered for nine days before finally feeling well enough to leave her bed…only to collapse and die within an hour of rising." Joffrey is killed by Criston in both the show and the book, but in "Fire & Blood," his death happens under the guise of a tourney. Solly McLeod and Fabien Frankel as Joffrey Lonmouth and Criston Cole in "House of the Dragon" season one. HBO In the show: While still teenagers, Lord Laenor Velaryon (Theo Nate) and Princess Rhaenyra Targaryen (Milly Alcock) agree to get married — not out of love, but duty to their families. They also secretly agree to keep their marriage open, since Laenor is gay, and Rhaenyra is having an affair with Ser Criston Cole (Fabien Frankel). Ser Joffrey Lonmouth (Solly McLeod), Laenor's lover, outs himself to Criston as the future king consort's boyfriend. Big mistake. Cole, embarrassed and infuriated by the exposure of his own secret tryst with Rhaenyra, lashes out and beats Joffrey to death in the middle of the wedding welcome feast. In the book: "Fire & Blood" historians recall that Joffrey was killed by Criston during a tourney, during which Criston had been "fighting in a black fury" against all his opponents. Joffrey's relationship with Laenor isn't mentioned as a possible motivation for his death: "Cole's favorite weapon was the morningstar, and the blows he rained down on Ser Laenor's champion cracked his helm and left him senseless in the mud. Borne bloody from the field, Ser Joffrey died without recovering consciousness six days later." Harwin and Lyonel canonically die in a fire at Harrenhal, but the show clarifies the arsonist. Ryan Corr as Ser Harwin "Breakbones" Strong in the sixth episode of "House of the Dragon." HBO In the show: House Strong controls Harrenhal, the largest castle in the Riverlands. It's where Ser Harwin Strong (Ryan Corr) and his father, Lyonel Strong (Gavin Spokes), meet their doom after leaving King's Landing in season one. "House of the Dragon" makes it clear who to blame: Larys Strong (Matthew Needham), a sneaky man who wants leverage on Queen Alicent (Olivia Cooke), hires two miscreants to set the castle ablaze in one of the show's most egregious acts of kinslaying. As an added bonus, with his older brother and father both dead, Larys inherits Harrenhal. In the book: Harwin and Lyonel are indeed killed in a mysterious fire at Harrenhal, but Larys is floated as just one of several suspects for the arson: "Shortly after their arrival, a fire broke out in the tower where they were sleeping, and both father and son were killed, along with three of their retainers and a dozen servants. The cause of the fire was never determined. Some put it down to simple mischance, whilst others muttered that Black Harren's seat was cursed and brought only doom to any man who held it. Many suspected the blaze was set intentionally. Mushroom suggests that the Sea Snake was behind it, as an act of vengeance against the man who had cuckolded his son. Septon Eustace, more plausibly, suspects Prince Daemon, removing a rival for Princess Rhaenyra's affections. Others have put forth the notion that Larys Clubfoot might have been responsible; with his father and elder brother dead, Larys Strong became the Lord of Harrenhal. The most disturbing possibility was advanced by none other than Grand Maester Mellos, who muses that the king himself might have given the command. If Viserys had come to accept that the rumors about the parentage of Rhaenyra's children were true, he might well have wished to remove the man who had dishonored his daughter, lest he somehow reveal the bastardy of her sons. Were that so, Lyonel Strong's death was an unfortunate accident, for his lordship's decision to see his son back to Harrenhal had been unforeseen." Like Aemma, Laena's cause of death isn't detailed in the book beyond childbirth. The show gives her a fiery end. Matt Smith and Nanna Blondell as Daemon Targaryen and Laena Velaryon in season one, episode six of "House of the Dragon." Ollie Upton/HBO In the show: Just like Aemma, Lady Laena Velaryon (Nanna Blondell) is unable to safely deliver her baby. A medieval C-section offers a tiny chance for the baby's survival, and Laena's husband, Daemon, is given that option. But Laena takes her fate into her own hands and opts for a dragonrider's death. While still in labor, she orders her dragon, Vhagar, to incinerate her. In the book: Laena officially dies of "childbed fever" and apparently doesn't make it to Vhagar in time: "During her final hour, it is said, Lady Laena rose from her bed, pushed away the septas praying over her, and made her way from her room, intent on reaching Vhagar that she might fly one last time before she died. Her strength failed her on the tower steps, however, and it was there she collapsed and died." Laenor fakes his own death in a big change from the book. Laenor's death in "House of the Dragon" is a ruse. Ollie Upton/HBO In the show: Even after marrying and having children with other people, Daemon and Princess Rhaenyra (Emma D'Arcy) are drawn to each other. They want to make it official as husband and wife, but Rhaenyra is trapped in a lavender marriage with Laenor (John MacMillan). Daemon and Rhaenyra seem to conspire to kill Laenor so they can finally be together — and that's what Laenor's mother long believed — but the show reveals that Laenor was in cahoots with the couple all along. Daemon kills a High Tide castle servant while Laenor and his lover, Qarl, stage a sword fight in the main hall. After another servant runs to get help, Daemon presumably helps Laenor dress the body of the dead servant in his own clothes, and they burn the servant's face beyond recognition. Laenor and Qarl sneak away to Essos with Daemon's help, leaving the realm to gossip about the tragic and fishy death of Rhaenyra's husband. In the book: Qarl stabs Laenor in broad daylight. Rumors spread that it was an act of passionate jealousy, or maybe a murder plot, but no damning evidence was ever collected: "Ser Laenor Velaryon, husband to the Princess Rhaenyra and the putative father of her children, was slain whilst attending a fair in Spicetown, stabbed to death by his friend and companion Ser Qarl Correy. The two men had been quarreling loudly before blades were drawn, merchants at the fair told Lord Velaryon when he came to collect his son's body. Correy had fled by then, wounding several men who tried to hinder him. Some claimed a ship had been waiting for him offshore. He was never seen again." In the book, Otto Hightower is executed as a traitor. The show makes his death more climactic by having Rhaenyra swing the sword. Rhys Ifans in "House of the Dragon." Ollie Upton / HBO In the show: After he's unceremoniously removed as Hand of the King by his grandson, King Aegon II Targaryen (Tom Glynn-Carney), viewers barely see Otto Hightower (Rhys Ifans) until his death. Otto tells his daughter, Alicent, that he's returning to House Hightower's seat in Oldtown, though she finds it odd that he doesn't answer any of her letters. Upon Rhaenyra's return to King's Landing, we discover that Otto never made it home. He was captured and imprisoned in the Red Keep's dungeon by none other than Larys, Aegon's master of whisperers. Otto is brought before Rhaenyra, and Daemon urges her to execute Aegon's former Hand as a traitor. "If you wish to rule," he says, "show them you do not waver." Unfortunately, Rhaenyra was raised to be a lady, not a knight. She has never beheaded anyone — let alone the father of her childhood best friend. She panics and misses Otto's neck on her first stroke, burying her sword in his shoulder blades. On her second stroke, Rhaenyra cuts off Otto's head, choking back sobs. In the book: When Rhaenyra takes the Iron Throne in "Fire & Blood," it's said that she set about "rewarding her friends and inflicting savage punishments on those who had served her half-brother." Otto is simply listed as the first traitor to die under Rhaenyra's rule, sans any grisly or personal details: "Queen Alicent was fettered at wrist and ankle with golden chains, though her stepdaughter spared her life 'for the sake of our father, who loved you once.' Her own father was less fortunate. Ser Otto Hightower, who had served three kings as Hand, was the first traitor to be beheaded." Read the original article on Business Insider

In this article, we are fine-tuning Gemma 4 for text on a reasoning dataset whose responses have been collected from DeepSeek-V4 Flash model. The post Gemma 4 Text Fine-Tuning appeared first on DebuggerCafe.

Grace Cong Sui and her family live in California; for Lunar New Year, she and her daughter traveled to China to visit relatives. Provided by Grace Cong Sui Grace Cong Sui is raising her daughter in Los Angeles without family nearby. During a two-month visit to Qingdao, China, she realized her family's parenting style differed from her own. The experience helped her continue to find her own parenting path between two cultures. My aunt chased her granddaughter around the living room with a spoon and a bowl of rice and vegetables. My niece's eyes stayed fixed on the TV as her mouth opened and closed automatically while my aunt slipped in bite after bite. My daughter sat beside them, watching quietly. I needed to say something. "You shouldn't feed her like that," I said. "It's not good for her." Before my aunt could respond, my mother shot me a look from across the room — the kind many Chinese mothers can give without saying a word. I fell silent. After nearly a decade abroad, this was the first time I'd brought my daughter back to China. We were in Qingdao, about halfway between Beijing and Shanghai, to celebrate Lunar New Year and spend time with my extended family. Cousins got to spend time together in China. Provided by Grace Cong Sui My family was thrilled to be together again I was relieved to step away from the constant demands of parenting, with no extended family nearby. For the first time in years, I wandered around the city streets on my own, caught up with old college friends, and sat through dinner without reminding anyone to chew. My parents happily took over with my daughter, a luxury my husband and I don't have in the US. Watching them with her transported me back to my own childhood in Shandong Province in the 1990s. I was raised by a village of relatives. My grandparents took turns living with us, someone was always waiting at the school gate, and every school holiday I stayed with them so my parents could focus on their careers. Her daughter spent two months in a Chinese school. Provided by Grace Cong Sui Parenting in the US I became a mother 3 years ago, thousands of miles away in Los Angeles. With no family nearby, I leaned heavily on parenting books, expert advice, and plenty of trial and error. Over time, I found my own approach. Once my daughter turned 2, I encouraged her to feed herself. I let her focus uninterrupted during independent play and reading, responded to tantrums with calm explanations instead of distractions, and limited her sugar intake. Help from my family in China We both loved being surrounded by family. They were always ready to help. I slipped easily into the rhythm of having extra hands around. Before I noticed my daughter was cold, someone had already found her another layer. Before I realized she needed bedding for kindergarten — a local Chinese school she attended while we were there — my mother had packed everything and set it by the door. But before long, I started feeling strangely displaced as a mother. "It's only ice cream," my mother would say whenever I tried to explain to my daughter why she shouldn't eat too many sweets. During quiet playtime, my father often interrupted to offer a puzzle or fruit. At meals, my mother encouraged "just one more bite," often feeding her herself. Whenever my daughter argued with my niece over a toy, my mother quickly redirected her attention instead of letting her work through her emotions. It wasn't how I had learned to parent. As a child in China, I never questioned those habits. My grandparents raised me the same way. Living in America changed how I think about parenting. My husband and I carefully established routines around sleep, meals, screen time, and the boundaries we wanted to set. Back in China, those routines gradually disappeared. My family cared for my daughter with love, but their way of helping often left me feeling as though my role as her mother had quietly slipped into the background. My daughter enjoyed the attention My daughter seemed happier than ever in China. At family gatherings, she bounced from lap to lap, proudly shouted "Xin Nian Kuai Le" ("Happy New Year") to relatives, and ate more than usual while competing with her cousin at mealtimes. Watching her, I felt both overwhelmed and deeply comforted. Back in the US, life quickly returned to normal. Once again, it was just my husband and me. She says her daughter quickly adapted to life in America after returning from the trip. Provided by Grace Cong Sui One evening, I placed dinner in front of her as usual. Instead of reaching for her spoon, she sat quietly. "Mommy, feed me," she said. I pointed to her utensils. "You can do it yourself." "No," she said more firmly. "Please feed me." Without thinking, I picked up my spoon. As I fed her, I caught myself doing the very thing I'd spent weeks resisting in China. Part of it was exhaustion. Part of it was the familiar worry that she wouldn't eat enough if I didn't help. In that moment, I realized how easily old habits can resurface. I had thought I'd left that style of parenting behind, but it was still part of me, shaped by the family who had raised me. My daughter still asks when we can go back to China to see her grandparents. She misses the constant attention and the feeling of being surrounded by family. I still believe my husband and I should make the decisions about how we raise her. But I no longer see my parents' approach as simply something to resist. It came from love, just as mine does. Now I'm trying to figure out what parts of each I want to pass on to my daughter as we navigate motherhood — and childhood — together. Read the original article on Business Insider

Lana Ng quit her job to become a stay-at-home mother in 2023. Lana Ng Lana Ng left her marketing job in 2023 to take care of her children. This year, she started experimenting with side gigs to chip away at the family debt. She said they've helped her feel more sane, not lose herself in motherhood, and recover her spark. This as-told-to essay is based on a conversation with Lana Ng, a 35-year-old stay-at-home mother based in Houston. It has been edited for length and clarity. I used to work as a marketing coordinator at a local hospital in Houston. I worked in a hybrid role, spending three days working from home and two days in the office, earning about $70,000 a year, and I enjoyed it. That's until I got pregnant with my son. I was so happy; my husband and I had been trying to get pregnant for months. I didn't really think about what would happen once he was born, as I just thought I'd go back to work. But the day before my maternity leave ended, my boss called me and said he was taking away my flexibility to work from home, and he wanted me in the office 40 hours a week. I was in a state of shock, but the choice was clear. I handed in my resignation letter almost immediately and became a stay-at-home mom in 2023. Working on 4 side hustles at once We were living on one income from 2023 to the beginning of 2026, during which we had our second child, a girl. When I gave up my salary, half of my husband's take-home pay went toward our mortgage, so the money was starting to feel a little tight. I started wishing I could bring in an extra couple of hundred dollars a month to help us cover our groceries. I saw all these women online making money through content creation and side hustles, which inspired me to try it too. Are you a side hustler? Do you use AI to help manage multiple workflows? Have you turned your side gig into your main job? We want to hear from you. Fill out this short form to share your story. So in January, I made a video on Instagram sharing how much debt my husband and I had in total and how much I planned to pay off this year, which was $18,000. This was everything except our mortgage, including our student loans, medical debt from having our kids, and some credit card debt. Over the last six months, I've dabbled in mystery shopping, creating user-generated content (UGC) for brands on social media, doing brand deals, and freelance graphic design work. I've also experimented with print-on-demand products sold through Etsy. UGC takes up most of my time, about 10 hours a week, which involves scripting, filming, and editing videos. Over the last few months, my earnings from side gigs have helped me reduce our debt from $18,000 to $13,100. Balancing side gigs and motherhood During the day, I work while my son and daughter nap. When they're both asleep at the same time, I usually have an hour or two during the day when I can work. That's really not a lot of time. But in the evening, from about 7 p.m. to midnight, I have time to do uninterrupted work because my husband comes home, looks after the kids, and handles their bedtime. My side hustles have affected the time I spend with my husband. It's either work or time with the kids, because we don't have house help. I'm thinking about having at least one day a week when I don't use social media at all so that I can focus on my marriage and family. What I love about working side gigs is that I don't have to answer to anyone, unlike in my previous job. If I wanted to take a day off because my kids are sick or to spend more time with my kids, I don't have to ask anybody. Feeling like myself again When I had my kids, I gave all of myself and my energy to them and lost myself in motherhood. I'd realize that at the end of the day, I didn't really do anything for myself. I also didn't love giving up my salary and living off one income. I understand we're a married couple and it's our money, but I wanted to bring in my own money so I could buy things for myself, and not have to rely on my husband. We budget hard, so every dollar has a job. At the beginning of the month, all the money is automatically assigned to something, like the mortgage, bills, groceries, or diapers. There's not much left over, and if we spend money on something for me, I start to feel really guilty. So I think the biggest achievement for me is that I can bring in some money, for myself and for my family. Content creation has been such a creative outlet for me, and I've always been a creative person, so it feels like my skills aren't going to waste. It's given me a little bit of my spark back. Read the original article on Business Insider

Allyson and Andrew Rappaport moved from the US to Costa Rica with their two sons. Provided by Allyson Rappaport. High taxes and concerns about school safety prompted Allyson and Andrew Rappaport to consider leaving the US. They moved from Connecticut to a car-free beach town in Costa Rica with their two sons. Living in Costa Rica allows the Rappaports to spend more quality time together as a family. For years, Allyson and Andrew Rappaport spent their days caring for animals at the veterinary practice they owned in New York, often working six days a week. At night, they returned to their cozy four-bedroom home in Ridgefield, Connecticut, where they lived with their two sons. Although they had built a comfortable life, they found the cost of living high and dreamed of slowing down. They were also worried about school safety and their kids growing up in what they saw as a highly competitive, consumption-driven culture. When the opportunity to sell their veterinary practice arose, the couple saw it as a chance to retire and make a lifestyle change. Have you relocated abroad for career opportunities, retirement, or lower living costs? We want to hear from you. Fill out this short form to share your story with a reporter. "We started soul-searching about what we really wanted our lives to look like and what we wanted our kids' lives to look like," Andrew, 65, told Business Insider. The couple knew they wanted to leave the cold behind, but they couldn't think of anywhere else in the US they wanted to live. Instead, they kept coming back to the idea of Costa Rica, a country they'd visited multiple times and loved. The couple fell in love with Las Catalinas, a car-free beach town in Costa Rica. Rani Zerafa/Getty Images When they returned for another family vacation, they checked out a few properties and found Las Catalinas, a car-free beachfront town on Costa Rica's Pacific coast. It felt like the perfect fit. "Immediately we're like, 'Oh my God, this is where we need to be,'" Allyson, now 50, told Business Insider. They didn't have to do much convincing with their sons, who were 6 and 8 at the time. Having already visited Costa Rica, both boys were excited about the move, she said. In 2020, they sold their Connecticut home for $915,000 and relocated permanently to Costa Rica. The car-free beach town they fell for The Rappaports, who consider themselves retired, now live in a four-bedroom home in Las Catalinas with their two sons. The couple bought the property for $1.6 million and spent about $200,000 renovating it. Provided by Allyson Rappaport. Their current residence is the second property they've owned in Costa Rica. They said they sold their first house in Las Catalinas for around $1.5 million before purchasing a larger home nearby for $1.6 million in 2021. They spent roughly $200,000 upgrading it. The home sits steps from the beach and features a pool and garden. Their annual property taxes in the US were about $15,000. Now, they pay around $5,000. A comparable beachfront house would likely have been out of reach in the US, Andrew said. "We would not have been able to afford that anywhere." One of the things the couple appreciates most about the town is its strong sense of community. In the US, they rarely interacted with their neighbors. Here, it's common to run into people they know while walking through town. Their sons attend an international school and have formed friendships with kids of different ages and backgrounds, Andrew said. "The walkability of the town has been so great for the kids, because they can just go out and we don't have to worry about them getting hit by a car," Allyson added. They say their kids get to have more freedom in Costa Rica than in the US. Provided by Allyson Rappaport. Back in Connecticut, their sons regularly participated in active-shooter drills at school, something Allyson found deeply unsettling. "Our kids were going to grow up in this environment where you have to worry about being shot in school," she said. "It's not some place I'd want to raise my kids." The reality of starting over The Rappaports are among many Americans who have chosen to make Costa Rica home. According to the US Embassy in Costa Rica, approximately 120,000 Americans live in the country. It helps that Costa Rica offers multiple residency options for foreigners, including visas geared toward retirees, digital nomads, and investors. The investor program allows foreigners to qualify for residency by making a minimum $150,000 investment in the country, including in real estate — the route the Rappaports took. Business Insider has spoken to several Americans who have relocated to Costa Rica in recent years. Many pointed to the country's affordability, slower pace of life, and outdoor lifestyle as reasons for making the move. Still, moving abroad came with a learning curve for the Rappaports. While English is widely spoken in Costa Rica, the family initially struggled with the language barrier. Now, they are fluent in Spanish. They miss some of the conveniences of American life, but say living abroad has made them more conscious consumers. Provided by Allyson Rappaport. The slower pace of life comes with trade-offs, too. Everyday tasks and bureaucratic processes often take longer than they would in the US, but the couple said it's simply part of adapting to a different culture. "Pura Vida, that's the local thing. Like, 'Oh, it's all good. Pura Vida!" Allyson said, referring to a Costa Rican phrase that's often used to express a laid-back, optimistic attitude toward life. Still, there are aspects of life in the US that the Rappaports miss. Specialty ingredients are sometimes hard to come by, and online shopping isn't nearly as convenient or accessible as it was back in Connecticut. That said, Allyson sees an upside: The family has become more mindful about what they buy. They now speak Spanish fluently and have since adapted to the slower pace of life. Provided by Allyson Rappaport. "You realize, maybe I don't really even need that," she said. More than anything else, the Rappaports say they get to spend a lot more time together. "Our kids have us growing up, both of us. We're at all the functions now," she said. Running errands also takes more time than it did in the US. A trip to the grocery store, a car service appointment, or other routine tasks can easily take several hours, Andrew said. "People say to me, 'Aren't you bored all the time?'" Andrew said. "We fill every day. There's really not a lot of downtime, yet we never feel rushed or pushed or stressed." On a typical chore-free day, that lack of urgency is easy to see. After making breakfast and getting the kids ready for school, Allyson heads to the tennis courts before the couple meets for lunch. In the afternoons, they're often at the beach, by the pool, or working out at the gym. The slower pace of life has benefited their marriage, too. "Even after seven years, it still feels like we're living on vacation," Allyson added. Read the original article on Business Insider

New York Times: Google's Earthquake Alerts system reached 11.4M+ people in Venezuela, giving them seconds or up to two minutes notice before two powerful earthquakes struck  —  Jose Flores was driving with his family to see “Toy Story 5” on Wednesday in Caracas, Venezuela, when a loud earthquake alert went off on his wife's Google Android phone.

<blockquote cite="https://blog.jonudell.net/2026/06/28/doctor-it-hurts-when-agents-create-unreviewable-prs-dont-do-that/"><p><strong><del>Human</del> Agent in the loop</strong></p> <p>I dislike the phrase “human in the loop” because it cedes authority to the machines. Let’s flip the narrative. It’s our loop, we work the same way we always have, now we recruit agents to join the team. An agent-assisted process need not be a black box that takes in prompts and emits features. [...]</p> <p>Let’s do agentic software development like that. Not as a loop we’ve been excluded from, instead as one we invite agents into.</p></blockquote> <p class="cite">&mdash; <a href="https://blog.jonudell.net/2026/06/28/doctor-it-hurts-when-agents-create-unreviewable-prs-dont-do-that/">Jon Udell</a>, “Doctor, it hurts when agents create unreviewable PRs.” “Don’t do that.”</p> <p>Tags: <a href="https://simonwillison.net/tags/jon-udell">jon-udell</a>, <a href="https://simonwillison.net/tags/coding-agents">coding-agents</a>, <a href="https://simonwillison.net/tags/generative-ai">generative-ai</a>, <a href="https://simonwillison.net/tags/agentic-engineering">agentic-engineering</a>, <a href="https://simonwillison.net/tags/ai">ai</a>, <a href="https://simonwillison.net/tags/llms">llms</a></p>