entry-level retrievalstateful case trajectoriesEMNLP 2026 Industry Track
RAFT (Retrieval-Augmented Framework for Troubleshooting Agents) treats each closed support case not as a static document but as a directed chain of timeline entries, and retrieves at the entry level. This surfaces historical cases whose intermediate states match the active case and returns the parent-case trajectory anchored at the matched state; an optional case-level graph links related cases through a configurable similarity representation.
Because public multi-stage troubleshooting data is scarce, the authors pair a synthetic benchmark built from Microsoft Learn Windows Server documentation with real Apache Jira issues carrying human duplicate labels. RAFT improves Case Hit over vanilla retrieval-augmented generation (RAG) and GraphRAG baselines at every stage of case progress, with statistically significant gains over the strongest baseline. The benchmark, implementation, and evaluation set are released.
Why read it A concrete advance in structured retrieval โ modelling the stateful, multi-stage nature of real troubleshooting instead of flattening cases into documents.
<6 GB27B parametersin-browser WebGPU
PrismML released Bonsai 2 27B, a 27-billion-parameter model compressed to under 6 GB with what the company describes as near-lossless quality. The footprint is small enough to run locally, including in the browser via WebGPU.
Aggressive compression of mid-size open-weight models is what makes local and edge inference practical on consumer hardware; the size-versus-quality trade-off is the number to watch.
Why read it Local and edge inference live or die on footprint โ a 27B model under 6 GB pushes capable models onto ordinary laptops and into the browser.
176 matched settingscontext-overflow preventionbash-only cost win
The authors fix a lightweight coding harness's execution loop and vary three components โ planning, action space, and context management โ across four models on SWE-Bench Verified and Terminal-Bench 2.1, evaluating 176 matched settings spanning five context-management strategies and four context-window budgets.
Findings: context management matters more as the context budget tightens, mostly by preventing context-overflow failures; staging rule-based elision before LLM-based summarization is the most efficient strategy, while making elided content recoverable adds machinery models rarely use; planning shifts from an accuracy scaffold for weaker models to a cost saver for stronger ones; and predefined tools help models with weak shell proficiency, whereas bash-capable models do fine with a bash-only interface at lower cost.
Why read it Component-level evidence for building coding-agent harnesses โ useful for anyone tuning agent loops rather than swapping models.
auto HSA overrideVulkan vs HIP benchzero install
ROCmFix is a single-file, zero-install script that auto-detects an AMD GPU, looks up the correct HSA_OVERRIDE_GFX_VERSION in a live community database, applies it to the shell configuration, and runs health diagnostics. It targets the common failure where local inference on AMD hardware throws HSA_STATUS_ERROR_INVALID_ISA or silently falls back to the CPU.
It also includes a 'bench' command that runs a short generation test against Ollama or LM Studio using Vulkan and then HIP/ROCm, reporting which backend is faster on the machine โ a practical answer to the recurring Vulkan-versus-ROCm question.
Why read it Local inference on AMD hardware is still a configuration minefield; tooling that automates the overrides and measures the backends removes real friction.
production serving stackfirst-party engineering account
Z.ai published a technical account of how it built the inference infrastructure behind its GLM models, covering the serving stack and the engineering trade-offs involved in running its own models at scale.
First-party write-ups of production inference stacks are relatively rare: they show the concrete systems decisions โ scheduling, batching, and hardware use โ that determine cost and latency.
Why read it A first-party look at a production large language model serving stack, relevant to anyone building or comparing inference infrastructure.
~4-month gapdrastically cheaperMozilla report
A Mozilla report concludes that China's open-weight AI models now trail the United States frontier by roughly four months, based on benchmark comparisons. The models still lag on some benchmarks but are drastically cheaper to use.
The gap matters for anyone deploying open-weight models: a four-month lag at a fraction of the cost changes build-versus-buy decisions for local and edge inference.
Why read it Quantifies the open-weight gap โ directly relevant to choosing open models for cost-sensitive local and edge deployments.
cut-point replay23 ยตs recording overheadbit-stable replay
Large language model agent failures are hard to reproduce because model responses are non-deterministic, tools read changing state, and multi-step trajectories rarely repeat. Chronicle records an agent run at its non-deterministic boundaries as immutable envelopes and can replay it from the record.
Its central operation, cut-point replay, serves a chosen subset of boundaries from the record and executes the complementary subset live with new code, turning a recorded incident into a regression test that runs in continuous integration. On a benchmark of six recorded failures, recording adds 23 microseconds per boundary crossing, full replay issues zero model calls and is bit-stable across 20 repetitions, and cut-point tests fail on faulty code and pass on guarded and benign changes for all six incidents.
Why read it Brings software-engineering discipline to agent reliability โ reproducible regression tests for systems that are inherently non-deterministic.
hypernetwork weight generationonline Bayesian updatefixed stored footprint
Scaling laws assume static pretraining data, but a deployed model faces a world where much of the useful information โ facts a user supplies, corrections they give โ arrives at run time and is normally stuffed into the prompt and discarded. This paper asks how an architecture could learn from live interaction by writing it into its weights. Taking inspiration from Mixture-of-Experts (MoE), it proposes the Infinite-Parameter LLM: a compact hypernetwork turns run-time data into a low-rank modulation of a shared base network, so feed-forward weights are generated from live data rather than stored in a fixed bank.
Where prior weight generators read the context once and freeze, the authors carry a Bayesian belief over the generator's latent code and update it online, so the effective weight is re-derived from that evolving belief as a session proceeds. The stored footprint stays fixed, yet the weights the model can compile are effectively infinite. The paper specifies an evaluation protocol that tests this against in-context learning and retrieval.
Why read it A genuinely novel architectural direction โ carrying run-time knowledge in generated weights rather than the prompt, with implications for continual learning and context economy.