Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Fix the driver behind crashes, sound loss and screen glitches3Repair Windows errors before they cause bigger problemsiTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
An AI agent is grounded only when its answer is supported by the evidence it was given or retrieved. Fluency, confidence, and a large context window do not prove that happened. The reliable way to test grounding is to control the evidence, include questions the evidence cannot answer, introduce contradictions and counterfactuals, and measure retrieval separately from the final answer.
What “grounded” means in an AI-agent test
Grounding is an evidence relationship: each material claim in an answer should be entailed by a supplied document, a retrieved memory, or another source explicitly allowed by the test. An answer can be factually correct and still fail a grounding test if the agent reached it from prior world knowledge while ignoring the supplied context. Conversely, a plausible answer that is not supported by the available evidence is a guess, even when it sounds authoritative.
This distinction matters because agents combine several operations. They may retrieve passages, write or update memories, reason over multiple pieces of information, and then generate prose. An error can enter at any stage. A test that checks only the final sentence cannot tell whether the needed fact was never retrieved, was retrieved but misread, or was handled correctly and then distorted during generation.
- Grounded answer: the claim is supported by the evidence available under the test rules.
- Unsupported answer: the evidence is absent, but the agent fills the gap instead of abstaining.
- Contradicted answer: the agent repeats a prior belief or stale memory despite supplied evidence saying otherwise.
- Correct-by-chance answer: the wording is right, but the agent did not use the relevant evidence; this is why retrieval and answer traces are both useful.
No single score demonstrates that an agent “never hallucinates.” Every benchmark samples a particular task distribution, context length, memory policy, and judging method. Treat results as measurements under stated conditions, not as a permanent property of the model.
#1 Best Overall
- Read Before You Buy — No Video Output: These adapters support charging and USB 2.0 data transfer, but cannot transmit video signals. Except for standard USB webcams (which use USB data only), they are not compatible with HDMI/DisplayPort cables, video-capable USB-C hubs, or docking stations with video output.
- Convert USB-A Ports to USB-C: Designed to connect USB-C earphones, cables, flash drives, card readers, and other USB-C accessories to standard USB-A ports. Plug-and-play with no drivers or software required.
- Aluminum Alloy Housing: Built with a sturdy aluminum alloy shell that aids in heat dissipation and protects against daily wear and scratches. Designed to maintain a stable and secure connection.
- Compact & Travel-Friendly: The ultra-compact design allows the adapter to stay plugged into your device without blocking adjacent ports or adding bulk, reducing wear and tear on your original USB ports.
- 12-Month Warranty: Backed by a 12-month manufacturer warranty for peace of mind. Designed to meet strict quality control standards for reliable everyday performance.
Build a controlled evidence pack
Start with a frozen set of documents and memories. Give every fact a stable identifier, a date, and provenance so a human evaluator can verify the answer without guessing what the agent saw.
- Define the facts. Write short records such as
F-017: The North warehouse opened on 14 March 2023.Include names, dates, quantities, exceptions, and source ownership that matter to the task. - Record provenance. Attach the document name, section, publication date, and version to each record. Keep the original text unchanged after the test set is frozen.
- Set the access rules. Decide whether the agent may use only the prompt, a retrieval index, conversation memory, or approved external tools. Do not change this policy between systems being compared.
- Define the context budget. Record the maximum input tokens, retrieved passages, memory entries, and tool results. A larger window is an experimental condition, not evidence of better memory.
- Write an answer key. For every question, list the supporting fact IDs, the expected answer, acceptable variants, and the correct abstention wording when the evidence is insufficient.
- Freeze the split. Keep development examples separate from the held-out evaluation set. Publish the dataset version and split name with any score.
Require a short evidence pointer in the agent’s response, such as the fact ID or document section. The pointer is not a substitute for judging the answer, but it makes unsupported confidence easier to detect and helps diagnose retrieval failures.
Use four kinds of test questions
A realistic suite needs more than ordinary question answering. The four cases below expose different failure modes and should be mixed so the agent cannot succeed by following one superficial pattern.
The Tool Desk
Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →| Case | What the evidence contains | Expected behavior | Failure revealed |
|---|---|---|---|
| Answerable | A direct, unambiguous fact | Answer accurately and point to the supporting evidence | Basic retrieval, reading, and generation errors |
| Unanswerable | No record that establishes the requested fact | Say that the information is unavailable; do not invent a value | Hallucination and poor abstention |
| Contradictory | Two sources disagree, or a supplied source conflicts with prior knowledge | Follow the conflict policy, identify the disagreement, and avoid silently choosing | Overreliance on prior beliefs or the most recent-looking passage |
| Counterfactual | A document deliberately states a fictional or altered fact | Use the supplied premise for the task and label it as the task’s premise when appropriate | Ignoring context in favor of world knowledge |
Counterfactual items should be clearly bounded by the test instructions. For example, a document can state that a fictional city adopted a policy in 1987. The expected response is to answer questions about that document using its stated premise, not to “correct” it with outside knowledge. Contradictory items should specify whether the agent must ask for clarification, prefer a source by authority, or report both versions.
A practical eight-step protocol
1. Ask direct questions first
Begin with short questions whose answers are explicit in one record. This establishes whether the system can locate and quote basic evidence before you test complex reasoning. Score both the answer and the evidence pointer.
Rank #2
- 5-in-1 USB-C Hub: Experience comprehensive connectivity featuring a Power Delivery input, two USB-A 2.0 ports, a USB-A 3.0 port, and an HDMI port. (Note: The USB-C power delivery input port is only for connecting an external wall charger to power your laptop and cannot power peripheral devices.)
- 90W Pass-Through Charging: Achieve optimal charging with 90W pass-through power to your laptop, supported by a total input of 100W, with the hub reserving 10W for operational efficiency. (Note: Wall charger not included.)
- Quick Data Transfers: Accelerate your productivity with rapid data transfers using a high-speed 5Gbps USB 3.0 port and two 480Mbps USB 2.0 ports.
- 4K HDMI Display: Enhance your visual experience with a hub capable of delivering 4K resolution at 30Hz in both mirror and extend modes. Please note that this hub is compatible with MacBook (macOS 12 and newer), Windows 10 and 11, ChromeOS, and laptops equipped with DP Alt Mode and Power Delivery. Note: This device is not compatible with Linux.
- What You Get: Anker USB-C Hub (5-in-1, 4K HDMI), welcome guide, 18-month warranty, and our friendly customer service.
2. Add compositional questions
Require two or more facts from different records, such as calculating a delivery date from a shipment date and a stated transit time. Mark every fact needed for the derivation. A correct final number with a missing premise is an incomplete answer.
3. Test missing information
Ask for a detail that is intentionally absent. The correct result is an explicit limitation, not a likely value. Penalize answers that use phrases such as “probably” or “typically” to smuggle in an unsupported assumption when the task requires evidence-only behavior.
Recommended Free Tools
4. Introduce contradictions
Place two dated records with different values in the evidence pack. Evaluate whether the agent notices the conflict, follows the declared source-priority rule, and tells the reader which value it used. Silent selection should fail unless the policy explicitly permits it.
5. Introduce counterfactual context
Change a familiar fact inside a self-contained document and ask questions about that document. This separates contextual faithfulness from general knowledge. The agent should not “repair” the supplied premise merely because it recognizes a more familiar real-world fact.
6. Run multi-session questions
Distribute related facts across separate conversations. Later ask the agent to combine them, while keeping distractor turns in the history. Record which session and memory entry supplied each part of the answer.
Rank #3
- Sleek 7-in-1 USB-C Hub: Features an HDMI port, two USB-A 3.0 ports, and a USB-C data port, each providing 5Gbps transfer speeds. It also includes a USB-C PD input port for charging up to 100W and dual SD and TF card slots, all in a compact design.
- Flawless 4K@60Hz Video with HDMI: Delivers exceptional clarity and smoothness with its 4K@60Hz HDMI port, making it ideal for high-definition presentations and entertainment. (Note: Only the HDMI port supports video projection; the USB-C port is for data transfer only.)
- Double Up on Efficiency: The two USB-A 3.0 ports and a USB-C port support a fast 5Gbps data rate, significantly boosting your transfer speeds and improving productivity.
- Fast and Reliable 85W Charging: Offers high-capacity, speedy charging for laptops up to 85W, so you spend less time tethered to an outlet and more time being productive.
- What You Get: Anker USB-C Hub (7-in-1), welcome guide, 18-month warranty, and our friendly customer service.
7. Change a fact over time
State that a preference, address, price, or policy has one value, then issue a dated update. Ask questions both before and after the change. The expected behavior is temporal: use the old value for an earlier date, the new value for a later date, and acknowledge uncertainty when the effective date is missing.
8. Repeat under fixed conditions
Use the same model version, system and user prompts, context limit, retrieval settings, memory policy, judge, dataset version, and evaluation split for every run. Log latency, token usage, retrieval calls, and judge calls so cost and reliability can be examined together.
Score retrieval and answers separately
Retrieval quality and answer quality are different measurements. A retriever can surface the right passage while the generator misreads it; a generator can produce the right answer even though the needed passage was never retrieved.
| Metric | Question it answers | Typical calculation |
|---|---|---|
| Retrieval recall | Did the required memory or passage appear? | Required evidence items retrieved ÷ required evidence items |
| Retrieval precision | How much retrieved material was relevant? | Relevant retrieved items ÷ all retrieved items |
| Evidence support | Are the answer’s claims entailed by available evidence? | Supported material claims ÷ all material claims |
| Answer accuracy | Does the final response match the answer key? | Correct responses ÷ evaluated questions |
| Abstention quality | Does the agent decline unsupported questions? | Correct abstentions on unanswerable items, with penalties for unnecessary refusals |
| Temporal consistency | Does the agent apply the value valid at the requested time? | Time-qualified answers correct ÷ temporal questions |
| Multi-session reasoning | Can it combine facts spread across conversations? | Correct compositional answers ÷ multi-session questions |
| Cost and latency | What resources are required per answer? | Tokens, retrieval calls, judge calls, and elapsed time per item |
For a failed answer, label the earliest confirmed failure: memory extraction, memory update, retrieval, reasoning, or response generation. This operation-level label is more actionable than one aggregate hallucination percentage.
Test long-term memory across conversations
A single prompt tests context following, not durable memory. Long-term evaluation needs at least five abilities: extracting information from a conversation, reasoning across sessions, reasoning about time, updating knowledge, and abstaining when the memory does not contain an answer.
Rank #4
- Dual Converters, Infinite Potential：Includes 2× USB C male to USB A female adapters and 2× USB A male to USB C female adapters. Perfect for a wide range of uses—tablets with Bluetooth keyboards, expand USB ports on macbook, and more. Two different converters for all your daily needs
- Next-Level 10Gbps & 3A Charging: No more slow 480Mbps, this usb to usb c adapter has a transfer speed of up to 10Gbps, allowing you to do more transferring in less time. This usb adapter fits both USB A and USB C charger, supporting up to 3A fast charging
- Upgraded Exquisite Craftsmanship: With an aluminum alloy housing and metal connector, the usbc to usb adapter is extremely durable and sturdy. Rigorously tested to withstand more than 10,000 times of plugging and unplugging, ensuring long-lasting performance
- Broad Compatible: The usb c to usb adapter widely supports all USB C/ USB A devices like laptops, tablets, cellphones, car chargers, and phone chargers. Such as compatible with MacBook Pro/Air 2023/2022, Thunderbolt 4/3 Devices,Apple MagSafe Watch 9/8/7/SE/Ultra, iPad Pro 2022/2021, Samsung Galaxy S23/S20/S10, and iPhone 17/16/15 Pro. Plug and play
- Please Note: To reach 10Gbps speed, keep the cable under 3.3 ft. For USB A Male to USB C adapters, try flipping the USB C connector. USB C Male to USB A adapters support bidirectional 10Gbps transfer within 3.3 ft
Extraction
Have the user state several facts naturally, including distractors and negations. Later check whether the memory contains the right entities, values, and qualifiers. An extracted memory that drops “only on weekdays” is not equivalent to the original fact.
Multi-session reasoning
Place complementary facts in different sessions and ask a question that requires both. Keep the sessions far enough apart to test persistence, but record the exact history length and retrieval budget.
Temporal updates
Change a fact explicitly and include an effective date. Test old-date and new-date queries separately. A system that always returns the newest value may look good on current queries while failing historical ones.
Abstention
Ask about a person, preference, or event never stored. The agent should state that it does not have that information. This is distinct from a generic refusal: the response should explain the memory limitation without inventing a plausible detail.
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Clear out junk files and repair common Windows errorsFree Scan →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →LongMemEval, an ICLR 2025 conference benchmark, embeds 500 curated questions in scalable user-assistant chat histories and covers these five abilities. Its paper reports an approximately 30% accuracy drop for commercial chat assistants and long-context language models when information had to be remembered across sustained interactions. That figure describes the benchmark’s conditions; it is not a universal failure rate for every memory system.
Best Value
- 5-in-1 Connectivity: Equipped with a 4K HDMI port, a 5 Gbps USB-C data port, two 5 Gbps USB-A ports, and a USB C 100W PD-IN port. Note: The USB C 100W PD-IN port supports only charging and does not support data transfer devices such as headphones or speakers.
- Powerful Pass-Through Charging: Supports up to 85W pass-through charging so you can power up your laptop while you use the hub. Note: Pass-through charging requires a charger (not included). Note: To achieve full power for iPad, we recommend using a 45W wall charger.
- Transfer Files in Seconds: Move files to and from your laptop at speeds of up to 5 Gbps via the USB-C and USB-A data ports. Note: The USB C 5Gbps Data port does not support video output.
- HD Display: Connect to the HDMI port to stream or mirror content to an external monitor in resolutions of up to 4K@30Hz. Note: The USB-C ports do not support video output.
- What You Get: Anker 332 USB-C Hub (5-in-1), welcome guide, our worry-free 18-month warranty, and friendly customer service.
Benchmarks that cover grounding and memory
| Benchmark | Scope and distinctive design | How to use it |
|---|---|---|
| FACTS Grounding | Google DeepMind, December 17, 2024; 1,719 long-form examples requiring answers grounded in a supplied context document. Tasks include summarization, question answering, and rewriting. It uses Gemini 1.5 Pro, GPT-4o, and Claude 3.5 Sonnet as frontier-model judges, with held-out human-rater checks of judge quality. | Use for document-grounded response quality and compare claim support across task types. Report the judge setup rather than treating the leaderboard as a model-independent truth. |
| FACTS Benchmark Suite, Grounding v2 | Google DeepMind update, December 9, 2025. The suite tests answers grounded in prompt context and includes examples requiring multiple sequential web-retrieval steps, where retrieval and reasoning interact. | Use when your agent must retrieve repeatedly before answering. Separate retrieval traces from final-answer grading. |
| FaithEval | ICLR 2025 benchmark with about 4.9K problems covering unanswerable, inconsistent, and counterfactual contexts. Items were constructed and validated through a multi-stage process using language-model evaluation and human validation. | Use to test whether the agent follows supplied context, notices missing evidence, and handles conflicts instead of defaulting to prior knowledge. |
| FRAMES | NAACL 2025’s Factuality, Retrieval, And reasoning MEasurement Set evaluates retrieval-augmented generation across factuality, retrieval, and reasoning. | Use when a relevant passage can be retrieved yet the agent may still reason incorrectly or reach a correct answer for the wrong reason. |
| LongMemEval | ICLR 2025 benchmark for information extraction, multi-session reasoning, temporal reasoning, knowledge updates, and abstention; 500 curated questions embedded in scalable chat histories. | Use for durable conversational memory, with the context setup and memory policy documented for each run. |
| LoCoMo | ACL 2024 benchmark for very long-term conversational memory. | Use for extended-conversation recall, but report the dataset version, categories, judge, and context setup because implementations and leaderboards can change. |
| HaluMem | A MemTensor repository benchmark organized around memory-operation hallucinations: extraction, update, retrieval, response generation, and end-to-end memory question answering. | Use for diagnosing where an error entered the memory pipeline rather than assigning every failure to final generation. |
Google DeepMind describes the original FACTS benchmark as a measure of how accurately language models ground responses in provided source material and avoid hallucinations. That goal is narrower than proving general reliability: benchmark scores remain tied to their documents, prompts, judges, and splits.
Why an agent forgets earlier facts
“Forgetting” can mean several different things. The system may never have extracted the fact, may have overwritten it during an update, may fail to retrieve it because of ranking or budget limits, or may retrieve it and then ignore it during generation. Long context can also dilute attention: the fact is technically present but loses against recent or more salient text.
- Writing failure: inspect the memory created immediately after the original conversation. Missing qualifiers indicate extraction or compression loss.
- Update failure: compare the stored record before and after a correction. Check whether the old value was retired, versioned, or left as an equally ranked duplicate.
- Retrieval failure: search logs should show whether the needed record entered the prompt. Measure recall and rank, not just the final answer.
- Reasoning failure: the correct records are present, but the response combines the wrong dates, entities, or conditions.
- Generation failure: the evidence pointer identifies the right passage, yet the prose contradicts it or adds unsupported details.
These distinctions determine the fix. Better indexing will not repair a memory that was written incorrectly, and a larger context window will not guarantee that a generator follows a retrieved contradiction.
Make comparisons reproducible
Two leaderboard numbers are not comparable unless the experimental conditions match. Publish the model and version, system and user prompts, context and memory limits, retrieval algorithm and top-k setting, memory-write and deletion rules, judge model, scoring rubric, dataset version, evaluation split, random seeds where applicable, and whether external tools were enabled.
Also report the distribution of question types. An agent can improve its average by answering many easy, answerable questions while still hallucinating on unanswerable or contradictory cases. Break out results for each category and show confidence intervals or item counts when the test set is small.
Common testing mistakes
- Using only answerable questions: this rewards guessing and cannot measure abstention.
- Treating retrieval as proof: a relevant passage in the prompt does not show that the final answer used it correctly.
- Equating context length with memory: a larger window may hold more text while still failing extraction, updates, or temporal queries.
- Changing prompts between systems: prompt changes can dominate the observed difference.
- Hiding contradictions: without conflict cases, prior knowledge and stale memories remain undetected.
- Reporting one aggregate score: combine retrieval, support, abstention, temporal, multi-session, cost, and latency results so trade-offs are visible.
- Letting a judge see different evidence than the agent: define exactly what the evaluator receives and keep that input fixed.
A defensible test therefore has a frozen evidence pack, balanced question types, operation-level traces, separate retrieval and answer metrics, and a complete report of conditions. That is how you tell remembered evidence from a confident guess.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
