Basalt — The Shape of AI Engineering

Section 04

Research & Content

Surfacing recent work that shaped how 2026 tools handle compression, context, multimodality and long-horizon agents, so teams can interpret and apply it to their own systems.

Sample size: 1,284

Foundational Model Advances

Sievecast: Streaming Vector Quantisation at Near-optimal Distortion

Sievecast is a streaming, data-oblivious quantiser for KV-cache compression and vector search that targets mean-squared error and inner-product distortion at once.

It rotates vectors before quantising so each coordinate carries similar information, then applies scalar quantisers coordinate-wise instead of slower vector schemes.
To correct the inner-product bias of MSE-optimal quantisation it adds a 1-bit corrective residual; the authors report quality-neutral KV compression at 3.4 bits per channel.
Long-context inference depends on compression that preserves the geometry attention actually uses, not just on storing vectors in fewer bits.

Recursive Prompt Machines

RPMs treat the prompt as external state, letting a model inspect, partition and recursively call itself over slices instead of pushing everything through one window.

The model works inside a sandboxed interpreter where the prompt lives as a variable, so it can read slices, run code and recurse over the parts that matter.
Across three long-document suites, RPMs handle inputs two orders of magnitude beyond the base context window and beat retrieval-style scaffolds.
Long-context scaling may come less from larger windows and more from giving models programmable ways to search and decompose external state.

Ledgerheads: Learning to Memorise at Inference Time

Ledgerheads is a family of long-context architectures pairing a limited attention window with a neural long-term memory that keeps learning while it runs.

Its memory module writes surprising inputs into its own weights with gradient updates, then applies momentum and weight decay as a learned forgetting rule.
The paper instantiates memory as context, as gating and as a layer; across language, reasoning, time-series and genomics the context variant does best on long dependencies.
Long-context modelling may benefit more from explicit write/forget/retrieve systems than from ever-larger attention windows alone.

Marlin V2: Visual Agentic Reasoning

Marlin V2 is an open-weight multimodal agentic model that jointly optimises text and vision, with a parallel agent-execution framework layered on top.

It mixes vision tokens early in pretraining, uses a native-resolution encoder for images and video, and follows with zero-vision SFT plus joint multimodal RL.
A trainable orchestrator spawns frozen specialist subagents and runs heterogeneous subtasks concurrently; the report cites 76.8% on a verified engineering benchmark.
Agentic performance may come less from one monolithic model working in sequence and more from trained foundations paired with learned orchestration.

Does Retrieval Really Struggle on Long Context?

Cachewell introduces KV-level retrieval for long-context tasks, treating the cache itself as the retrieval surface rather than the raw text.

Inputs are split into fixed-size KV pages with bookmark tokens; a trained page retriever selects important pages per layer while the rest live off-device.
Across three long-document benchmarks, Cachewell beats standard retrieval pipelines and other efficient long-context methods.
The framework reframes retrieval as choosing which cached representations to keep, rather than which raw tokens to stuff into the prompt.

Rethinking Ensembles of Agents

Self-Blend examines whether diverse model ensembles are actually necessary for strong ensemble-of-agents performance.

Instead of querying several different models, Self-Blend repeatedly samples one strong model and aggregates its responses, trading cross-model diversity for in-model diversity.
Experiments on two preference benchmarks show Self-Blend outperforming traditional ensembles when proposer quality is high.
A sequential variant aggregates in sliding windows to stay inside context limits while scaling the number of samples.

Application-Layer Innovations

Prismyard Context-1: Training a Self-Editing Search Agent

Context-1 is a 20B agentic search model derived from an open base and designed to act as a retrieval subagent rather than answer the question directly.

It decomposes a task into subqueries, searches over multiple turns, and selectively prunes its own context to drop irrelevant documents.
Trained on more than 8,000 synthetic tasks with a curriculum shifting from recall-heavy exploration to high-precision retention, it improves prune accuracy from 0.824 to 0.941.
Multi-hop retrieval may work best as a specialist problem, where small purpose-trained models plan, explore and actively edit context.

Cadence 2 Technical Report

Cadence 2 is a domain-specialised model for agentic software engineering, built for long-horizon coding work while staying fast enough for interactive use.

Its recipe combines continued pretraining on a code-heavy mix and a 256k context extension with targeted coding SFT, then large-scale RL in a production tool-use harness.
The report introduces an internal benchmark drawn from real engineering sessions rather than curated public repos; Cadence 2 scores 61.3 on it.
Coding agents improve most when the model, training environment, reward design and evaluation all match the actual engineering workflow.

Terrace: Reflective Prompt Evolution Can Outperform RL

Terrace is a reflective prompt-evolution method that optimises instructions using execution traces instead of updating model weights.

The system samples rollouts, has the model analyse its own traces in natural language, and proposes new prompts; a frontier keeps candidates that win on different data subsets.
Across four tasks, Terrace matches or beats policy-gradient RL with up to 35x fewer rollouts.
The work treats prompts as an external optimisation layer, using natural-language reflection rather than heavyweight RL.

Foldstep: Single-Agent RL for Deep Web Research

Foldstep is a reinforcement-learning framework for training one web-research agent that decides when to search, browse or execute code.

The agent uses three minimal tools — search, fetch and a stateless interpreter — plus a self-managed memory call that lets it control long-horizon context.
Length-normalised RL stabilises multi-step optimisation and prevents degenerate, repetitive tool use.
Results on open-ended research benchmarks highlight that context management and planning are the core bottlenecks, not model size.

Thrift-1: Constant-Memory Long-Horizon Agents

Thrift-1 is an RL framework that trains agents to work over long multi-turn tasks while keeping memory use close to constant.

At each step previous memory and new observations merge into a compact internal state token and older context is discarded; a masked-trajectory scheme rebuilds valid rollouts.
Thrift-1-7B matches or beats much larger baselines on tasks with up to 16 sequential objectives while cutting memory use by roughly 3.7x.
Long-horizon behaviour can come from learned internal state handling rather than from expanding windows or bolting on external memory.

Quarrel-R1: Training Models to Reason and Search Together

Quarrel-R1 trains models to interleave step-by-step reasoning with live search-engine queries.

The framework uses a structured template — think, search, information, answer — where policy updates apply only to model-generated segments, treating search as environment.
Evaluated across seven QA datasets, Quarrel-R1 delivers large gains over strong retrieval baselines, including on multi-hop tasks.
The paper positions targeted, RL-trained search behaviour as an alternative to static top-k retrieval and hand-crafted tool chains.

The Shape of AI Engineering

Navigate the Report

Engineering Team Velocity

Agent Tool Adoption

Model Growth Trends

Research & Content

Engineering Team Velocity

Agent Tool Adoption

Model Growth Trends

Research & Content

Foundational Model Advances

Sievecast: Streaming Vector Quantisation at Near-optimal Distortion

Recursive Prompt Machines

Ledgerheads: Learning to Memorise at Inference Time

Marlin V2: Visual Agentic Reasoning

Does Retrieval Really Struggle on Long Context?

Rethinking Ensembles of Agents

Application-Layer Innovations

Prismyard Context-1: Training a Self-Editing Search Agent

Cadence 2 Technical Report

Terrace: Reflective Prompt Evolution Can Outperform RL

Foldstep: Single-Agent RL for Deep Web Research

Thrift-1: Constant-Memory Long-Horizon Agents

Quarrel-R1: Training Models to Reason and Search Together

Automatically review PRs against your team's own standards