Specifications face a basic tradeoff: leave details out, and important questions go unanswered; record every detail separately, and the specification becomes large and prolix. We investigate whether a language model can serve as a compact, living specification oracle by learning facts about a target and answering questions about it directly. We compare two ways of storing the learned facts: external text notes and changes to the model's weights. Across four families of 596-fact worlds and two Qwen2.5 model sizes, weight-only oracles benefited substantially more from structure: with the 7B model, their accuracy integrated across storage capacities was 18.5 percentage points higher on structured than unstructured worlds, compared with 1.1 points for note-sheet oracles. This advantage came at a substantial storage cost, with the smallest adapter requiring approximately 175 KiB compared with a maximum note budget of 16 KiB. Adapted weights therefore exploited latent structure more successfully, while external notes required substantially less object-specific storage.
Figures & tables
Property
Note-sheet oracle
Weight-only adapter oracle
Base decoder
Frozen Qwen model
Frozen Qwen model plus active adapter
Stored program
UTF-8 note text
Trainable LoRA parameters
Swept capacity
Allocated byte budget
Adapted layer prefix / trainable parameters
Reported final storage
UTF-8 bytes in the final note
Trainable parameters × bytes per stored parameter
Study input
Compact fact table, current note, feedback
Object-specific QA/fact training data
Selection signal
Dev accuracy
Dev accuracy
Table 1: Oracle contract for the two systems.
Figure 1: Held-out accuracy across capacity for the four semantic sweeps. Adapter capacity is adapted layer count; note capacity is allocated UTF-8 note budget. Markers show seed means, and shaded bands show the observed seed range, not confidence intervals.
Figure 2: Structure advantage, ΔAUC , computed within each model/oracle system as structured-family AUC minus independent-family AUC. Hollow points are seed-level contrasts; filled diamonds and bars show the mean and seed-level 95% Student- t interval. This is the cleanest visual summary because it does not require adapters and notes to share a byte axis.
Figure 3: Storage accounting. Notes use final UTF-8 note length; adapters use trainable parameters × 2 bytes. The plotted ranges show why a single accuracy-per-byte ratio would be misleading: note and adapter observations are separated by roughly an order of magnitude in persistent payload size.
Figure 4: Note-utilization diagnostic. The 7B 32 KiB extension uses only three matched seeds and should be treated as exploratory; among those 12 matched worlds, mean accuracy changes only from 69.3% to 69.6%, with improvements, stalls, and regressions all present. The 14B note sweep uses the 16 KiB budget more fully and reaches high raw accuracy without producing a large structure advantage.
Memory-augmented language models combine persistent or temporary memory with neural computation to retain, retrieve, and use information beyond the active context.
Existing benchmarks for large language models (LLMs) in hardware design evaluate downstream artifacts such as generated RTL, assertions, or testbenches. When a model fails such a benchmark, the failure is ambiguous: it may have misread the specification, or it may have understood the specification and failed to write the code. We present SpecRead, a benchmark that isolates specification comprehension from generation ability. SpecRead v2.1 contains 385 questions over 10 open-source OpenTitan IP blocks: exact retrieval, cross-section reasoning, contradiction detection in mutated specifications, and spec-RTL consistency checking, plus 82 controls (41 distractor, 41 consistent-RTL). Type-4 items are built from real RTL mutations; we retain only mutations that Icarus Verilog simulation shows to change observable behavior. A with-spec vs. without-spec ablation suggests the questions require the excerpt, not training recall alone (without-spec accuracy 3/20 on the t1/t2 subset), though memorization of the source text may still help spot mutations. As an initial characterization with a small model, Ministral-3B scores 33.2% overall (128/385; macro average 39.0%): 55.2% on retrieval, 51.7% on cross-section reasoning. On the two contradiction-focused types, the verdict-plus-location measure gives 48.0% (t3) and 63.3% (t4), with a 51.2% false-positive rate on distractors and 100% on consistent-RTL controls. Layered scoring shows the model locates contradictions well (78.9-81.6% location accuracy) but scores lower on their category (43.9-49.7%). A structured "rule-table" prompting intervention lowers accuracy on every question type except t2 (tied). SpecRead is automatically scorable by deterministic checks, with gray-zone cases counted wrong under the conservative main scoring. The benchmark is regenerable for type-3 items via mutation injection, and built exclusively from public sources.
Large language models increasingly serve as persistent knowledge workers, with in-context memory - facts stored in the prompt - as the default strategy. We benchmark in-context memory against Knowledge Objects (KOs), discrete hash-addressed tuples with O(1) retrieval. Within the context window, Claude Sonnet 4.5 achieves 100% exact-match accuracy from 10 to 7,000 facts (97.5% of its 200K window). However, production deployment reveals three failure modes: capacity limits (prompts overflow at 8,000 facts), compaction loss (summarization destroys 60% of facts), and goal drift (cascading compaction erodes 54% of project constraints while the model continues with full confidence). KOs achieve 100% accuracy across all conditions at 252x lower cost. On multi-hop reasoning, KOs reach 78.9% versus 31.6% for in-context. Cross-model replication across four frontier models confirms compaction loss is architectural, not model-specific. We additionally show that embedding retrieval fails on adversarial facts (20% precision at 1) and that neural memory (Titans) stores facts but fails to retrieve them on demand. We introduce density-adaptive retrieval as a switching mechanism and release the benchmark suite.
We study reasoning for accessing world knowledge stored in a language model's parameters. For example, recalling that Canberra is Australia's capital may benefit from thinking through major cities and the concept of purpose-built capitals. While reasoning language models are trained via reinforcement learning to produce reasoning traces on tasks such as mathematics, they may not reason well for accessing their own world knowledge. We first find that models do not generate their best world knowledge reasoning by default: adding a simple "think step-by-step" cue demonstrates statistically significant improvement in knowledge recall but not math. Motivated by this, we propose training models to reason over their parametric knowledge using world-knowledge question answering as a verifiable reward. After reinforcement learning on TriviaQA (+9.9%), performance also improves on Natural Questions, HotpotQA, SimpleQA, and StrategyQA by 4.2%, 2.1%, 0.6%, and 3.0%, respectively. Reasoning models are under-optimized for parametric knowledge access, but can be easily trained to reason better.