Monday, August 17, 2026

Persistent mobile assistants need memory spanning a year of heterogeneous, cross-app, multimodal activity, but existing long-term-memory benchmarks assume cloud storage and clean dialogue histories. MobileMem synthesizes year-scale interaction trajectories from real smartphone usage metadata and interviewed personas, then evaluates nine memory systems end-to-end on multi-hop, temporal, preference, and unanswerable questions.

• The KEME synthesis engine alternates top-down temporal planning — recursively expanding a persona and time horizon into an event graph, then into sessions — with bottom-up experience evolution that revises unexpanded future events and updates persona attributes with message-level evidence links, using real OPPO app-usage metadata as immutable "knowledge anchors"; the text track covers seven app templates while MobileMem-Omni adds screenshots, synthetic portraits, and bilingual dialogue in trajectories exceeding 2M tokens.
• Systems that preserve raw conversational detail win: A-MEM scores 79.68 overall and HippoRAG2 78.85 with a GPT-4.1-mini backbone (HippoRAG2 reaching 80.06 on GPT-5.4-mini), while a 128k-token long-context baseline reaches only 54.51 and compress-and-overwrite designs like Mem0 (35.63) and LangMem (24.79) collapse.
• Memory construction cost varies by roughly 4x for similar accuracy — A-MEM spends 5.46M tokens per trajectory on GPT-4.1-mini and 11.17M on GPT-5.4-mini (which extracts far more keywords per memory unit), whereas HippoRAG2 matches or beats it at about 2.8M.
• Temporal reasoning is the consistent weak spot, topping out at 72.04 against 86.30 on single-hop, and adversarial unanswerable questions invert the whole ranking: LangMem, last overall, leads at 76.42 because stronger retrievers surface weakly relevant distractors that convince the model an answer exists.
• The main caveats are that trajectories are LLM-synthesized by GPT-5.2 from only two volunteers in the text track and eight real plus eight virtual personas in Omni, global consistency is checked by manual sampling rather than any automatic metric, only about 44% of fine-grained profile fields ever surface in the generated data, and — despite the on-device framing — every evaluated system runs cloud-hosted backbones with cost reported purely in tokens, never in storage, latency, or power.

arXiv:2608.13606 PDF

Predictive-coding architectures like JEPA are usually described in terms of representation matching rather than probabilistic inference, leaving unclear what generative model, if any, they correspond to. The claim here is that a fully stochastic temporal JEPA — specifically PIB-VJEPA, where the current representation, the future target, and the transition are all distributions — instantiates the encode–transition–emit pipeline of a hidden Markov model, with the context encoder acting as an amortized filtering distribution, the predictor as the latent transition, and a decoder or inverse target encoder as the emission.

• The correspondence is graded rather than binary, split into four levels — shared computational roles, an emission-complete latent state, sequence-level HMM equivalence, and model-and-objective equivalence — with a theorem giving four sufficient conditions for the third level: first-order Markov latent dynamics, a valid state-to-observation conditional, transition-consistent latent marginals, and a history encoder that coincides with the induced filtering posterior.
• MCJEPA (Markov-Chain JEPA) makes the idea concrete by replacing the neural predictor with a learned row-stochastic transition matrix A over K categorical states, so an h-step forecast is just q_t A^h and matrix powers give exact Chapman–Kolmogorov path consistency — every decomposition of the same horizon yields an identical predictive distribution, which is what makes hierarchical planning over composed short- and long-horizon steps well-defined.
• Because the encoder, target encoder, and transition are trained jointly against a KL latent-matching loss, they can agree through degenerate solutions, so training adds an occupancy term KL(q̄_B || Unif(K)) against single-state collapse plus a per-sample entropy penalty against uniform-assignment collapse — two regularizers whose weights must be balanced, since too much occupancy pressure forces artificial state usage and too much entropy pressure causes premature hard assignments.
• The construction generalizes cleanly along both axes of latent Markov dynamics: side-information-conditioned matrices A_φ(ξ_t) for action- or time-dependent transitions, continuous-state Gaussian or flow-based kernels, continuous-time generators via exp(Δt Q_φ) for irregular sampling, and deterministic temporal JEPA recovered as the degenerate Dirac-kernel limit as the transition covariance goes to zero.
• The honest limits are substantial: standard JEPA training optimizes latent target matching plus information-bottleneck regularization, not observation-sequence likelihood, so the stated conditions are sufficient but not necessary and are not guaranteed by ordinary JEPA training; compressive target encoders are generally not invertible, the implicit Bayes-rule emission p(x|z) ∝ p_data(x) q_θ(z|x) is a static correspondence that need not be tractable for generation; and the empirical support comes from 4 controlled experiments on synthetic processes rather than any real video or large-scale benchmark.

arXiv:2608.13621 PDF

ReAct agents only reason during the Thought phase, leaving the Action–Observation interval — the wait for a tool to execute and return — as a recurring window in which no reasoning happens at all. Second Thought is a training-free inference framework that forks four auxiliary reasoning branches the moment a Thought concludes, decodes them concurrently with the main loop, and appends whatever they produced to the observation message so the next turn starts with the extra deliberation already in hand.

• The four branches each take a complementary angle on the same live trajectory — Check audits the just-finalized plan for unverified assumptions, Recall resurfaces constraints fading from the context window, Rehearse pre-computes conditional next steps, and Alternative drafts fallback strategies — so merging is plain concatenation rather than voting or ranking, and all four share the main thread's prompt prefix KV cache.
• Because the window closes unpredictably when the observation arrives, every branch emits a stream of atomic thoughts: self-contained units of ≤25 words wrapped in <thought> tags with no cross-references, so cancelling mid-generation discards at most the unit in flight, and each buffer is truncated at its last closed tag and capped at 5 thoughts per dimension.
• Across SWE-Bench Pro, Terminal-Bench 2.1, and τ³-bench with DeepSeek-V4-Flash, Qwen3.6-Plus, and MiniMax-M3, turn count falls in all nine model–benchmark pairs and main-thread decoding in six by up to 43% (roughly 20% on average there), while Pass@1 is statistically unchanged in seven pairs and significantly higher in two (+12.4 and +10.2 points, both on Terminal-Bench 2.1); a paired wall-clock replay of 50 SWE-Bench Pro instances turns this into 10.9% lower median per-task latency (256.9s to 229.0s), with branch contention costing only 2.8s against 27.9s saved.
• Against s1 budget forcing — a compute-matched control that spends the same extra reasoning tokens on the main thread's own thought — Second Thought reaches strictly higher Pass@1 with 1.3× to 3.2× less sequential decoding in all four settings where the control is applicable, and a replay ablation shows harvested thoughts substitute for work the agent would otherwise do inline (next-turn reasoning grows from 196.2 to 316.5 tokens when they are removed).
• The main costs are monetary and structural: four branches raise per-task API spend by 66.4% to 181.5% (almost entirely cached prefix reads, reducible to 16.3–35.5% by keeping only the Alternative branch), the default configuration is window-starved — letting branches run to completion on the critical path reaches 56.7% Pass@1 versus 52.0%, meaning the truncated version captures only 41% of the attainable gain — and on τ³-bench, where windows are short and failures stem from retrieval and policy adherence rather than planning, gains top out at +3.1 points.

arXiv:2608.13667 PDF

Eight years of language-model progress are measured against public benchmarks, API prices, and one reproducible small-model experiment, rather than against impressions. The picture that emerges is a capability curve still compounding fast, a cost curve collapsing at a similar rate, and a frontier that has fragmented into task-specific leaders — so the practical unit of deployment becomes a routing policy rather than a single model.

• A log-linear fit to SWE-bench Verified release scores from October 2024 to July 2026 puts annual growth in the odds of resolving a real GitHub issue at ~5.8× (R²=0.78, n=14), carrying the independent frontier to 96–97% and implying the benchmark saturates within about a year, the same lifecycle GLUE and MMLU already completed.
• Input-token prices fell ~60× from GPT-3's $60/M (2020) to GPT-5.6 Luna's $1/M (2026), and in OpenAI's own launch table Luna wins 7 of 10 agentic and professional-work benchmarks against the eleven-week-old GPT-5.5 flagship at one-fifth the price — while trailing sharply on long-context recall (41.3% vs 81.5% on MRCR) and hardest academic reasoning.
• The mid tier is squeezed out: a Pareto analysis of fifteen GPT-5.6 effort settings finds every Terra configuration dominated by a cheaper-or-smarter Luna or Sol setting, and effort itself has an interior optimum — Opus 5 peaks on Frontier-Bench at xhigh (44.4%) and drops to 43.3% at max.
• No model leads everywhere — Opus 5 tops frontend preference (1,712 Elo) and ARC-AGI-3 (30.2%, ~4× the prior 7.8% record), Fable 5 leads SWE-bench Pro at 80.0% against Sol's 64.6%, and Sol leads terminal work — yet a two-model router pairing Sol with Fable 5 recovers the entire +2.4-point gain of a six-model oracle on the fourteen-benchmark suite.
• The inference-time experiment is honest about its own weakness: a frozen Qwen2.5-1.5B configuration on 100 locked GSM8K items scores 58/100 greedy versus 62/100 for four-sample plurality (paired McNemar p=0.481, not significant) against a 79/100 any-sample oracle, so selection — not sampling — is the bottleneck; a post-hoc confidence model reaches 0.833 AUC and 47/50 correct in its top-ranked half, but it was designed after labels were visible and cross-validated on those same items.
• Provenance is the standing caveat throughout: mid-2026 numbers mix vendor claims with independent harnesses (GPT-5.5 scores 85.1% vendor-reported versus 82.6% reconstructed), the oracle router cheats by knowing benchmark identity, and targeted gains may not transfer — Opus 5's ARC-AGI-3 record collapses to a statistical tie on the held-out Witness puzzles.

arXiv:2608.13675 PDF

Circuit discovery is the most mature instrument mechanistic interpretability could offer to satisfy EU AI Act Annex IV documentation duties, but its output is only useful as evidence if two competent analysts using the same tool on the same system file compatible claims. A pre-registered multiverse over seven defensible analytic axes, with each discovered circuit mapped through deterministic code into a structured regulatory statement, finds that the filed claim flips across 73.2% of specification pairs.

• The design crosses seven axes — discovery objective, ablation operator, corruption distribution, evaluation metric, size threshold, prompt variant, seed and granularity — with every level taken from a published implementation, yielding 15,840 specifications on GPT-2 small and the indirect object identification task, of which 7,561 survived the metric-relative threshold rule and produced a claim.
• The modal claim commands only 41.1% of the space, failing a proposed filability criterion (modal share ≥ 1 − α) at every tolerance a conformity body would plausibly accept, and the top two claims contradict rather than merely differ: 41.1% attribute the behaviour to early layers and 28.5% to late layers.
• No single control rescues the filing — standardising the evaluation metric, the most influential axis, leaves the flip rate at 59.4%, while removing circuit size from the claim entirely and holding size fixed still leaves 27.1% (95% CI 0.255 to 0.286).
• The instability is structural and functional, not verbal: median pairwise Jaccard overlap between circuits is 4% and per-example functional agreement is uncorrelated at Cohen's kappa 0.015, which directly rebuts the "phantom specialisation" objection that discovery merely samples behaviourally equivalent subgraphs.
• Caveats are substantial and the authors foreground them — one model and one task (a Pythia-160m replication is pre-registered but unrun), a 52.3% discard rate skewed heavily by metric, one of the library's seven documented discovery objectives that crashes on its own canonical task, and a pooled-versus-within-size reversal showing discovered circuits are in fact more stable than a size-matched random null (0.2746 against 0.4230) once size is held constant.

arXiv:2608.13754 PDF

Practical LLM scaling grows the training-token budget in step with model size, but high-quality domain data (code, math, curated text) cannot be scaled the same way, so its share of the mixture is progressively diluted. Repeating that scarce data counteracts the dilution at the cost of overfitting risk, and this work charts where the optimum sits as a function of model size, domain, and corpus size.

• The setup sweeps repetition counts for a fixed domain while holding the tokens-per-parameter ratio (TPP) constant and scaling the token budget proportionally with parameters, using final per-domain validation loss as the criterion for the optimal number of passes.
• Contrary to the usual intuition that larger models memorize faster and should therefore see repeated data less often, the optimal repetition count mildly increases with model size at fixed TPP, meaning repeat budgets tuned at small scale are a floor rather than a ceiling.
• Across domains, the optimal repetition count is strongly negatively correlated with a domain's final validation loss — domains the model already fits well can absorb more passes, while high-loss domains saturate sooner.
• The amount of unique domain data is only weakly related to the optimal repetition count, undercutting the common heuristic of setting epochs from corpus size alone.
• The headline practical claim — that repetition counts tuned on small proxy models at matched TPP estimate the right setting for large models — is supported by trend direction rather than by reported absolute numbers or correlation coefficients here, and rests on validation loss rather than downstream benchmark quality, so how far the mild upward trend extrapolates beyond the tested size range remains open.

arXiv:2608.14071 PDF

Forecasting hourly returns for 1,000 US equities produces predictions that are nearly flat and rank stocks poorly by cross-sectional correlation — a failure the authors call forecast collapse — even though the same models forecasting trading volume in the same setting behave normally. The core claim is that collapse tracks target predictability, and that the standard squared-error objective is structurally unable to deliver both calibrated amplitude and usable cross-sectional ranking.

• The phenomenon is characterized across time-series foundation models, twelve deep-learning forecasters, and 97 public benchmark configurations, isolating two distinct causes: low predictability caps the amplitude any calibrated point forecast can take, while per-series training objectives never identify cross-series structure at all.
• This yields a calibration-ranking tradeoff — minimizing squared error drives forecasts toward flatness, whereas directly optimizing cross-sectional correlation recovers ranking but inflates forecast amplitude by more than an order of magnitude.
• CalibRank, the proposed objective, mixes the two terms to balance calibration against ranking, and on Finance1K it nearly triples cross-sectional correlation while holding forecast amplitude close to the target, improving correlation on every model tested.
• The volume-versus-returns contrast is the cleanest evidence that this is about signal-to-noise rather than architecture, since identical models and data pipelines collapse on one target and not the other.
• The broader methodological point is that conventional per-series metrics such as MSE or MAE can look healthy while the cross-series structure that downstream ranking decisions actually depend on has already degenerated.

arXiv:2608.14106 PDF

Mobius-v0 breaks the Transformer's layer-wise binding between feed-forward knowledge storage and attention-based reasoning, replacing per-layer FFNs with a single globally shared memory that every attention "reasoner" can query. The claim is that this decoupling buys both better knowledge compression and a shorter path to an answer, since latents can be refined against the full knowledge pool in a few layers rather than traversing every layer per token.

• Sharing one oversized key-value memory across layers gives the model an implicit backward residual connection — deep reasoners can reach shallow-layer knowledge, not just the reverse — plus dynamic latent reasoning in which hidden states iterate against the repository and decode several tokens at once; at scale the shared FFN is partitioned MoE-style with sparse activation to keep the cost tolerable.
• Trained from scratch as a 7B-A1B MoE on 1TB tokens, Mobius reaches the Transformer baseline's final MMLU score using only 62.6% of the data, which the authors attribute to less cross-layer redundancy in how knowledge is stored.
• Intern-S2-Mobius-35B, continually pre-trained from Qwen3.5-35B-A3B on 1TB tokens plus SFT and RL, averages 67.88 vs 65.05 on general benchmarks (AIME 2026 95.31, HMMT 2026 85.51) and 52.14 vs 18.20 on scientific ones like Mol-Instructions and MolecularIQ, while delivering roughly 4× end-to-end inference speedup.
• The speedup comes almost entirely from shorter chains of thought rather than faster forward passes — on an MMLU-Pro linear algebra item the model reaches the same correct answer in 516 tokens against 2,364, skipping the baseline's 1,147 tokens of repeated derivation and checking.
• Weaknesses are real and acknowledged: the converted model loses to its own base on UGD hard (73.02 vs 78.02) and HLE (19.11 vs 22.40), the mechanisms behind both the data efficiency and the shorter reasoning traces are hypotheses rather than established results, expert routing still carries a block-diagonal prior inherited from the source checkpoint, and the compositional-generalization evidence in the appendix comes from an unreleased newer architecture.

arXiv:2608.14290 PDF

Video generators can now fabricate convincing footage of wars, disasters, and public emergencies, but existing detection benchmarks are built on generic web video and say little about whether detectors hold up on socially consequential events. RA-Bench closes that gap by anchoring 1,830 real crisis clips across 10 social-risk categories and conditioning nine image-to-video generators on each clip's first frame, yielding 17,886 videos used to stress-test three detector families, generation properties, and survival through social sharing.

• Each real anchor is captioned with Gemini-3.1-Pro-Preview, and that caption plus the anchor's first frame is fed to four open-source and five closed-source generators, so every fake shares scene semantics and a genuine opening frame with its real counterpart — mimicking the realistic attack where a true crisis photo seeds a fabricated video.
• Detection collapses out of domain: the seven traditional detectors fall from published AUCs of 67.6–98.6% to source-level means of 43.9–57.3%, with 26 of 63 detector–generator pairs scoring below chance, and the public ranking barely transfers (Spearman 0.26), so no single detector leads across generators.
• MLLM detectors fail differently — Gemini-3.1-Pro-Preview tops the zero-shot models at only about 63% BAcc while smaller Qwen3.5 variants flip from 19.7% to 100% fake-recall purely on prompt format, and fine-tuned Skyra turns out to key on a timestamp artifact: swapping absolute timestamps for frame indices over identical frames drops it from 68.5% to 54.4–54.9% BAcc, while BusterX++ calls almost everything real (FakeR 4.1–9.1%).
• The videos that fool people are the same ones that fool machines: 20 reviewers flagged 68.6% of open-source clips but only 52.9% of closed-source ones (Seedance2.0 40.7%, Kling 45.1%), and on the 633-clip RA-Bench-HumanProof subset that all five reviewers called real, Gemini reaches just 54.7/54.5% BAcc and traditional detectors average 47.5% AUC — worse than a coin flip.
• Real-world circulation is the hardest blow: the full RA-Bench-LastMile dissemination simulation cuts mean fake-recall across five fine-tuned configurations from 46.0% to 1.4%, and more real-image conditioning (T2V → first-frame → first+last-frame I2V) steadily suppresses fine-tuned MLLM FakeR from 70.5% to 42.7% to 28.3%, though the study reports associations within one benchmark rather than causal effects and closed-source coverage is uneven because provider safety filters rejected some prompts.

arXiv:2608.14391 PDF

Interactive game world models usually autoregress pixels or latents, leaving pose, geometry, and occlusion to be maintained implicitly by the same generative sequence, so errors compound over long rollouts. Marionette instead predicts an explicit 276-dimensional articulated world state, hands exact geometry to a zero-parameter renderer, and leaves only appearance to a video-diffusion model.

• Dynamics are split into a compact decision model (ActionGPT, ~2.5M parameters, one discrete action token per frame per entity) and a larger animation model (PoseGPT, ~25M parameters, producing the 258-D body pose), after which a deterministic bridge integrates roots, places joints in metric world space, and rasterizes a three-channel geometry buffer that Wan2.2-Fun-5B renders into photorealistic RGB.
• Because the action is an explicit token, control is applied by overwriting it: feeding a temporally shuffled action stream degrades root-aligned joint error by 31% (0.272 m to 0.357 m) across 48 held-out segments, with the correct forced stream matching free-run quality at 0.272 m against 0.281 m.
• Routing every frame through the predicted state costs no measurable fidelity, with an FVD of 831 against 799 for recorded pose and 975 for an end-to-end pixel-autoregressive baseline trained on the same footage, though the bootstrap intervals for the baseline comparison overlap.
• Long-horizon failures live in the state and can be repaired there without touching the observation model: a free rollout drifts the two characters 21.2 m apart (recorded sessions stay near 5 m) with a third of frames showing ground penetration, while a terrain collider cuts the collision-frame ratio from 0.337 to 0.114 (66%) and a 6 m separation cap holds the pair at 5.1 m for 7% more foot-skate.
• The main limits are scope and drift: the dynamics model is trained on a single monster type's 173-action vocabulary while the appearance model spans 27 monsters, appearance identity decays across diffusion chunks since only a first frame anchors it, and a reported negative result shows a differentiable penetration penalty was gamed by stretching the skeleton (bone-length error rising from ~3% to ~13%), which is why constraints are imposed as inputs and post-hoc projections instead.

arXiv:2608.14530 PDF