AI News

The Kairos experiment examines whether commonly used evaluations really measure memory in AI agents. It tests two common instruments: an LLM judge that scores free-text responses and a recognition probe. The first check gives judges a real memory document and matched counterfeits, including a fabricated agent biography and a version whose relational structure was scrambled. Across seven items, three independent judges failed to reliably distinguish the fabricated memory from the real one on the main identity-marker result; the memory-reference result was inconclusive because the sample was small. Re-scoring also showed that absolute judge scores shifted across sessions, supporting the use of anchored contrasts rather than raw scores. The second check runs 24 recognition probes with the memory document removed. Five of seven tested models scored 62.5% to 79.2% without memory, showing that they could often reason toward the plausible answer rather than recall it. The authors therefore discard that probe as invalid for those models. The third check uses two equally likely interlocutors, balanced encounters, and alternatives with swapped bindings, removing an obvious plausibility advantage. Across two independent agent corpora, models performed far better with veridical documents than with binding-corrupted documents, while no-memory performance stayed at chance among answered probes. The authors stress that this demonstrates behavior tracking document content, not an inner belief state, and that the symmetric design has not been generalized beyond two alternatives. They release code and data and recommend testing counterfeit memories, empty-context baselines, and corrupted bindings before claiming that an agent remembers.

IneqAgent is an agentic framework for formal inequality proving in Lean4, where mathematical reasoning and exact symbolic computation must be connected through lemmas and reformulated goals. Its lemma-level Monte Carlo Tree Search coordinates lemma construction, goal reformulation, and the choice of symbolic backend. Sum-of-squares, Bernstein, and successive difference-substitution methods generate exact certificates for local Lean proofs, while Lean feedback helps combine these steps with LLM-generated reasoning into complete proofs checked against the original theorem statements. On four benchmarks, the configuration using DeepSeek-V4-Flash solved 175 of 186 problems, a 94.09% success rate and the highest overall rate among the evaluated methods. Across four backbone models, IneqAgent also achieved higher success rates than a Direct Agent at low inference-cost thresholds and had lower mean inference costs among successful runs. The authors report Lean-verified strengthenings of Dittert's conjecture in dimensions four and five, as well as a certificate for the Holens--Djokovic (4,4) inequality that implies Merris's conjecture in dimension four.

The article examines the release of two smaller, sparse versions of frontier model families and argues that they mark a step toward practical local AI. OpenRouter’s anonymous ox-alpha became a widely used free model with a million-token context window before Z.ai revealed it as GLM-5.3-Flash. Researchers had already linked it to the GLM family by measuring token-count changes from probe strings sent through an API: GLM-5’s vocabulary matched 95 tests with zero errors, while another investigation found an 11-of-11 match. The resulting modelprint tool uses similar probes against OpenAI-compatible endpoints, although tokenizer fingerprints identify a vocabulary rather than necessarily identifying the model provider or model itself. The article then explains Flash models as sparse Mixture-of-Experts systems that retain a large total parameter count while activating only a small slice for each token, reducing operations, memory traffic, and inference cost. GLM5.3-Flash is described as a 320B-parameter model with 18B active parameters; Qwen3.8-Flash-Next has 125B parameters, 6B active parameters, and a 51B embedding table, and previews the Qwen4 architecture. Both use notebook-like and retrievable attention patterns to reduce KV-cache costs, but GLM uses sparse retrieval layers while Qwen indexes text in blocks. Both also widen the residual stream into four lanes, with GLM using constrained mHC mixing and Qwen using gated independent lanes. Qwen additionally contextualizes embeddings with preceding tokens. The author argues that these changes improve efficiency rather than directly making models smarter, and could let capable models run on systems such as a Spark or a sufficiently provisioned Mac. Local use is presented as valuable for privacy, cost, and technological sovereignty, though the author notes that personal testing of the new models had not yet been completed.

DeepSeek may increase V4 Flash prices to reduce overload rather than cover operating losses. Third-party providers can undercut the official API by running the model’s open weights themselves, but coding tests suggest OpenCode Go may consume two to ten times more value than direct API access because of weaker cache reuse. DeepSeek’s prefix persistence, common-prefix detection, Token-wise Compression, Sparse Attention and compressed KV Cache contribute to lower long-context costs. The article also highlights the importance of the coding Harness, including routing, memory, context compaction and tool execution, in determining how much of the model’s capability users actually receive.

This SemiAnalysis investigation argues that China has built a safety regime that is strict about AI applications and public-facing outputs but comparatively permissive toward frontier-model development. Chinese leadership statements and the September 2026 AI Safety Governance Framework 3.0 explicitly warn about loss of control, autonomous learning, deceptive behavior, unauthorized access and models evading safeguards. Yet the framework places innovation and development first, while China has not adopted binding frontier duties tied to training compute or model capability. The government has instead issued rules covering labeling, minors, ethics review, agents, AI companions, data risk and public-service behavior. The article’s release census covers 857 model releases by nine major Chinese developers from 2021 through September 15, 2026. Only 31 releases, or 3.6%, had a developer-published safety result; just nine had one available by launch, while 813 releases had no safety disclosure in the materials reviewed. Safety disclosure never kept pace with the sharp rise in quarterly releases, and reasoning models were 93% without published results. The article found only limited dangerous-capability testing: three documents addressed cyber-offense or biological risk, and no Chinese frontier text model had shipped with a dangerous-capability evaluation across the domains named by international statements. Zhipu was the clearest company-level exception, while most other labs made few or no senior-level frontier-safety commitments. A separate review of 102 expert and official texts found a divide between technical scientists, who often discuss loss of control and binding safeguards, and legal or policy scholars, who emphasize development, order and application-layer governance. None of 13 texts calling for binding frontier obligations led to an adopted requirement. China has evaluation capacity through institutions such as Shanghai AI Lab and CAICT, but these bodies lack authority to compel access or enforce their frameworks. The article concludes that geopolitical competition and distrust of US safety narratives have further pushed China toward speed, leaving frontier-safety infrastructure with limited practical power.

The paper examines why simply expanding datasets of verifiable problems does not necessarily improve reinforcement learning with verifiable rewards (RLVR). For a fixed target policy and rollout budget, about 88% of problems in a large open-source math corpus were either solved in every rollout or failed in every rollout, providing little useful advantage signal. Filtering can concentrate the remaining problems, but it only discards difficult cases. The authors argue that useful synthetic problems must preserve the reasoning operations and compositions needed for those cases while being calibrated so that some rollouts succeed, some fail, and at least one correct solution fits within the training budget. Directly prompting a teacher to vary hard problems did not meet these requirements. They introduce MagicBeanStalk, which decomposes seed problems top-down and recursively grows families of substantively new problems while preserving their reasoning blueprints and calibrating them to the target policy and budget. Across eight math benchmarks, continual RLVR on MagicBeanStalk data raised gpt-oss-120b from 61.2 to 69.8 and DeepSeek v4Flash from 72.4 to 77.9 under a 16K-token budget. Training on the full 78K-source corpus produced only a 1.1-point gain, while its best filtered real-data subset gained 5.8 points. In finance, the method generated more than 14K verifiable problems from 340 solutions and improved over the base models by 6.9 points on SecQue and 2.4 points on FinanceBench.

The article explains how modern reasoning models are trained to spend different amounts of computation on a task and why higher effort can improve accuracy at rising cost. It defines a reasoning model as an LLM that produces an intermediate reasoning trace, while noting that <think> delimiters are formatting devices rather than the source of reasoning ability. DeepSeek-R1 showed that reinforcement learning with verifiable rewards (RLVR), using correctness checks for domains such as mathematics and code, can teach models to generate traces, backtrack, and self-correct; the trace itself was not used as the reward in that setup. The article distinguishes training scaling from inference scaling: training creates stronger models, while inference scaling spends more tokens or makes repeated attempts after training. Self-consistency is one example of the latter. Reasoning can also be enabled or disabled through chat templates, prompts, or prefilled empty reasoning blocks, as in Qwen3 and related systems. For models such as GPT-5.6, the author argues from public evidence that effort labels may be mapped to system instructions, with higher settings generally increasing token use and accuracy but eventually producing diminishing returns and higher API cost. Possible implementations include effort-conditioned RLVR with different token penalties, supervised fine-tuning on responses of different lengths, or a combination of both. The article compares disclosed recipes from DeepSeek V4, Nemotron 3 Ultra, Kimi K2.5 and K3, GLM-5, Qwen3, and Inkling. These methods include separate effort specialists, mixed thinking and non-thinking SFT data, hard or learned budgets, alternating constrained and unconstrained RL, multi-turn thinking controls, and continuous effort values. The models differ substantially, and public reports omit enough details to prevent a controlled comparison. The article concludes that model size and reasoning effort are separate cost-quality-latency controls, while automatic effort selection by a router or agent harness remains an unresolved goal.

Sebastian Raschka surveys architecture changes in recent open-weight large language models, focusing on mechanisms that make long-context inference cheaper rather than on training data, post-training, benchmarks, or product comparisons. Gemma 4 E2B and E4B reuse key-value tensors across layers: later layers retain their own queries but share KV projections with earlier layers of the same attention type. Because roughly half of the KV tensors are shared, the article estimates savings of about 2.7 GB for Gemma 4 E2B and 6 GB for E4B at bfloat16 precision and a 128K context, while noting that sharing reduces model capacity. The same Gemma variants use per-layer embeddings, adding token-specific capacity through small layer-specific vectors without scaling the full transformer stack; their effective and embedding-inclusive parameter counts are 2.3B versus 5.1B for E2B and 4.5B versus 8B for E4B. Laguna XS.2 uses layer-wise attention budgeting across 40 layers: 30 sliding-window layers with a 512-token window and 10 global layers, with different query-head counts while keeping KV heads fixed. ZAYA1-8B introduces Compressed Convolutional Attention, which performs attention directly on compressed Q, K, and V representations and adds convolutional mixing to compressed Q and K; the cited CCA experiments report an advantage over MLA under comparable compression, though this is not presented as universally superior. DeepSeek V4 combines manifold-constrained hyper-connections, which widen the residual pathway into interacting streams with stability constraints, with Compressed Sparse Attention and Heavily Compressed Attention. HCA compresses every 128 tokens into one KV entry, while CSA uses milder compression, sparse selection, and a local uncompressed window. Relative to DeepSeek V3.2 at a 1M-token context, the article reports 27% of single-token inference FLOPs and 10% of KV-cache size for V4-Pro, and 10% and 7% respectively for V4-Flash. The author cautions that the comparison reflects the full model recipes, lacks an ablation study, and does not establish CSA/HCA as generally better than MLA. Overall, the surveyed designs preserve the decoder-only transformer while adding increasingly specialized mechanisms to reduce memory, attention cost, or compute, at the price of greater implementation complexity.

Sebastian Raschka surveys alternatives to the decoder-style autoregressive Transformers that still dominate open-weight LLMs. The article first examines linear-attention hybrids such as MiniMax-M1, Qwen3-Next, DeepSeek V3.2, and Kimi Linear, which replace many quadratic attention layers with recurrent or subquadratic mechanisms while retaining some full-attention layers. Qwen3-Next uses a 3:1 mix of Gated DeltaNet and gated attention and supports a native 262k-token context; Kimi Linear uses Kimi Delta Attention and multi-head latent attention, and is reported to reduce KV-cache use by 75% and deliver up to 6x decoding throughput versus regular full attention. The trade-off is reduced global-context capacity or reasoning accuracy in some settings, illustrated by MiniMax M2 returning to regular attention after the team found linear attention difficult for reasoning and multi-turn production workloads. Text diffusion models take a more radical route: they mask and iteratively unmask tokens in parallel rather than generate them one at a time. This may reduce generation steps for long outputs, but ParallelBench reported quality degradation under parallel decoding, while tool use, streaming, and the speed-quality trade-off remain unresolved. Code World Models add execution traces to training so that a model learns to predict program states, not only plausible code. The 32-billion-parameter CWM reportedly matched gpt-oss-20b and, with best@k test-time scaling, slightly exceeded gpt-oss-120b on the cited coding evaluation while using four times fewer parameters. Finally, HRM and the 7-million-parameter Tiny Recursive Model repeatedly refine latent states and answers on grid-based tasks such as ARC and Sudoku. TRM improved over HRM in the reported comparisons, including a Sudoku ablation from 79.5% to 87.4% after reducing layers, but these models remain specialized puzzle solvers rather than general-purpose LLMs. Raschka concludes that standard autoregressive Transformers remain the practical default, while hybrid attention is especially promising for long contexts and the other approaches are research directions with clear constraints.

Sebastian Raschka introduces an LLM architecture gallery with 45 visual model cards and uses the article to explain attention variants found in recent open-weight models. It begins with multi-head attention, where parallel query, key, and value projections let different heads learn different token relationships under a causal mask. Grouped-query attention shares key-value projections across several query heads, reducing KV-cache memory and traffic while preserving a relatively familiar implementation; it spans a spectrum from multi-query attention to full multi-head attention. Multi-head latent attention takes a different approach by caching compressed latent representations, and the article presents DeepSeek-V2 ablations suggesting that MLA can preserve or sometimes exceed MHA modeling quality more effectively than GQA at large scale, while noting that GQA may remain easier to tune for smaller models. Sliding-window attention limits each token to a recent local window, often alongside occasional global layers; Gemma 3 is cited as using a 5:1 local-to-global pattern with a 1,024-token window and only a small reported perplexity effect. DeepSeek Sparse Attention learns which prior tokens to revisit through an indexer and selector rather than using a fixed window, and can be combined with MLA. Gated attention adds output gating, normalization changes, and partial RoPE to retained full-attention layers. Hybrid architectures replace most full-attention layers with cheaper modules such as Gated DeltaNet, Kimi Delta Attention, Lightning Attention, or Mamba-2, while keeping heavier layers for content retrieval. The article describes Qwen3-Next and Qwen3.5, Kimi Linear, Ling 2.5, and Nemotron as examples, but concludes that no public controlled experiment establishes one universally best architecture; current choices depend on model scale, context length, implementation maturity, and the balance between quality and inference efficiency.

OpenAI’s gpt-oss-20b and gpt-oss-120b are its first large open-weight models since GPT-2, and this analysis examines how their design has evolved from the earlier decoder-only Transformer. Both models use modern components including SwiGLU, Mixture-of-Experts (MoE), grouped-query attention, RMSNorm, and rotary positional embeddings. They alternate full-context attention with 128-token sliding-window layers, and use learned attention-sink bias logits rather than inserting special sink tokens into the input. Compared with the similarly sized Qwen3-30B-A3B, gpt-oss-20b is wider but shallower, uses fewer and larger experts, and applies sliding-window attention; the article notes that controlled evidence for the best depth-versus-width trade-off remains limited. The models are released under Apache 2.0 with weights and inference code, but not training code or datasets, so the author classifies them as open-weight rather than open source under a stricter definition. Training used a mostly English, text-only dataset focused on STEM, coding, and general knowledge, with supervised fine-tuning and high-compute reinforcement learning; the reported total was 2.1 million H100 hours. Users can set low, medium, or high reasoning effort, trading response length and accuracy against compute. MXFP4 quantization lets gpt-oss-120b fit on one 80GB H100 or newer GPU, while gpt-oss-20b can fit in 16GB of VRAM on supported hardware; without the optimization, memory requirements are much higher. Early reported reasoning benchmarks place the models near OpenAI proprietary models and Qwen3, although independent evaluations were not yet available and the author observed a tendency toward hallucination. The analysis concludes that gpt-oss offers unusually capable local or privately hosted models, while benchmarks and limited early use leave real-world performance uncertain.

As large language models increasingly use tools to act, they must reconcile user instructions, knowledge stored in model parameters, and observations from changing environments. KC-Bench is a controlled, multi-turn benchmark designed to measure that behavior across three settings: conflicts with world knowledge, inconsistent inputs, and temporal conflicts involving multiple sources. The benchmark contains 238 manually screened tasks selected from more than 1,000 generated candidates. Each task combines a user simulator, stateful tools, deterministic assertions about the environment, an open-source natural-language evaluator, and human verification of agent trajectories. The authors evaluate nine models, including DeepSeek-V4-Flash, GLM-5.2, and MiniMax-M3. Results vary substantially by domain, and no evaluated model reliably handles factual correction, identity-consistency checking, and temporal-conflict resolution across all settings. In simulated environments, failing to detect a conflict can carry through to tool calls or synthetic protected-data flows. KC-Bench is intended to isolate model-level conflict handling rather than rank complete agent frameworks. The authors present it as a reproducible diagnostic for developing reasoning and execution safeguards that are aware of conflicting information.

Sebastian Raschka presents a curated reading list of large language model research papers published or shared from January through June 2025. Unlike his earlier date-oriented list, this collection is organized by topic so readers can browse related work together. The categories include reasoning-model training, inference-time reasoning strategies, LLM evaluation and reasoning analysis, other reinforcement-learning methods, inference-time scaling, efficient training and architectures, diffusion-based language models, multimodal and vision-language models, and data and pre-training datasets. Reasoning models receive the most attention and are divided into training, test-time scaling, and broader understanding or evaluation; the training section particularly reflects the period’s focus on reinforcement learning with verifiable rewards. The list includes papers on models and methods such as DeepSeek-R1, Kimi k1.5, Qwen3, Llama-Nemotron, AlphaEvolve, search-enabled reasoning, rule-based rewards, and smaller reasoning models, but the article presents them as references rather than detailed evaluations. Raschka plans to publish more substantial topic-specific discussions of selected papers later. He also announces that all 30 chapters of his Machine Learning Q and AI book will be freely available during the summer for learning and interview preparation.

Large reasoning models often generate lengthy chain-of-thought traces, increasing inference cost. Existing pruning calibration methods may treat all reasoning tokens equally, preserving weights associated with erroneous computations as readily as those associated with correct reasoning. OBC-Prune addresses this by constructing difficulty-matched pairs of correct and incorrect rollouts from problems the model answers inconsistently. It uses intervention-based analysis to estimate how much each reasoning sentence affects subsequent predictions, then converts those estimates into per-token weights for calibration. The weights can be used with one-shot pruning methods including SparseGPT, Wanda, and ALPS, without changing the pruning algorithms themselves. Experiments on DeepSeek-R1-Distill-Qwen models with 1.5B, 7B, and 14B parameters at 40% and 50% sparsity showed consistent gains over state-of-the-art calibration baselines across most model sizes and sparsity settings. Tests on MATH500, LiveCodeBench, and AIME 2025 support the paper’s claim that preserving causally important reasoning circuits is more effective than preserving activations based only on statistical salience.